{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"pygments_lexer":"ipython3","nbconvert_exporter":"python","version":"3.6.4","file_extension":".py","codemirror_mode":{"name":"ipython","version":3},"name":"python","mimetype":"text/x-python"}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"markdown","source":"# Exploring the landscape of AI Ethics\n\n# Introduction\n\nAs AI becomes ubiquitous in our lives moving forward in this decade,\nfocusing on the ethical implication of AI is not just essential but\nextremely pressing. Several ethical guidelines and principles have been\nreleased around the world by governments, organizations, and\ncompanies [\\[1\\]](#ref-jobin2019global). Among these ethical\nconsiderations, there are common principles that should be promoted in\nthe development of AI systems. These principles form an initial\nconsensus of features or components that should be embedded in AI\nsystems in order to make their use more socially acceptable. The\nfigure <a href=\"#fig:eth_pr\" data-reference-type=\"ref\"\ndata-reference=\"fig:eth_pr\">1</a> showcases the most common principles\nfound in existing ethical guidelines [\\[1\\]](#ref-jobin2019global).\nThese principles include: *fairness*, which promotes bias-free and\nnon-discrimination in AI systems when used to make decisions in\nhigh-stake scenarios; *privacy and data protection*, to ensure privacy\nis guaranteed for sensitive data of individuals used in the process.\n*Responsibility and Accountability*, to promote processes and rules to\nenforce ethical considerations throughout the entire lifecycle of AI\nsystems to limit unintended outcomes; *Safety and robustness*, to\npromote the robustness and reliability of AI systems in different\nreal-world scenarios. *Transparency and explainability*: to uncover\ninformation about how the system works and explain the decisions made;\n*Environmental impact*, to study the impact the growing deployment and\nuse of AI systems might have in the environment along with how these\nsystems can be leveraged for environmental protection. All these\nprinciples play a key role in fostering trust between all stakeholders\nin the AI system lifecycle. In this report, we walk through these\nprinciples and present recent efforts, solutions, and regulations for\nimplementing each ethical guideline in the AI lifecycle.\n\n<center>\n<figure id=\"fig:eth_pr\">\n<img\nsrc=\"https://www.linkpicture.com/q/responsible_ai_fairness_robustness-Page-2.drawio.png\"\nstyle=\"width:400px\" />\n<figcaption>Figure 1: Ethical principles of responsible AI</figcaption>\n</figure>\n</center>\n\nAs this report focuses on a sociotechnical issue, we covered each\nguideline from both social and technical perspectives. We cover the\nsocial impact and value of each guideline and discuss technical\ncontributions from the literature to address them.\n\n# 1 Privacy and Data Protection\n\nEnsuring the security and privacy of machine learning models have become\na crucial issue as they are widely used in various fields. In order to\nprovide trustworthy AI, it is important to include safety, privacy, and\nsecurity in their lifecycle. Based on some definitions, safety means\nreducing the probability of expected and unexpected\nharm [\\[2\\]](#ref-varshney2019trustworthy). According to this\ndefinition, a machine-learning model should be trained and released to\nbe robust against different kinds of uncertainty; in other words, in\ncase of an accident, the model should be able to continue its expected\nnormal behavior. To prevent safety problems, first, the designer should\nbe able to specify the correct objective function and have a method to\nevaluate it; and second, sufficient data, time, and infrastructure\nshould be available to train and evaluate the\nmodel [\\[3\\]](#ref-amodei2016concrete). Two main attributes of AI safety\nare safe exploration and robustness to distributional\nshift [\\[3\\]](#ref-amodei2016concrete). If one of these attributes is\nlacking, the model will be vulnerable to different kinds of attacks\nagainst their privacy and security.\n\nSeveral studies have indicated that machine learning models are\nsusceptible to attacks on their privacy and security at different stages\nof their lifecycles. Since these models are trained on vast amounts of\ndata, Model training is sometimes outsourced, or pre-trained models are\nobtained from untrusted sources, which makes them vulnerable to attacks\nduring the training phase. Additionally, machine learning models\nprovided as a service contain valuable information about their training\ndata and hyperparameters, making them attractive targets for attacks\nduring the test and deployment phases [\\[4\\]](#ref-xue2020machine).\nAttacks on privacy commonly aim to compromise the confidentiality of\nvarious machine learning model components, while security attacks target\ntheir integrity and availability.\n\n## 1.1 Privacy and Security attacks overview\n\nDifferent attacks are possible depending on the model architecture and\nattackers’ capabilities. In terms of security, attackers may aim to gain\naccess to the model, steal its information or disrupt its normal\nfunctioning. For example, an attacker may target a spam detector to make\nit unable to classify spam correctly. This can be executed by poisoning\ntraining data during training [\\[5\\]](#ref-barreno2010security).\n\nThe attacks can be categorized into three main categories: black-box,\npartial white-box, and white-box attacks, according to the attacker’s\nknowledge of the machine learning\nmodel [\\[6\\]](#ref-pitropakis2019taxonomy). In black-box attacks, the\nattacker has no information about the training dataset or model’s\narchitecture; in white-box attacks, the attacker has full access to the\nmodel and all information about it, including the training dataset,\nmodel parameters, model architecture, prediction vectors, etc. Partial\nwhite-box attacks stand between these extremes, meaning the attacker has\nsome information about the model’s architecture and training data\ndistribution. The most common security and privacy attacks against\nmachine learning models are described in this section.\n\n-   **Membership Inference**: The membership inference is an attack in\n    which an attacker attempts to determine if a particular data sample,\n    $x$, is part of a model $M$’s training dataset. This attack is often\n    carried out using black-box techniques to query the model. Different\n    querying techniques are used to optimize the attack to gain more\n    information about the membership of individual records in the\n    training dataset. One of the first Membership inference attacks,\n    implemented by [\\[7\\]](#ref-shokri2017membership), was performed on\n    Google and Amazon’s APIs that provide machine learning as a\n    service(MLaaS). They could achieve high accuracy in guessing the\n    membership of their queried data points in the target models’\n    training dataset.  [\\[8\\]](#ref-quan2022amplification) showed that\n    having additional knowledge about the model or training dataset\n    distribution can improve the attack’s success rate. For example,\n    suppose the attacker knows the model was trained on an image dataset\n    with ten classes. When querying the target model, they can use this\n    information to balance their queries among all classes.\n\n-   **Model Inversion**  \n    In a model inversion attack, the adversary tries to get information\n    from the target model to reconstruct some representation of its\n    input dataset. The first category aims to generate an actual data\n    reconstruction [\\[9\\]](#ref-he2019model). In contrast, the second\n    group of attacks tries to create class representatives or probable\n    values of sensitive features that may not belong to the training\n    dataset [\\[10\\]](#ref-fredrikson2014privacy). Several attacks have\n    been implemented based on different assumptions. Some attacks\n    assumed to have information about data and sensitive features, and\n    some had query access to the model to get a prediction for an input\n    $X$. Two main categories of this attack are performed.\n\n-   **Property Inference**: A property inference attack attempts to\n    deduce information about the characteristics of a training dataset\n    that are not explicitly represented in the features. Revealing such\n    properties can result in privacy breaches since they may be\n    considered confidential. Property inference attacks are designed to\n    identify dataset-wide properties [\\[11\\]](#ref-ateniese2015hacking)\n    or detect common characteristics among a subset of the training\n    data [\\[12\\]](#ref-melis2019exploiting). For instance, a\n    classification model may be trained to distinguish between criminals\n    and non-criminals. An attacker can estimate the proportion of men\n    and women in the dataset by conducting a property inference attack.\n    This information was not meant to be disclosed and was learned by\n    the model unintentionally.\n\n-   **Model Extraction**: In a model extraction\n    attack [\\[13\\]](#ref-tramer2016stealing), the attacker aims to\n    develop a model replicating the target model’s behavior. This is\n    typically done when the attacker lacks information about the target\n    model’s architecture and training dataset. To achieve this, the\n    attacker generates a training set for the attack model by sending\n    queries to the target model and uses the predictions generated by\n    the target model as labels of its data points. A successful model\n    extraction attack, as shown by [\\[13\\]](#ref-tramer2016stealing),\n    allows the attacker to produce a model that can be used for\n    inference or to extract information about the training dataset.\n    However, if the attacker has some knowledge about the training set’s\n    distribution, the attack’s success rate can be improved. Selecting\n    the data to query the target model is a critical aspect of the\n    attack methodology, which significantly impacts the accuracy and\n    fidelity of the attack model.\n\n-   **Poisoning**: In certain situations, a machine learning model\n    designer may not generate or thoroughly examine the data used to\n    train the model. Generally, this may occur when data generation is\n    outsourced to third parties, or pre-trained models are fine-tuned\n    for a specific task. In such cases, the model becomes vulnerable to\n    poisoning attacks [\\[14\\]](#ref-dalvi2004adversarial) that target\n    the training dataset. In a poisoning attack, the attacker injects\n    malicious data into the training dataset, causing the model to learn\n    patterns unrelated to the classification task specified for the\n    model. These patterns can be exploited during the inference phase as\n    a backdoor, resulting in incorrect decisions being made by the model\n    when presented with data containing these malicious patterns.\n\n-   **Evasion**: Evasion attacks [\\[15\\]](#ref-huang2011adversarial)\n    occur when an adversary attempts to cause a machine learning model\n    to misclassify a data sample during the inference phase. This attack\n    typically occurs when the model is trained and deployed. The\n    attacker aims to generate a data sample similar to the original ones\n    but misclassified by maximizing a loss function based on their\n    attack objectives. There are two main categories of evasion attacks:\n    targeted and untargeted. In a targeted attack, the adversary wants\n    the manipulated sample to be classified as a specific class. On the\n    other hand, in an untargeted attack, the attacker is not concerned\n    with which class the manipulated input is classified in; Their\n    objective is to misclassify the input. The manipulated data samples\n    created in evasion attacks are known as adversarial examples in the\n    literature. Several techniques, such as projected gradient descent,\n    have been developed to generate adversarial\n    examples [\\[16\\]](#ref-deng2020universal). A typical example of an\n    evasion attack is in spam detection, where a spammer creates an\n    email that the spam detector cannot classify correctly.\n\n-   **Manipulation** It is critical for the model’s users to understand\n    how it makes its decisions. Manipulation attacks occur when the\n    attacker tries to explain the model decisions based on some criteria\n    that are not made. Here, explanations help users understand complex\n    models’ behavior (explanations are described in section\n    <a href=\"#explainability\" data-reference-type=\"ref\"\n    data-reference=\"explainability\">2</a>). But to gain the users’\n    trust, it is not enough to have an explanation; it should be robust\n    too. Manipulation attacks are the type of attacks that exploit the\n    fragility of the explanations. It means that while the model makes\n    decisions based on some criteria, if manipulated, explanations can\n    pretend the model is using other, more reasonable\n    measures [\\[17\\]](#ref-ghorbani2019interpretation). This type of\n    attack is performed to achieve different goals. One of its main\n    objectives is farewashing [\\[18\\]](#ref-aivodji2019fairwashing),\n    which means the model designer has manipulated the explanation\n    methods to hide the unfairness of their model’s decisions (Fairness\n    is described in detail in section\n    <a href=\"#bias-fairness\" data-reference-type=\"ref\"\n    data-reference=\"bias-fairness\">3</a>).\n\n## 1.2 An Overview of Defense Techniques\n\nIn response to the security and privacy attacks, some researchers have\nproposed defense strategies for the attack they discovered. Some of the\nmost common defense techniques are described here. Differential privacy\nis one of the most popular defenses against privacy attacks on machine\nlearning models [\\[19\\]](#ref-dwork2014algorithmic). Differential\nprivacy guarantees that no more privacy risks will be introduced to the\ndata after being used to train the model compared to when they are not\nused for this purpose. One of the main formalizations of differential\nprivacy is $\\epsilon$-differential privacy, which means that if two\nmodels are training on the same datasets ($D,D'$) that are only\ndifferent in one record, for every $S$ in the domain of the model:\n\n$$Pr(M(D)\\in S) \\leq exp(\\epsilon) Pr(M(D')\\in S)$$\n\nAnother technique usually used to prevent privacy attacks is\nk-anonymity [\\[20\\]](#ref-sweeney2002achieving) which guarantees that\neach data instance is indistinguishable among at least $k-1$ other data\npoints. As a countermeasure to the evasion attacks, adversarial\ntraining [\\[21\\]](#ref-goodfellow2014explaining) has been suggested to\nmake the model robust by exposing it to some adversarial examples in the\ntraining phase, allowing it to learn correct behavior against them.\n\n# 2 Transparency and Explainability\n\nAs mentioned in section\n<a href=\"#chap:privacy-security\" data-reference-type=\"ref\"\ndata-reference=\"chap:privacy-security\">1</a>, it is necessary for AI\nusers to understand how the model works and makes its decisions.\nExplainability, a crucial aspect of machine learning, bridges the gap\nbetween predictive accuracy and human understanding. Explainability has\nemerged as a critical aspect of machine learning models, addressing the\nneed to understand the reasoning behind their decisions. While\npredictive accuracy has long been the primary focus, the growing demand\nfor transparency in models has motivated academic research in\ninterpreting black-box neural networks. Explanations can potentially\nallow the deployment of machine learning models while maintaining high\nethical standards.\n\nIn various areas of AI applications, the need to have explanations for\nAI model decisions has raised and played an important role in motivating\nresearchers to focus on making complicated models explainable.\n\nIn the context of medical research, the use of an explainable model for\nscreening patients goes beyond simply identifying high-risk individuals\naccurately. It is equally important for the researcher to understand the\nunderlying causes of diseases like cancer. By employing an explainable\nmodel, the researcher gains insights into the factors and features that\ncontribute to the prediction, the relationship between various risk\nfactors and the identification of significant biomarkers, genetic\npredispositions, lifestyle factors, or environmental exposures that\ncontribute to a diagnosis. There is an urgent need for explanation in\nthe finance section. When a loan application is rejected or a credit\nscore is calculated, auditors, regulators, and customers need to\nunderstand the factors contributing to these outcomes. The transparency\noffered by interpretability helps stakeholders identify biases, errors,\nor potential discriminatory practices in the decision-making process and\nidentify the specific features or variables that significantly influence\nthe risk profiles. Additionally, interpretability aids in fraud\ndetection by allowing investigators to understand how the AI models\nidentify and flag suspicious transactions. It empowers auditors,\nregulators, and customers to understand and validate AI-driven\ndecisions, leading to a more accountable and fair financial ecosystem.\n\n#### Emergent Behaviours\n\nAlphaGo [\\[22\\]](#ref-silver2016mastering), a neural network model\ntrained to play the game GO, played a move known as \"move 37\" that\ndefied conventional wisdom and went against traditional Go strategies\nthat had been developed over centuries of human gameplay. Initially,\nmany human players and experts were perplexed by the move, as it seemed\nto disregard established principles of Go. However, as the game\nprogressed, it became clear that AlphaGo had a deeper understanding of\nthe game and could see the potential of unconventional moves to human\nplayers. Move 37 played by AlphaGo turned out to be a pivotal move in\nthe game, setting the stage for a highly complex and strategic battle.\nUltimately, AlphaGo won the game, showcasing its ability to challenge\nand surpass human intuition in Go. Move 37 and similar cases emphasize\nthe need to explore and develop techniques for explaining AI\ndecision-making processes.\n\n#### For-profit Business\n\nExplainability enhances the understanding of customer preferences and\nbehavior. By providing transparent explanations, businesses can gain\nvaluable insights into customer interactions with AI systems, enabling\nthem to personalize offerings, improve customer experience, and tailor\ntheir products and services to better meet customer needs. This\npersonalized approach can lead to increased customer satisfaction and\nloyalty. It further enables businesses to effectively communicate with\ncustomers about the usage of AI systems. By providing transparent\nexplanations, businesses can educate customers on how AI-driven\ndecisions are made, alleviating concerns, and fostering a better\nunderstanding of the technology. This communication builds trust and\nstrengthens customer relationships.\n\n#### Regulations\n\nThe rapid advancement of AI technologies has led to their widespread\nadoption across various domains. Regulators and policymakers recognize\nthe need for mechanisms to shed light on AI models’ inner workings,\nenabling stakeholders to understand and assess the justifiability of\nAI-driven decisions. Regulatory bodies across industries have placed a\nstrong emphasis on transparency, fairness, and non-discrimination when\nit comes to the use of AI systems. Explanations play a pivotal role in\nmeeting these regulatory requirements by providing a means to elucidate\nhow AI models arrive at their predictions or decisions. Moreover, legal\nand ethical concerns surrounding AI technologies necessitate the\nintegration of explanations into AI systems.\n\n## 2.1 Stakeholders in XAI\n\nMost experts agree that explainability and/or interpretability are\ncrucial for artificial intelligence (AI) and machine learning systems.\nHowever, there isn’t a universal understanding of what “explainable” and\n“interpretable” mean. As a result, analyzing the opinion of stakeholder\ncommunities surrounding explainable AI is of great importance. The\nmajority of stakeholder related debates merely distinguish between end\nusers and system developers. This can be seen in\n[\\[23\\]](#ref-mcdermid_artificial_2021) through the following points\n\n-   Prediction-recipients who are directly impacted by the ML-based\n    system’s predictions despite not using it themselves since the\n    prediction is typically mediated by an expert user.\n\n-   End users who utilize the ML-based system directly, and are\n    consequently directly impacted by it.\n\n-   Expert users who directly use the ML-based system though they are\n    not immediately impacted by its forecasts. Since they could be held\n    responsible (both legally and ethically) for the results of\n    predictions put into action, they are indirectly impacted.\n\n-   Attorneys and the courts are interested in determining who is\n    responsible for damage caused by an ML-based system.\n\n-   The Financial Services Authority, the Vehicle Certification Agency,\n    and the Medical and Healthcare Products Regulation Agency are\n    regulatory bodies that are not the system’s immediate users nor its\n    direct beneficiaries. They, however, safeguard the interests of\n    prediction recipients and end users [\\[24\\]](#ref-burton_mind_2020).\n\nA compressed version of the latter was suggested by\n [\\[25\\]](#ref-hoffman_stakeholder_2021), where the authors show a\n“Users Chart” of stakeholder groups as developed by the Defense Advanced\nResearch Projects Agency (DARPA). They summarized some “Sensemaking\nNeeds” and “Explanation Requirements” for some Stakeholder Groups (such\nas Policy Makers and Regulators, Developers: AI Experts, test Operators,\nand lastly Operations: Military, Legal, and Transportation).\n\n## 2.2 Properties of Explanations\n\nExplanations, in order to be able to respond to their stakeholders’\nneeds, should involve several properties:\n\n#### Clarity\n\nrefers to the quality of understanding the explanation. It involves\npresenting explanations in a clear, concise, and interpretable manner. A\nclear explanation avoids unnecessary jargon or technical complexity,\nmaking it accessible to the intended stakeholders. It should be\nstructured and organized to facilitate understanding and promote\neffective communication of underlying concepts or factors.\n\n#### Fidelity\n\nis a crucial property of explanations, emphasizing the importance of\nproviding correct and truthful information. An accurate explanation\naligns with the underlying data or model, ensuring that the explanation\nreflects the actual reasons and factors influencing a decision.\nExplanations must not misrepresent or distort the information but rather\nprovide a faithful representation of the relevant aspects of the data.\n\n#### Completeness\n\nrelates to the extent to which an explanation provides a comprehensive\naccount of the relevant factors or features contributing to a decision\nor result. A complete explanation includes all the necessary\ninformation, leaving no significant gaps or missing components. It\nshould cover both the primary factors and any secondary or indirect\nfactors that may have influenced the outcome. A complete explanation\nhelps to avoid ambiguity or misunderstanding by providing a holistic\nview of the situation.\n\n#### Consistency\n\nemphasizes an explanation’s coherence and logical consistency. A\nconsistent explanation should not contain contradictory statements or\nconflicting information. It should maintain a coherent narrative that\naligns with the underlying data or model. Consistency ensures that the\nexplanation is internally coherent and does not introduce confusion or\nambiguity in the interpretation of the provided information.\n\n#### Causality\n\nexplores the causal relationships between variables or factors. A causal\nexplanation seeks to identify and explain the cause-effect relationships\nthat lead to a particular outcome or prediction. It provides insights\ninto the mechanisms and processes that drive the observed phenomena.\nCausal explanations help to uncover the underlying reasons behind a\ndecision or result, shedding light on the factors that directly or\nindirectly influence the outcome.\n\n#### Transperent\n\nrelates to the openness and accessibility of an explanation. A\ntransparent explanation is understandable, interpretable, and accessible\nto the intended audience. It avoids unnecessary complexity or\nobfuscation and allows individuals to examine and verify the reasoning\nbehind a decision or outcome. Transparency promotes trust,\naccountability, and scrutiny, enabling individuals to assess the\nreliability and fairness of the provided explanation.\n\n#### Contextuality\n\nconsiders an explanation’s relevance and contextual appropriateness. A\ncontextual explanation takes into account the specific circumstances,\nbackground knowledge, and contextual factors that may impact the\ninterpretation and understanding of the explanation. It adapts the level\nof detail, language, or content to suit the specific context or\naudience, ensuring that the explanation is meaningful and relevant\nwithin the given context.\n\n#### Granular\n\nexplanations can vary in their level of granularity. They can range from\nhigh-level summaries to detailed explanations at the feature or instance\nlevel. The granularity of explanations should align with the users’\nneeds and their level of understanding, striking a balance between\nsimplicity and depth.\n\n#### User-Centric\n\nemphasizes the importance of tailoring explanations to the needs,\npreferences, and cognitive abilities of the intended users. A\nuser-centric explanation is designed to effectively communicate\ninformation to the target audience, taking into account their background\nknowledge, expertise, and information processing capabilities. It\nconsiders the user’s perspective and provides explanations adapted to\ntheir specific requirements and level of understanding. User-centric\nexplanations enhance the usability and utility of the provided\ninformation.\n\n## 2.3 Explainability Techniques\n\n#### Simple models\n\nA subset of algorithms that inherently produce interpretable models is a\nstraightforward approach to achieving interpretability.\n\n1.  Linear regression [\\[26\\]](#ref-hastie2009elements) is a widely used\n    algorithm for modeling the relationship between a dependent variable\n    and one or more independent variables. The model assumes a linear\n    relationship and provides interpretable coefficients that represent\n    the impact of each independent variable on the dependent variable.\n\n2.  Logistic regression is an extension of linear regression used for\n    binary classification problems. It models the probability of a\n    binary outcome based on the values of input variables. The\n    coefficients in logistic regression indicate the strength and\n    direction of the relationship between the input variables and the\n    binary outcome. Various extensions of linear regression, such as\n    ridge regression, lasso regression, and elastic net regression,\n    offer improved performance and interpretability. These methods\n    introduce regularization techniques to handle multicollinearity and\n    prevent overfitting.\n\n3.  Decision trees are intuitive and interpretable models that make\n    predictions by recursively splitting the data based on different\n    input features. Each split represents a decision point, and the\n    resulting tree structure can be easily understood and visualized.\n    Decision trees can also capture non-linear relationships between\n    variables.\n\n4.  Decision rules [\\[27\\]](#ref-holte1993very) are concise,\n    human-readable statements that describe how the input variables\n    contribute to the prediction. They are often extracted from decision\n    trees or learned directly using algorithms. Decision rules provide a\n    transparent representation of the decision-making process.\n\n5.  The RuleFit algorithm combines the strengths of decision trees and\n    linear regression. It first builds a decision tree ensemble to\n    capture complex interactions between variables. Then it fits a\n    linear model using the original input features and the leaf node\n    indicators from the decision tree ensemble. The resulting model\n    incorporates both non-linear patterns and interpretable linear\n    components.\n\nWhile these models offer interpretability, it is important to note that\ntheir simplicity and transparency come at the cost of potential\nlimitations in capturing complex relationships and handling\nhigh-dimensional data. Additionally, the choice of model depends on the\nspecific problem, data characteristics, and the desired level of\ninterpretability.\n\n#### Global Model Agnostic Methods\n\n1.  Partial Dependence plots [\\[28\\]](#ref-friedman2001greedy) visualize\n    the average effect of a single feature on the model’s predictions\n    while holding other features constant. They provide a global\n    understanding of how changes in a specific feature influence the\n    model’s output.\n\n2.  Accumulated local effects plots [\\[29\\]](#ref-apley2020visualizing)\n    capture the accumulated effects of a feature by integrating the\n    local effects of changing the feature values over its entire range.\n    They are particularly useful when features are dependent on each\n    other.\n\n3.  Feature interaction [\\[30\\]](#ref-friedman2008predictive) quantifies\n    the extent to which the joint effects of multiple features influence\n    the prediction. It measures the interaction between features and\n    provides insights into how they interact to contribute to the\n    model’s predictions.\n\n4.  Functional Decomposition [\\[31\\]](#ref-hooker2004discovering) is a\n    fundamental concept in interpretability. It involves breaking down\n    the complex prediction function of a model into smaller, more\n    interpretable parts. This technique helps understand the\n    contributions of different components or sub-functions to the\n    overall prediction.\n\n5.  Permutation feature importance [\\[32\\]](#ref-fisher2019all) measures\n    the importance of a feature by randomly permuting its values and\n    evaluating the increase in loss. It assesses the impact of each\n    feature on the model’s performance and helps identify the most\n    influential features.\n\n6.  Global surrogate models replace the original complex model with a\n    simpler, interpretable model that approximates its behavior. These\n    surrogate models are trained to mimic the predictions of the\n    black-box model and provide insights into its decision-making\n    process.\n\n#### Local Model Agnostic Methods\n\n1.  Individual Conditional\n    Expectation [\\[33\\]](#ref-goldstein2015peeking) curves provide a\n    fine-grained view of how changing a specific feature affects the\n    model’s prediction for an individual instance. Unlike PDP, which\n    shows the average effect, ICE curves show the predicted outcome for\n    each instance as the feature value varies. They offer a more\n    detailed understanding of the relationship between a feature and the\n    prediction at an individual level.\n\n2.  Local Interpretable Model-agnostic Explanations\n    (LIME) [\\[34\\]](#ref-ribeiro2016should) is a method that explains\n    individual predictions by approximating the complex model locally\n    with a simpler, interpretable model. It generates a surrogate model\n    that is more easily explainable and uses it to understand the\n    reasoning behind the prediction for a specific instance. LIME\n    provides insights into how different features contribute to the\n    prediction for that instance.\n\n3.  Scoped rules, also known as\n    anchors [\\[35\\]](#ref-ribeiro2018anchors), are rule-based\n    explanations that identify specific feature values that anchor or\n    lock a prediction in place. These rules define conditions under\n    which the model’s prediction remains unchanged, providing\n    interpretable criteria for decision-making. Anchors help in\n    understanding the important factors that drive a particular\n    prediction.\n\n4.  Counterfactual explanations [\\[36\\]](#ref-mothilal2020explaining)\n    explore what changes in the feature values would be required to\n    achieve a desired prediction outcome. Counterfactual explanations\n    help understand the model’s decision boundaries and provide insights\n    into how different features impact the predicted outcome by\n    identifying the necessary modifications to the features.\n\n5.  Shapely values [\\[37\\]](#ref-sundararajan2020many) are an\n    attribution method that fairly allocates the prediction value to\n    individual features. Inspired by cooperative game theory, Shapley\n    values assign contributions to each feature based on its impact on\n    predictions when combined with other features. They provide a more\n    nuanced understanding of the feature’s importance in the model’s\n    predictions.\n\n6.  SHAP [\\[38\\]](#ref-lundberg2017unified) is a computation method for\n    Shapley values that combines the individual feature contributions to\n    explain the model’s predictions. It not only provides insights at\n    the feature level but also proposes global interpretation methods by\n    considering combinations of Shapley values across the data. SHAP\n    allows for a comprehensive understanding of the model’s behavior and\n    feature importance in a broader context.\n\n## 2.4 Large Language Models and Explainability\n\nIn a rising number of different tasks, large language models (LLMs) have\nshown outstanding prediction performance. However, these models have\nbecome known as “black boxes” due to their inability to be understood\nclearly. Due to their opacity, they are no longer useful in high-risk\nfields like medicine and policymaking. Explainability is essential for\nLLMs since it enables users to comprehend how the model generates its\npredictions [\\[39\\]](#ref-singh_explaining_nodate),\n[\\[40\\]](#ref-peng2023check).\n\nIn this line,  [\\[41\\]](#ref-cifka_black-box_2023) developed a\nmodel-agnostic explanation technique based on tracking the model’s\npredictions as a function of the number of context tokens available for\ncausal (autoregressive) language models. Each time a new token is\nintroduced, there is an increment in context length, and the authors\nproposed a metric, *Differential Importance Score* to quantify this\nchange. These scores appear to have the ability to find long-range\ndependencies (LRDs), which is particularly intriguing because they are\nintended to highlight information not already covered by shorter\ncontexts, unlike attention maps, for instance.\n\nDeep neural networks are trained to recognize very particular structural\nand perceptual attributes of inputs. Techniques for locating neurons\nthat react to certain idea categories, such as textures, are readily\navailable in computer vision. Nevertheless, the scope of these methods\nis constrained, labeling only a tiny portion of the neurons and\nbehaviors in every network. To solve this, Hernandez E., et\nal [\\[42\\]](#ref-hernandez2021natural) proposed **MILAN**\n(**m**utual-**i**nformation-guided **l**inguistic **a**nnotation of\n**n**eurons) which generates descriptions (that capture categorical,\nrelational, and logical structure in learned features) of neuron\nbehavior in vision models using patch-level information about visual\ncharacteristics.\n\nAnother approach to black box language models is through text modules.\nSingh C., et al  [\\[39\\]](#ref-singh_explaining_nodate) introduced the\n<u>S</u>ummarize <u>a</u>nd <u>S</u>core (SASC) approach that takes a\ntext module as input and outputs a natural language explanation of the\nmodule’s selectivity coupled with a reliability score. They show better\ninterpretability for LLMs may be attained by the SASC, which can enhance\nautomated analysis of LLM submodules such as attention heads. Bills S.,\net al  [\\[43\\]](#ref-bills2023language) offer a SASC-like technique for\nexplaining individual neurons in an LLM by forecasting token-level\nneuron activations. They used an automated approach to solve the issue\nof scaling an interpretability approach to each neuron in an LLM which\nis expected to assess the trustworthiness of the models before the\ndeployment. The method clarifies how textual patterns trigger neuron\nactivation through; Explaining neuron’s activation using GPT-4,\nSimulating activations conditioned on the explanation, and Scoring the\nexplanation  [\\[44\\]](#ref-singh2022explaining),\n[\\[45\\]](#ref-zhong2022describing).\n\n# 3 Fairness and Equity\n\nAs discussed in sections\n<a href=\"#chap:privacy-security\" data-reference-type=\"ref\"\ndata-reference=\"chap:privacy-security\">1</a> and\n<a href=\"#explainability\" data-reference-type=\"ref\"\ndata-reference=\"explainability\">2</a>, Adding transparency to the AI\nsystems, makes different stakeholders equipped to find the system’s\ndeviations from desired behavior. Fairness is one of the main\nrequirements in which many AI systems fail. Fairness can be defined as\nthe absence of any prejudice or favoritism towards an individual or a\ngroup of individuals based on their inherent acquired characteristics\nsuch as race, gender, religion, etc. There are numerous examples of AI\napplications that exhibited unwanted discriminatory behaviors. Alarming\nfor the need to take action to overcome the potential bias that might be\nembedded in AI systems.\n\n#### The Compas Software\n\nOne of the most prominent examples is the *Correctional Offender\nManagement Profiling for Alternative Sanctions* (Compas) system. A\nsoftware used in the US to assess the recidivism risk of defendants.\nJulia Angwin investigated[1] the software and showed that compared to\nwhite defendants, black defendants are predicted as twice as likely to\nre-offend although they do no subsequent offenses; While white\ndefendants with subsequent offenses are predicted as low risk compared\nto their black counterparts. In other words, the software had a higher\nfalse positive rate for black defendants and a higher false negative\nrate for white defendants.\n\n#### Amazone gender bias hiring system\n\nAnother example is the AI-based hiring system[2] used by Amazon for\nassessing the resumes of job applicants. It was observed that the\nevaluations assigned to applicants’ resumes exhibited gender bias. Women\nreceived lower scores for technical positions compared to their male\ncompared. In fact, due to the underrepresentation of women in technical\njob positions, the software inherently linked male applicants to higher\nscores. Although gender information was not explicitly used in the\nresumes, the system was able to infer it via women’s related information\nin the resume, such as \\*women’s chess club captain\\* or attendance to\nall women’s colleges.\n\n#### Google Translate gender bias\n\nGender bias in machine translation was shown[3] when translating text\nbetween gender-neutral languages such as Turkish and other languages\nsuch as English and French. When translating a sentence, it is observed\nthat the system often assigns male pronouns (he, him) to terms like\nDoctor and Software. In contrast, female pronouns are predominantly\nassigned to words like Nurse or Hometaker.\n\n#### Twitter chatbot Tay\n\nMicrosoft launched a Twitter chatbot[4] named *Tay* that can answer\ntweets from users. The bot was trained on tweets crawled from the web,\nand the goal was to experiment with the potential conversational AI.\nAcross initial interactions, the chatbot started throwing racist,\nantisemite, or sexist content. The tool was turned off after a few\nhours, and this drama served as a good example of the importance of\navoiding training AI models with uncurated content from the web.\n\n#### Bias in the facial recognition system\n\nFace recognition algorithms claim to achieve high classification\naccuracy, surpassing 90%. However, these results do not apply uniformly.\nA growing body of research reveals significant variations[5] in error\nrates among different demographic groups, particularly in the accuracy\nof female individuals, Black individuals, and those aged 18-30. In a\nsignificant project called \"Gender Shades\" conducted\nby [\\[46\\]](#ref-buolamwini2018gender), researchers employed an\nintersectional approach to evaluate gender classification algorithms,\nincluding those developed by IBM and Microsoft. Participants were\ncategorized into four groups: darker-skinned females, darker-skinned\nmales, lighter-skinned females, and lighter-skinned males. Remarkably,\nall three algorithms performed the poorest in identifying darker-skinned\nfemales, with error rates up to 34% higher than lighter-skinned males.\nThese findings have been further supported by the National Institute of\nStandards and Technology (NIST), which independently assessed 189 face\nrecognition algorithms and confirmed that the technologies are least\naccurate when identifying women of color.\n\nThese non-exhaustive examples of AI bias have triggered the need for\ndeveloping more inclusive AI models in order to make their use more\nsocially acceptable. A biased AI system does not only have a negative\nimpact on the end-users but can the organization deploying the system\ncan suffer reputation damage, users distrust, and juridical liability.\nInitial important steps to mitigate these issues focused on\nmathematically defining and quantifying bias AI systems\n\n## 3.1 Definitions of Bias and Fainess\n\nThe concept of fairness has a variety of definitions depending on the\ndomain\nconsidered[\\[47\\]](#ref-shaheen_ai_2021)–[\\[49\\]](#ref-laurim2021computer).\nIt can be defined according to *political philosophy, areas of\neducation, in the legal domain and according to the general public’s\nperception* [\\[50\\]](#ref-mehrabi_survey_2022).unfairness or\ndiscrimination is divided into two categories;\n\n-   Disparate treatment: Which is defined as “ intentionally treating\n    someone differently based on his/her membership in a protected class\n    (direct discrimination)” [\\[51\\]](#ref-pessach_review_nodate)\n\n-   Disparate impact: which is “ negatively affecting members of a\n    protected class more than others even if by a seemingly neutral\n    policy (indirect discrimination)”\n    [\\[51\\]](#ref-pessach_review_nodate)\n\n<center>\n<figure id=\"fig:disparate_treatment\">\n<img src=\"https://www.linkpicture.com/q/Disparate-Treatment-3.png\"\nstyle=\"width:520px\" />\n<figcaption>Figure 2: Disparate Treatment vs Disparate Impact <a\nhref=\"https://www.qualtrics.com/experience-management/employee/disparate-treatment/\">[Source]</a></figcaption>\n</figure>\n</center>\n\n\nFigure <a href=\"#fig:disparate_treatment\" data-reference-type=\"ref\"\ndata-reference=\"fig:disparate_treatment\">2</a> shows the difference\nbetween disparate impact and disparate treatment. In fact, AI systems\ncan exhibit *disparate treatment* if the model heavily relies on\nsensitive attributes to make predictions. However, in general,\ndiscriminatory outcomes of AI systems are not intended, but due to\ndifferent sources of bias, the system will provide *disparate outcome*s\nover different demographic groups considered.\nTable <a href=\"#tab:disparate_treatment\" data-reference-type=\"ref\"\ndata-reference=\"tab:disparate_treatment\">1</a> shows examples of\ndecisions making processes that exhibit disparate impact and disparate\ntreatment.\n\n| Types of Discrimination | Disparate Treatment                                                                                                                                                                                                                                    | Disparate Impact                                                                                                                                                 |\n|:------------------------|:-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|:-----------------------------------------------------------------------------------------------------------------------------------------------------------------|\n| Age                     | Laying off employees over the age of 40 in favor of retaining younger employees                                                                                                                                                                        | Ninety percent of the sales agents subjected to the hiring freeze initiative were 40 years of age or older                                                       |\n| Sex                     | Refusal to accommodate pregnancy-related lifting restriction for one employee. At the same time, accommodating the restrictions of other non-pregnant employees who were injured on the job and who were similar in their ability or inability to work | Conducting strength testing as a requirement for workers to be hired for various jobs, causing an unlawful discriminatory impact on female workers seeking jobs. |\n| Race                    | Employees are subject to a hostile work environment and different treatment based on their race/nationality                                                                                                                                            | A broad criminal background check that negatively impacts a particular racial group, even though it is not a business necessity                                  |\n\nTable 1: Examples of decision-making processes with disparate treatment and\nDisparate\nTreatment [\\[Source\\]](https://www.qualtrics.com/experience-management/employee/disparate-treatment/)\n\n## 3.2 Source of bias and Fairness notions in AI\n\nUnfairness in machine learning originates from three main sources of\nbiases: biases due to the data, those from the algorithm, and user\ninteraction. Machine learning systems and artificial intelligent systems\nare data-driven since they rely on data in order to be trained, making\nthem an integral part of the system. As a result, if the algorithm is\ntrained on a biased dataset, then these biases are likely to be\nportrayed in the model’s outcome.\n\n#### Source of bias\n\nThere are a wide variety of sources of bias in data and some important\nones are highlighted here as reported by\n[\\[50\\]](#ref-mehrabi_survey_2022);\n\n-   O**mitted-Variable Bias** (OVB): OVB occurs when the dataset fails\n    to incorporate one or more relevant variables/features. The authors\n    in [\\[52\\]](#ref-wilms_omitted_2021) showed that in a model which\n    explains the relationship between dependent and independent\n    variables, omitting a relevant variable leads biased estimates. They\n    also showed that OVB leads to statistical relationships that can be\n    indicated as larger, smaller, or opposite to their actual value,\n    which inflates error rates.\n\n-   **Measurement Bias**. Also called Reporting or Recall bias, which is\n    a result of how important features are measured\n    [\\[50\\]](#ref-mehrabi_survey_2022).\n\n-   **Aggregation Bias** occurs when it is erroneously believed that\n    individual data points follow the trends found in aggregated data.\n    Aggregation bias frequently happens in research because it is\n    sometimes assumed incorrectly that patterns that exist at an\n    aggregate level must also appear at an individual level. Sadly, as\n    the preceding illustration showed, this is not always true. A\n    study’s results may derive incorrect conclusions due to aggregate\n    bias, which is misleading. This kind of bias is especially damaging\n    when it comes to the correlations between different variables.\n\n-   **Representation Bias** occurs when certain segments of the target\n    population are underrepresented in the training data and,\n    consequently, don’t generalize well. Data representation bias may be\n    due to biases introduced after the data was obtained, either\n    historically, cognitively, or statistically, or it may result from\n    how (and where) the data was initially\n    collected [\\[53\\]](#ref-shahbazi2023representation). Selection bias\n    can cause representation bias, which occurs when just a small\n    percentage of the population is sampled, the population of interest\n    has changed, or the population of interest differs from the\n    population used to train the model. For instance, if a poll\n    measuring the illegal drug use of teenagers only includes high\n    school students and leaves out homeschooled children or dropouts, it\n    may be biased [\\[54\\]](#ref-suresh2021framework). The skewness of\n    the underlying distribution is another possible explanation for\n    representation bias. Let’s say that adults aged 18 to 60 years old\n    are the target demographic for a specific medical dataset. Within\n    this community, there are minority groups; for instance, pregnant\n    women may constitute only 5% of the target population. Because the\n    model has fewer data points to learn from for the group of pregnant\n    people, it is susceptible to being less robust even with perfect\n    sampling and an identical\n    population [\\[54\\]](#ref-suresh2021framework).\n\n-   **Algorithmic Bias**. Another common source of bias can be the\n    algorithm itself. The machine learning model, trained on a biased\n    dataset, can reproduce and amplify the biases in the model’s output.\n    Even if trained using an unbiased dataset, machine learning\n    algorithms throughout their architecture have the ability to\n    demonstrate biased behavior [\\[50\\]](#ref-mehrabi_survey_2022). It\n    arises solely as a result of the design characteristics and model\n    architecture, such as the choice of the regularizer, and loss\n    functions.\n\nHaving identified the origins of bias in the AI lifecycle, auditing AI\nsystems for biases assessment requires metrics to quantify them and to\nevaluate the efficiency of intervention methods. In this regard, various\nfairness metrics (definitions) have been defined to capture different\naspects of fairness.\n\n#### Fairness definitions\n\nFairness notions can be categorized into group, individual and subgroup\ntypes [\\[55\\]](#ref-alves2023survey). Group fairness suggests that\ndifferent groups are treated equally. In its widest sense, group\nfairness splits a population into groups defined by *protected\nattributes/features* (such as gender, religion, caste) and desires some\nstatistical quantities to be equal across different groups.\n[\\[50\\]](#ref-mehrabi_survey_2022), [\\[56\\]](#ref-weerts2023fairlearn)\ndiscuss a wide variety of group fairness metrics and the following are\nthe most significant ones:\n\n-   **Demographic Parity**: Demographic parity seeks to guarantee that a\n    model’s predictions are not related to one’s belonging to a\n    vulnerable group. Demographic parity refers to equal selection rates\n    for each group in the binary classification\n    scenario [\\[55\\]](#ref-alves2023survey),\n    [\\[57\\]](#ref-pagano_bias_2023). Equal selection, for instance, in\n    the context of a resume screening approach, would imply that the\n    proportion of candidates chosen for a job interview should be the\n    same across groups.\n\n-   **Equalized Odds**: the Equalized Odds aims to guarantee that a\n    machine learning model works equally effectively for all groups. It\n    is more stringent than demographic parity because it demands that\n    groups have the same true positive and false positive rates as well\n    as independent predictions from the machine learning model\n    regardless of membership in sensitive groups\n    [\\[57\\]](#ref-pagano_bias_2023),\n    [\\[58\\]](#ref-birzhandi_application_2023). This distinction is\n    crucial because even if a model achieves demographic parity (i.e.,\n    its predictions are not dependent on a subject’s membership in a\n    sensitive group), it may nevertheless provide more false positive\n    predictions for a particular group.\n\n-   **Equal Opportunity**: this can be understood as requiring that both\n    protected and unprotected group members have an equal chance of\n    being allocated to a positive result if they belong to a positive\n    class [\\[50\\]](#ref-mehrabi_survey_2022). In other words, the equal\n    opportunity definition states that the true positive rates for\n    protected and unprotected groups should be equal.\n\n-   **Disparate Impact (DI)**: this notion can also be viewed as the\n    ratio between the two groups’ rates of accurate predictions and so a\n    high value of the ratio guarantees that the percentage of accurate\n    predictions is consistent across groups. Nevertheless, one of the\n    major drawbacks of *disparate impact* and *demographic parity* is\n    that a perfectly accurate classifier may be viewed as being unfair\n    when the proportion of real positive outcomes of the various groups\n    are noticeably different [\\[51\\]](#ref-pessach_review_nodate).\n\n-   **Individual Fairness**: In contrast to group fairness, individual\n    fairness is focused on how each individual is\n    treated [\\[59\\]](#ref-li2021tutorial). This notion requires similar\n    individuals to receive similar outcomes from the model. Individual\n    fairness is beneficial because it is a highly specific way of\n    defining fairness and also because people tend to care more about\n    individuals than large\n    groups [\\[60\\]](#ref-barocas-hardt-narayanan).\n\n    **Fairness Through Unawareness**: This notion states that a model is\n    fair as long it is not trained using the sensitive\n    attributes [\\[50\\]](#ref-mehrabi_survey_2022). However, a\n    significant weakness of this notion is its failure to consider\n    non-sensitive features that may correlate with sensitive ones. When\n    the model uses these non-sensitive features as proxies for the\n    unused sensitive features, it can result in discriminatory outcomes.\n\n-   **Counterfactual fairness**: it concerns the root causes of\n    differences. A sensitive trait would be replaced in practice,\n    affecting everything that occurred due to that sensitive feature\n    down the line. In the hiring scenario, one would alter a sensitive\n    attribute, such as race, if counterfactual fairness were applied. As\n    a result, subsequent events, like education, could alter. Based on\n    the counterfactual, a classifier might decide whether to hire the\n    candidate.\n\n## 3.3 Fairness enhancing methods\n\nFairness-enhancing methods are grouped into three main categories based\non the stage of the pipeline where the fairness constraint is enforced,\ni.e., at the data level before training the model(pre-processing\ntechniques), during the model training (in-processing techniques), or\nafter training the model (post-processing techniques). The\ntable <a href=\"#tab:sota_fairness\" data-reference-type=\"ref\"\ndata-reference=\"tab:sota_fairness\">2</a> showcases most state-of-the-art\nfairness-enhancing techniques with their open source implementation.\n\n### Pre-processing techniques\n\nA model that relies on sensitive attributes (e.g., gender, race,\nnationality) to make predictions can lead to discrimination or unfair\nresults. Pre-processing techniques are used to remove the influence of\nsensitive attributes from the data before training the model. The main\nadvantage of these techniques is that they are model-agnostic. The\ntransformed dataset or representation learned can then be used in\ndownstream tasks (classification, regression, etc) without any change to\nprovide \"fairer\" outcomes. We group approaches to mitigate biases at the\ndata level into three main categories:\n\n-   Fair representation learning: learn a fair representation of the\n    data that obfuscates information about the sensitive\n    attributes [\\[61\\]](#ref-zemel2013learning),\n    [\\[62\\]](#ref-lowd2005adversarial).\n\n-   Dataset transformation: Modify the training data by relabeling or\n    reweighing data points [\\[63\\]](#ref-kamiran2012data).\n\n-   Sampling: find a distribution close to the empirical distribution of\n    the dataset subject to fairness\n    constraints [\\[63\\]](#ref-kamiran2012data).\n\n \n\n### In-processing techniques\n\nIn-processing techniques are used when we have access to the model\ntraining, and it is not costly to retrain an existing model. In a\nnutshell, the loss function is transformed to add a loss/regularization\nterm that penalizes the model’s disparities across groups. Therefore,\nthe model is forced to optimize for accuracy and fairness. The\nclassification problem becomes a constrained optimization problem where\nthe goal is to minimize the classification error (maximize the accuracy)\nwhile satisfying a given fairness constraint. However, this optimization\nproblem is nonconvex and difficult to enforce. Therefore, existing\nin-processing techniques are reformulated in different ways or dual\nproblems are solved. They can be grouped as follows:\n\n#### Reduction approach\n\nThe Exponential Gradient approach [\\[64\\]](#ref-agarwal2018reductions)\nfor fairness transforms any binary classification problem into a\ncost-sensitive classification problem, that can yield a randomized\nclassifier having the lowest error while satisfying fairness\nconstraints.\n\n#### Adversarial-based approach\n\nA popular application of adversarial learning to enforce fairness\nconstraints to a model is adversarial debiasing. Adversarial\ndebiasing [\\[65\\]](#ref-zhang2018mitigating) is an adversarial-based\napproach that enforces the independence between the classifier outcome\nand the sensitive attributes.\n\n#### Regularization-based approach\n\nRegularization is generally used in ML to prevent the model from\noverfitting on the training data, by penalizing the model’s weights\nusing $L_1$ or $L_2$ norms [\\[66\\]](#ref-kamishima2011fairness). In the\nauthors introduced *prejudice remover regularizer*, a regularization\nterm for fairness that minimizes the mutual information between the\nmodel’s output and the sensitive attributes.\n\n#### Sapmling-based approach\n\nFairBatch [\\[67\\]](#ref-roh2020fairbatch) is a batch selection process\nthat enforces a given fairness metric by sampling mini-batches in a way\nto transform the Empirical Risk Minimization problem into a weighted EMR\nto incorporate fairness constraints.\n\n### Post-processing techniques\n\nFairness-enhancing post-processing techniques involve methods that treat\nthe model as a black box and enforce fairness constraints over the\nmodel’s output.\n\nMost existing post-processing methods consist of post hoc modification\nof the model’s outputs in order to satisfy a given fairness metric. In\nparticular, [\\[68\\]](#ref-hardt2016equality) formalized an optimization\nproblem over the model’s output to derive a classifier that satisfies\nthe fairness constraint while minimizing the classification loss. The\nderived classifier depends on four parameters that measure the\nprobability of positive outcomes given the current classifier output and\nthe sensitive attribute. The optimization problem is thus defined as a\nconstrained linear optimization problem. When the model output is\ncontinuous (a score function) the derived classifier is based on a\nthreshold of each demographic group such that it maximizes the\nclassification loss while satisfying fairness constraints, i.e., equal\nopportunity and equalized odds. Similar methods are proposed in the\nliterature, and they differ mainly in the way the optimization problem\nis defined.\n\n| Algorithms                                                    | Pre-process | In-process | Post-process | Implementation                                                                                                                                                     |\n|:--------------------------------------------------------------|:-----------:|:----------:|:------------:|:-------------------------------------------------------------------------------------------------------------------------------------------------------------------|\n| Threshold Optimizer [\\[68\\]](#ref-hardt2016equality)          |   $\\circ$   |  $\\circ$   |  $\\bullet$   | [ThresholdOptimizer](https://fairlearn.org/v0.8/api_reference/fairlearn.postprocessing.html)                                                                       |\n| Exponentiated Gradient [\\[64\\]](#ref-agarwal2018reductions)   |   $\\circ$   | $\\bullet$  |   $\\circ$    | [ExponentiatedGradient](https://fairlearn.org/v0.8/api_reference/fairlearn.postprocessing.html)                                                                    |\n| Learning fair representation [\\[61\\]](#ref-zemel2013learning) |  $\\bullet$  |  $\\circ$   |   $\\circ$    | [LFR](https://aif360.readthedocs.io/en/latest/modules/generated/aif360.algorithms.preprocessing.LFR.html)                                                          |\n| Adversarial debaising [\\[65\\]](#ref-zhang2018mitigating)      |  $\\bullet$  |  $\\circ$   |   $\\circ$    | [AdversarialFairnessClassifier](https://fairlearn.org/v0.8/api_reference/fairlearn.adversarial.html)                                                               |\n| Calibrated fairness [\\[69\\]](#ref-pleiss2017fairness)         |   $\\circ$   |  $\\circ$   |  $\\bullet$   | [CalibratedPostprocessing](https://aif360.readthedocs.io/en/latest/modules/generated/aif360.algorithms.postprocessing.CalibratedEqOddsPostprocessing.html)         |\n| FairBatch [\\[67\\]](#ref-roh2020fairbatch)                     |   $\\circ$   | $\\bullet$  |   $\\circ$    | [FairBatch](https://github.com/yuji-roh/fairbatch)                                                                                                                 |\n| Reweighing[\\[63\\]](#ref-kamiran2012data)                      |  $\\bullet$  |  $\\circ$   |   $\\circ$    | [Reweighing](https://aif360.readthedocs.io/en/latest/modules/generated/aif360.algorithms.preprocessing.Reweighing.html#aif360.algorithms.preprocessing.Reweighing) |\n| Prejudice Remover [\\[66\\]](#ref-kamishima2011fairness)        |   $\\circ$   | $\\bullet$  |   $\\circ$    | [PrejudiceRemover](https://aif360.readthedocs.io/en/latest/modules/algorithms.html#aif360-algorithms-inprocessing)                                                 |\n\nTable 2: State-of-the-art fairness enhancing techniques.\n\n### Pre-In-Post process: where should we enforce fairness?\n\nEach group of fairness-enhancing methods has its pros and cons. There\nare different settings where they can be applied and settings where\ntheir use is more challenging or not possible. We summarize the\nadvantages and disadvantages of each group of methods as follows:\n\n-   Pre-processing methods can work with any type of classifier and\n    machine learning tasks. As the fairness intervention is done at the\n    data level, the downstream task can be of any type, however, it\n    becomes difficult to control the tradeoffs, and the algorithmic bias\n    that might arise in the downstream task is not controlled.\n\n-   In-processing methods allow control over the fairness-accuracy\n    tradeoff that the model can achieve. Having access to the\n    optimization problem with fairness constraints provides more\n    flexibility in the tradeoffs, however, there is little flexibility\n    over the type of models used, i.e., the constraint optimization is\n    model specific.\n\n-   Similarly to pre-processing techniques, post-processing methods can\n    be applied to any type of model (classifier), which is treated like\n    a black box. The output of the model is modified to satisfy a given\n    fairness metric. However, changing the model’s output comes at a\n    significant cost of accuracy. Moreover, these methods can yield\n    unfair outcomes against certain individuals as the model output is\n    changed to satisfy a certain fairness metric.\n\nOverall, as shown by Friedler, Sorelle A., et\nal. [\\[70\\]](#ref-friedler2019comparative) there is no consensus in the\nliterature about which group of methods performs best. None of the\nmethods consistently outperforms others, and their performances depend\non the fairness metric and datasets.\n\n# 4 Responsibility, Accountability, and Regulations\n\nAs it was demonstrated in previous sections, AI systems are generally\ncomplex \"black-boxes\" sociotechnical systems that can sometimes produce\nunintended outcomes, which have raised issues of accountability\nestablishment when something goes wrong. Moreover, understanding how new\ntechnologies intertwine with the upcoming AI regulations and how they\nreflect on AI ethics could go a long way. In this section, we delve into\nthe ethical considerations of AI and the current landscape of AI\nregulation.\n\n## 4.1 Accountability\n\nWe have seen in previous sections that AI systems can sometimes go\nwrong, for example, the Twitter bot throwing hateful and racist content,\ngender bias in the Amazon hiring system, or adversarial environments\nthat mislead a deep neural network model. In general, when something\ngoes wrong in an organization or a company, someone is held accountable\nfor it. However, the AI system itself cannot take responsible for\nunpredicted outcomes. A question that naturally arises when AI systems\nprovide outcomes that have adverse effects or harm individuals, is who\nis accountable for the issues. Accountability is a well-studied notion\nin different fields (politic, law). It provides moral principles that\nguide the ethical conduct of people or organizations as they bear the\nresponsibility for their actions. Accountability is, therefore, a key\ncomponent of trust in society, organizations, and the professional\nmilieu. In the context of AI, accountability is a meta-component of\ntrustworthiness that ensure that ethical principles are promoted and\nenforced throughout the lifecycle of AI projects.\n\n#### Accountable for what?\n\nAccording to Virginia Dignum accountability means the decisions of the\nAI system are explainable and derivable from the decision-making\nmechanisms used [\\[71\\]](#ref-dignum2019responsible) and is a set of\ncomponents guided by moral values that are part of a large\nsocio-technical system. [\\[72\\]](#ref-millar2018accountability) see\naccountability as \"answerability\" for decisions, actions, products, and\npolicies. From this perspective, it is required that accountability is a\npractice that the management board of a company could enforce by making\nall the stakeholders who develop the system understand that they bear\nthe responsibility for the decisions of the system. Therefore,\npolicymakers, data scientists, and AI engineers should carefully choose\nand justify the choices made during the design, development, and\ndeployment of AI systems [\\[73\\]](#ref-novelli2023accountability).\n\n#### Accountable to whom?\n\nThe design, development, and deployment of AI systems involve multiple\nstakeholders. While a system in which decisions can be explained can\nhelp to detect the causes or reasons for the incidents, it does not\nnecessarily provide an answer to who is most responsible for the\nunintended outcomes. Establishing causal accountability remains\nchallenging, and unintended negative outcomes might bear legal\nliability. Companies or organizations can acknowledge incidents caused\nby the use of their systems and can pay for reparation. However,\nmonetary compensation could not be enough in society and requires\nsomeone to get punished, i.e., to take legal responsibility for the\npotential harm. [\\[74\\]](#ref-ammanath2022trustworthy) says that instead\nof looking for people to blame, companies should recognize the legal\nresponsibility of sociotechnical systems around AI and all stakeholders\nshould use it as an \"additional motivation to own their individual\nresponsibility\".\n\nPromoting accountability throughout the entire lifecycle of AI systems\nfaces the challenge of developing processes and rules to enforce it\nwhile not hindering innovation. The enforcement of accountability can\nalso be baked by laws and regulations that define a conformity\nassessment of the outcome of sociotechnical systems in order to hinder\npotential flaws of the system [\\[72\\]](#ref-millar2018accountability),\n[\\[74\\]](#ref-ammanath2022trustworthy). Accountability, therefore,\nbecomes a core component of trustworthiness in a socio-technical system.\nThat is, if a system provides discriminatory outcomes, leaks people’s\nsensitive data, or is not robust to different kinds of environments\nwhere it is deployed, someone would bear the responsibility of\njustifying and mitigating it.\n\n## 4.2 AI Ethics and Regulation\n\nTo better understand the AI regulation landscape, we talked with Matt\nHodges, a political technologist and the Executive Director of Zinc\nLabs. During the interview, when asked about how he sees data security\ncoming into play with political campaigns, he highlighted the\nsignificance of privacy and data security concerns about AI\napplications, particularly in political campaigns. He acknowledged that\nML had been employed in campaigns for a long time, whether through\nin-house teams, vendors, or consulting firms. However, he emphasized\nthat privacy and data security have become crucial with the increasing\nuse of AI tools. Hodges provided an example, stating that if a media or\npress department campaign worker uses AI tools to draft a press release,\nthere is a risk if those tools lack good data security and privacy\npolicies, such as an upcoming policy announcement affecting millions of\npeople, could leak if typed into an insecure online text box.\n\nHowever, we have initiatives around the world that focus on the\ndevelopment and promotion of trustworthy AI and the regulations around\nit. One exciting initiative to look at is the National AI Initiative,\none of the components of the United States’ approach to AI regulation.\nLooking at the National AI Initiative Act of\n2020 [\\[75\\]](#ref-NAIIA2020), it aims to promote and support research\nand development in trustworthy AI in both the public and private\nsectors. They also aim to develop technical standards and guidelines\nthat facilitate the evaluation of bias in artificial intelligence\ntraining data and applications. Despite many of these efforts, the\nUnited States does not have a comprehensive federal law governing AI.\nDifferent states or even problem statements have taken different\napproaches to AI regulation, with some enacting laws related to specific\napplications of AI, such as autonomous vehicles, and others focusing on\nbroader issues, such as data privacy and security. For autonomous\nvehicle regulation in the United States, the National Conference of\nState Legislatures (NCSL) provides a comprehensive database\n[\\[76\\]](#ref-NCSLAutonomus2023) that tracks autonomous vehicle bills\nintroduced in all 50 states and the District of Columbia. And for\nbroader issues such as data privacy and security, the National\nConference of State Legislatures (NCSL) documents state privacy laws in\nvarious areas [\\[77\\]](#ref-NCSLPrivacy2022)\n\nOn the other hand, the European Union (EU) has taken significant steps\nto regulate AI and protect data. General Data Protection Regulation\n(GDPR) [\\[78\\]](#ref-GDPR2018), enacted in May 2018, is a set of rules\ndesigned to give EU citizens more control over their personal data.\nElements of GDPR, like Data Minimisation, can play a huge role in what\nkind of data AI models can store and use for training. Additionally, the\nEuropean Commission introduced the Artificial Intelligence Act in April\n2021 [\\[79\\]](#ref-EUAIA2021) and actively working towards establishing\ncomprehensive AI regulations. \"The vote sets up the so-called trilogue\nnegotiations, the final phase of the EU’s process, and paves the way for\nthe likely adoption of Europe’s — and the world’s — first comprehensive\nAI regulatory framework in early 2024.\" [\\[80\\]](#ref-Morse2023)\n\nIn conclusion, AI’s ethical implications and regulations are crucial to\ntake into consideration. According to Matt Hodges, AI can empower\ntechnical staff and enhance productivity by giving them \"superpowers\"\nthrough advanced code generation. When asked what he finds exciting\nabout AI, he states, \"I’m actually pretty excited about the ability to\nstep up the game of technical staff here... The people that we do\nrecruit, if we can make them give them superpowers with this new code\ngeneration, then it actually probably makes it a little bit more likely\nthat they’re successful because they can get more done and do some more\ncomplicated tasks in less time.\"\n\n## 4.3 Generative AI Ownership\n\nFor a long time, creativity has been considered one of the most\ndistinguishing attributes of humans. Being able to create language, art,\nscience, and technology is what made humans the governors of our planet\nEarth. But with the advances in computer science and machine learning,\npeople tried to add creativity to the machine learning models, and it is\nwhere Generative AI appears. Regarding probabilistic modeling,\ngenerative models are machine learning models that describe generating a\ndataset. Sampling from such a model generates new data belonging to the\ndesired distribution [\\[81\\]](#ref-foster2022generative).\n\n### Who is the author?\n\nWhile current samples of generative models can create high-quality\nproducts like meaningful text [\\[82\\]](#ref-openai2023), different types\nof art, including paintings [\\[83\\]](#ref-Photoleap2023),\nmusic [\\[84\\]](#ref-soundraw2023), and executable computer codes, the\nfirst generative models which caught the news attention were\nStyleGAN [\\[85\\]](#ref-karras2019style) which created hyper-realistic\nimages from human faces and GPT-2 [\\[86\\]](#ref-radford2019language)\nthat could finish a paragraph of meaningful text related to its received\nopening sentence.\n\nWhile the number of generative AI users is increasing daily, one\nimportant question about generative AI ownership still needs to be\nanswered. Consider a situation in which an employee who is expected to\nadd some lines of code to their company’s product asks a generative\nmodel to create the code and add that code to the company’s product\ninstead of doing the task themselves. The question that arises here is\nwho owns the product? Does the company lose copyright over its product\nbecause a machine learning model generates one part? What about art\ngenerated by generative AI? In 2022, Jason Allen won the Colorado State\nFair’s art competition using an AI-generated\npicture [\\[87\\]](#ref-Harwell2022). He claimed he had not broken any\nrules. On the contrary, Boris Eldagsen rejected the prize he won at the\nSony world photography award, revealing that AI had generated the\nwinning photo [\\[88\\]](#ref-Grierson2023). As a prevention from such\nevents, the Grammys forbids AI-generated works from participating in\ntheir competitions [\\[89\\]](#ref-King2023). Debated about the ownership\nof Generative AI products remains unsolved, and people taking each side\nof this debate have satisfying reasons.  \n\n### Rules and Regulations\n\nSince the debate on generative AI ownership has not concluded. In\ncontrast, generative AI is being used more daily. Thus, it needs some\nregulation to decide who is responsible for and who owns the products\ngenerated by AI. Based on Federal Register\nguideline [\\[90\\]](#ref-federalregister2023), the work generated by AI,\nusing only one prompt by the human user, is unprotected. At the same\ntime, if it is a product of interaction between the human and machine\nwhere a human makes some modification, selection, or arrangement on the\nAI product, it is protected by copyright law. Regulations in the EU and\nUK take a different side toward AI-generated product ownership: a\nproduct can be subject to copyright if it is “the author’s own\nintellectual creation.” According to Copyright Designs and Patents Act\n1988 (CDPA) [\\[91\\]](#ref-CDPA1988), since computer programs cannot be\nconsidered as the owner of their product, “the person by whom the\narrangements necessary for the creation of the work are undertaken” is\nknown as the author of AI products.\n\nAs can be seen from different viewpoints of regulations toward\ngenerative AI ownership, there has yet to be a finalized committed\ndecision. Some consider AI models as an extension of their human users,\nand the human owns their products; others take model owners as the\ncreator of the model’s products, and the last group considers joint\nownership between creators and the model owners. Since the legal\nlandscape around AI ownership is still evolving, a long way exists to\nachieve commitment. But what are generative AI’s economic and social\nimpacts apart from the ownership of AI products? Based on AI ACT, when\nproducts like images, voice, art, etc., are generated by AI and\npublished, it should be exposed that it is AI-generated, especially when\nthey resemble existing persons, places, or\nevents [\\[92\\]](#ref-AIact2022). Otherwise, AI-generated products can\neasily deceive and manipulate people in various political, social, and\ncultural domains. Malicious use of AI to defame or create popularity for\nspecific individuals or groups is another important concern toward using\nGenerative AI. In this regard, AI Act added an account for\ngeneral-purpose AI that “may lead to discriminatory outcomes and the\nexclusion of certain groups.”\n\nProduct manufacturers or fashion designers can utilize AI-generated\ncontent to promote their products without hiring specialized\nindividuals. However, there is also the possibility of misusing\nindividuals’ identities in such cases, and it is necessary to consider\nthis aspect in the regulations related to artificial intelligence\nproducts. In this regard, another issue that arises is the economic\nimpact of using generative AI. Despite the significant concerns raised\nabout generative AI’s impact on the future of jobs, the economic\nconsequences of its use and the possibility of further economic\ninequality due to AI utilization exist. Since launching and utilizing\ngenerative AI requires high-level hardware capabilities and software\nknowledge, equal access to it is impossible for all countries and all\nindividuals within a society. This can significantly increase social and\neconomic inequality between developing and developed countries and\nbetween organizations and individuals with financial and informational\nresources and those without them.\n\nThe lack of consistent regulation for generative AI added to the\nconcerns about the harms unregulated AI can bring to\nsociety [\\[93\\]](#ref-bender2021dangers) has resulted in an open\nletter [\\[94\\]](#ref-futurefl2023) initiated by Future for Life and\nsigned by more than 3000 people in the technology industry, including\npioneer deep learning scientist, Yoshua Bengio [6], asking for a\nsix-month pause on giant AI experiments by all companies until\nsufficient policies are made and enforced. On the other hand, since the\nAI act is forcing more strict regulations for using generative AI, some\nEuropean companies claim that it may “jeopardize technological\nsovereignty” and signed an open letter [\\[95\\]](#ref-openletter2023)\nasking the EU reconsider its plans to let European companies participate\nin the advancement of AI.  \n\n# 5 Environmental Impact\n\nWhile machine learning algorithms are not inherently bad for the climate\nand environment, their implementation can have certain implications that\nneed to be considered. One significant factor is the high energy\nconsumption associated with training and running machine learning\nmodels, particularly when GPUs (Graphics Processing Units) are used for\naccelerated computation. From an environmental viewpoint, the following\naspects have to be considered.\n\n#### Energy Consumption and Carbon Emissions\n\nMachine learning algorithms, particularly deep learning models, require\nextensive computational resources for training. This process involves\nperforming numerous complex calculations on large datasets. Graphics\nProcessing Units (GPUs) are commonly used for their parallel processing\ncapabilities, which accelerate the training process. However, GPUs are\npower-hungry devices that consume substantial amounts of electricity.\nThe energy consumption of AI algorithms is a significant concern due to\nthe scale at which these algorithms are deployed. Training\nstate-of-the-art models can take weeks or even months, consuming a\nsubstantial amount of energy during that time. The energy consumption is\ndirectly proportional to the size and complexity of the model and the\namount of data being processed. As machine learning applications become\nmore prevalent across industries, the collective energy consumption\nassociated with training and running these algorithms\nincreases [\\[96\\]](#ref-sarswatula2022modeling).\n\n#### Carbon Emissions\n\nThe energy consumed by machine learning algorithms has direct\nimplications for carbon emissions. The majority of electricity\nproduction worldwide still relies on non-renewable energy sources, such\nas coal, natural gas, and oil. These fossil fuels release carbon dioxide\n(CO2) and other greenhouse gases when burned for energy generation,\ncontributing to climate change and global warming. Since machine\nlearning algorithms heavily depend on GPU usage, the electricity\nrequired to power these GPUs often comes from grids that rely on fossil\nfuel-based power plants. As a result, the carbon emissions associated\nwith electricity production for running machine learning infrastructure\ncan be significant. The carbon footprint of machine learning algorithms\nincludes both the direct emissions from power consumption and the\nindirect emissions from the fossil fuel extraction, production, and\ntransportation required to meet the energy\ndemands [\\[97\\]](#ref-anthony2020carbontracker).\n\n#### Electronic Waste and Disposal Challenges\n\nThe rapid advancement of machine learning technology necessitates\nfrequent hardware upgrades, including GPUs. As newer and more powerful\nGPUs are introduced to the market, older models become obsolete. This\ncycle of hardware replacement results in electronic waste generation.\nThe electronic waste consists of discarded electronic devices, including\nGPUs, that are no longer in use. Improper handling and disposal of\ne-waste pose significant environmental risks. Electronic devices contain\nhazardous materials, such as lead, mercury, cadmium, and brominated\nflame retardants, which can contaminate the environment if not managed\nproperly. E-waste often ends up in landfills, where toxic substances can\nleach into the soil and water, posing threats to ecosystems and human\nhealth.\n\nAddressing the e-waste challenge requires responsible management\npractices. This includes recycling electronic components to recover\nvaluable materials and properly disposing of hazardous waste. However,\nmanaging e-waste from machine learning infrastructure can be\nparticularly challenging due to the rapid obsolescence of hardware and\nthe need for specialized recycling processes to handle complex\nelectronic components [\\[98\\]](#ref-ewaste).\n\n## Using AI for mitigating Climate Change\n\nContrary to the potential harms of climate change due to AI, recent\nresearch [\\[99\\]](#ref-cowls2021ai) has gained interest in harnessing AI\nto combat climate change. By leveraging AI’s capabilities, climate\nmodeling and prediction, optimization of resource management and energy\nefficiency, advancement of renewable energy solutions, and enhancement\nof climate change adaptation and resilience can be improved. The\npotential of AI to address the pressing issues of climate change offers\na promising path toward sustainability and a greener future.\n\n#### Improved Climate Modeling and Prediction\n\nAI has revolutionized climate modeling by providing powerful tools to\nprocess vast amounts of data and make accurate predictions . Machine\nlearning algorithms can analyze historical climate data, satellite\nimagery, and environmental variables, unveiling complex climate system\ndynamics. This enhanced understanding enables us to anticipate extreme\nweather events, assess the impacts of human activities on the\nenvironment, and predict long-term climate trends. AI-driven climate\nmodeling empowers policymakers with valuable insights to develop\neffective mitigation and adaptation strategies, fostering a more\nsustainable future.\n\n#### Optimized Resource Management and Energy Efficiency\n\nAI has the potential to transform resource management and enhance energy\nefficiency across various sectors. By employing AI, especially RL\nalgorithms, smart grids can analyze real-time data on energy\nconsumption, demand, and weather conditions. This analysis facilitates\nefficient energy distribution, minimizes wastage, and encourages the\nintegration of renewable energy sources. Machine learning techniques can\noptimize energy-intensive processes in industries, transportation\nsystems, and buildings, resulting in significant energy savings.\n\n#### Advancements in Renewable Energy Solutions\n\nRenewable energy plays a pivotal role in mitigating climate change, and\nAI can accelerate its adoption. AI algorithms can optimize the\nintegration and operation of renewable energy systems like solar, wind,\nand hydroelectric power. By analyzing geographical and climatic data, AI\ncan identify optimal locations for solar panels and wind turbines,\nmaximizing energy output. Additionally, AI can aid in developing\nadvanced materials and technologies for efficient energy storage,\naddressing one of the primary challenges in renewable energy adoption.\nBy harnessing AI, the shift to a renewable energy-powered world reduces\ndependence on fossil fuels and mitigates climate change.  \n\nIt is important to note that efforts are being made to mitigate the\nenvironmental impacts of using AI. For example, some researchers are\nexploring energy-efficient training methods, optimizing algorithms to\nreduce computational requirements, and utilizing renewable energy\nsources for powering machine learning infrastructure. Additionally,\nadvancements in hardware design and the development of specialized chips\nfor AI tasks (TPUs) aim to improve energy efficiency.\n\nIn summary, to ensure the responsible use of machine learning\nalgorithms, stakeholders must consider the environmental implications\nand strive to adopt sustainable practices. This can include optimizing\nalgorithms for efficiency, using energy-efficient hardware, considering\nrenewable energy sources for computing infrastructure, and promoting\nresponsible e-waste management. By balancing technological advancements\nwith environmental considerations, we can harness the benefits of\nmachine learning while minimizing its potential negative impacts on the\nclimate and environment.\n\n# References\n\n<span class=\"csl-left-margin\">\\[1\\]\n</span><span class=\"csl-right-inline\">A. Jobin, M. Ienca, and E. Vayena,\n“The global landscape of AI ethics guidelines,” *Nature machine\nintelligence*, vol. 1, no. 9, pp. 389–399, 2019.</span>\n\n<span class=\"csl-left-margin\">\\[2\\]\n</span><span class=\"csl-right-inline\">K. R. Varshney, “Trustworthy\nmachine learning and artificial intelligence,” *XRDS: Crossroads, The\nACM Magazine for Students*, vol. 25, no. 3, pp. 26–29, 2019.</span>\n\n<span class=\"csl-left-margin\">\\[3\\]\n</span><span class=\"csl-right-inline\">D. Amodei, C. Olah, J. Steinhardt,\nP. Christiano, J. Schulman, and D. Mané, “Concrete problems in AI\nsafety,” *arXiv preprint arXiv:1606.06565*, 2016.</span>\n\n<span class=\"csl-left-margin\">\\[4\\]\n</span><span class=\"csl-right-inline\">M. Xue, C. Yuan, H. Wu, Y. Zhang,\nand W. Liu, “Machine learning security: Threats, countermeasures, and\nevaluations,” *IEEE Access*, vol. 8, pp. 74720–74742, 2020.</span>\n\n<span class=\"csl-left-margin\">\\[5\\]\n</span><span class=\"csl-right-inline\">M. Barreno, B. Nelson, A. D.\nJoseph, and J. D. Tygar, “The security of machine learning,” *Machine\nLearning*, vol. 81, pp. 121–148, 2010.</span>\n\n<span class=\"csl-left-margin\">\\[6\\]\n</span><span class=\"csl-right-inline\">N. Pitropakis, E. Panaousis, T.\nGiannetsos, E. Anastasiadis, and G. Loukas, “A taxonomy and survey of\nattacks against machine learning,” *Computer Science Review*, vol. 34,\np. 100199, 2019.</span>\n\n<span class=\"csl-left-margin\">\\[7\\]\n</span><span class=\"csl-right-inline\">R. Shokri, M. Stronati, C. Song,\nand V. Shmatikov, “Membership inference attacks against machine learning\nmodels,” in *2017 IEEE symposium on security and privacy (SP)*, 2017,\npp. 3–18.</span>\n\n<span class=\"csl-left-margin\">\\[8\\]\n</span><span class=\"csl-right-inline\">P. Quan, S. Chakraborty, J. V.\nJeyakumar, and M. Srivastava, “On the amplification of security and\nprivacy risks by post-hoc explanations in machine learning models,”\n*arXiv preprint arXiv:2206.14004*, 2022.</span>\n\n<span class=\"csl-left-margin\">\\[9\\]\n</span><span class=\"csl-right-inline\">Z. He, T. Zhang, and R. B. Lee,\n“Model inversion attacks against collaborative inference,” in\n*Proceedings of the 35th annual computer security applications\nconference*, 2019, pp. 148–162.</span>\n\n<span class=\"csl-left-margin\">\\[10\\]\n</span><span class=\"csl-right-inline\">M. Fredrikson, E. Lantz, S. Jha,\nS. Lin, D. Page, and T. Ristenpart, “Privacy in pharmacogenetics: An\nend-to-end case study of personalized warfarin dosing,” in *23rd\n$\\{$USENIX$\\}$ security symposium ($\\{$USENIX$\\}$ security 14)*, 2014,\npp. 17–32.</span>\n\n<span class=\"csl-left-margin\">\\[11\\]\n</span><span class=\"csl-right-inline\">G. Ateniese, L. V. Mancini, A.\nSpognardi, A. Villani, D. Vitali, and G. Felici, “Hacking smart machines\nwith smarter ones: How to extract meaningful data from machine learning\nclassifiers,” *International Journal of Security and Networks*, vol. 10,\nno. 3, pp. 137–150, 2015.</span>\n\n<span class=\"csl-left-margin\">\\[12\\]\n</span><span class=\"csl-right-inline\">L. Melis, C. Song, E. De\nCristofaro, and V. Shmatikov, “Exploiting unintended feature leakage in\ncollaborative learning,” in *2019 IEEE symposium on security and privacy\n(SP)*, 2019, pp. 691–706.</span>\n\n<span class=\"csl-left-margin\">\\[13\\]\n</span><span class=\"csl-right-inline\">F. Tramèr, F. Zhang, A. Juels, M.\nK. Reiter, and T. Ristenpart, “Stealing machine learning models via\nprediction APIs.” in *USENIX security symposium*, 2016, vol. 16, pp.\n601–618.</span>\n\n<span class=\"csl-left-margin\">\\[14\\]\n</span><span class=\"csl-right-inline\">N. Dalvi, P. Domingos, S. Sanghai,\nand D. Verma, “Adversarial classification,” in *Proceedings of the tenth\nACM SIGKDD international conference on knowledge discovery and data\nmining*, 2004, pp. 99–108.</span>\n\n<span class=\"csl-left-margin\">\\[15\\]\n</span><span class=\"csl-right-inline\">L. Huang, A. D. Joseph, B. Nelson,\nB. I. Rubinstein, and J. D. Tygar, “Adversarial machine learning,” in\n*Proceedings of the 4th ACM workshop on security and artificial\nintelligence*, 2011, pp. 43–58.</span>\n\n<span class=\"csl-left-margin\">\\[16\\]\n</span><span class=\"csl-right-inline\">Y. Deng and L. J. Karam,\n“Universal adversarial attack via enhanced projected gradient descent,”\nin *2020 IEEE international conference on image processing (ICIP)*,\n2020, pp. 1241–1245.</span>\n\n<span class=\"csl-left-margin\">\\[17\\]\n</span><span class=\"csl-right-inline\">A. Ghorbani, A. Abid, and J. Zou,\n“Interpretation of neural networks is fragile,” in *Proceedings of the\nAAAI conference on artificial intelligence*, 2019, vol. 33, pp.\n3681–3688.</span>\n\n<span class=\"csl-left-margin\">\\[18\\]\n</span><span class=\"csl-right-inline\">U. Aı̈vodji, H. Arai, O. Fortineau,\nS. Gambs, S. Hara, and A. Tapp, “Fairwashing: The risk of\nrationalization,” in *International conference on machine learning*,\n2019, pp. 161–170.</span>\n\n<span class=\"csl-left-margin\">\\[19\\]\n</span><span class=\"csl-right-inline\">C. Dwork, A. Roth, and others,\n“The algorithmic foundations of differential privacy,” *Foundations and\nTrends in Theoretical Computer Science*, vol. 9, no. 3–4, pp. 211–407,\n2014.</span>\n\n<span class=\"csl-left-margin\">\\[20\\]\n</span><span class=\"csl-right-inline\">L. Sweeney, “Achieving k-anonymity\nprivacy protection using generalization and suppression,” *International\nJournal of Uncertainty, Fuzziness and Knowledge-Based Systems*, vol. 10,\nno. 5, pp. 571–588, 2002.</span>\n\n<span class=\"csl-left-margin\">\\[21\\]\n</span><span class=\"csl-right-inline\">I. J. Goodfellow, J. Shlens, and\nC. Szegedy, “Explaining and harnessing adversarial examples,” *arXiv\npreprint arXiv:1412.6572*, 2014.</span>\n\n<span class=\"csl-left-margin\">\\[22\\]\n</span><span class=\"csl-right-inline\">D. Silver, A. Huang, C. J.\nMaddison, A. Guez, L. Sifre, G. Van Den Driessche, J. Schrittwieser, I.\nAntonoglou, V. Panneershelvam, M. Lanctot, and others, “Mastering the\ngame of go with deep neural networks and tree search,” *nature*, vol.\n529, no. 7587, pp. 484–489, 2016.</span>\n\n<span class=\"csl-left-margin\">\\[23\\]\n</span><span class=\"csl-right-inline\">J. A. McDermid, Y. Jia, Z. Porter,\nand I. Habli, “[Artificial intelligence explainability: The technical\nand ethical dimensions](https://doi.org/10.1098/rsta.2020.0363),”\n*Philosophical Transactions of the Royal Society A: Mathematical,\nPhysical and Engineering Sciences*, vol. 379, no. 2207, p. 20200363,\nOct. 2021.</span>\n\n<span class=\"csl-left-margin\">\\[24\\]\n</span><span class=\"csl-right-inline\">S. Burton, I. Habli, T. Lawton, J.\nMcDermid, P. Morgan, and Z. Porter, “[Mind the gaps: Assuring the safety\nof autonomous systems from an engineering, ethical, and legal\nperspective](https://doi.org/10.1016/j.artint.2019.103201),” *Artificial\nIntelligence*, vol. 279, p. 103201, Feb. 2020.</span>\n\n<span class=\"csl-left-margin\">\\[25\\]\n</span><span class=\"csl-right-inline\">R. Hoffman, G. Klein, S. T.\nMueller, M. Jalaeian, and C. Tate, “[The Stakeholder Playbook for\nExplaining AI Systems](https://doi.org/10.31234/osf.io/9pqez),”\nPsyArXiv, preprint, Nov. 2021.</span>\n\n<span class=\"csl-left-margin\">\\[26\\]\n</span><span class=\"csl-right-inline\">T. Hastie, R. Tibshirani, J. H.\nFriedman, and J. H. Friedman, *The elements of statistical learning:\nData mining, inference, and prediction*, vol. 2. Springer, 2009.</span>\n\n<span class=\"csl-left-margin\">\\[27\\]\n</span><span class=\"csl-right-inline\">R. C. Holte, “Very simple\nclassification rules perform well on most commonly used datasets,”\n*Machine learning*, vol. 11, pp. 63–90, 1993.</span>\n\n<span class=\"csl-left-margin\">\\[28\\]\n</span><span class=\"csl-right-inline\">J. H. Friedman, “Greedy function\napproximation: A gradient boosting machine,” *Annals of statistics*, pp.\n1189–1232, 2001.</span>\n\n<span class=\"csl-left-margin\">\\[29\\]\n</span><span class=\"csl-right-inline\">D. W. Apley and J. Zhu,\n“Visualizing the effects of predictor variables in black box supervised\nlearning models,” *Journal of the Royal Statistical Society Series B:\nStatistical Methodology*, vol. 82, no. 4, pp. 1059–1086, 2020.</span>\n\n<span class=\"csl-left-margin\">\\[30\\]\n</span><span class=\"csl-right-inline\">J. H. Friedman and B. E. Popescu,\n“Predictive learning via rule ensembles,” 2008.</span>\n\n<span class=\"csl-left-margin\">\\[31\\]\n</span><span class=\"csl-right-inline\">G. Hooker, “Discovering additive\nstructure in black box functions,” in *Proceedings of the tenth ACM\nSIGKDD international conference on knowledge discovery and data mining*,\n2004, pp. 575–580.</span>\n\n<span class=\"csl-left-margin\">\\[32\\]\n</span><span class=\"csl-right-inline\">A. Fisher, C. Rudin, and F.\nDominici, “All models are wrong, but many are useful: Learning a\nvariable’s importance by studying an entire class of prediction models\nsimultaneously.” *J. Mach. Learn. Res.*, vol. 20, no. 177, pp. 1–81,\n2019.</span>\n\n<span class=\"csl-left-margin\">\\[33\\]\n</span><span class=\"csl-right-inline\">A. Goldstein, A. Kapelner, J.\nBleich, and E. Pitkin, “Peeking inside the black box: Visualizing\nstatistical learning with plots of individual conditional expectation,”\n*journal of Computational and Graphical Statistics*, vol. 24, no. 1, pp.\n44–65, 2015.</span>\n\n<span class=\"csl-left-margin\">\\[34\\]\n</span><span class=\"csl-right-inline\">M. T. Ribeiro, S. Singh, and C.\nGuestrin, “\" why should i trust you?\" explaining the predictions of any\nclassifier,” in *Proceedings of the 22nd ACM SIGKDD international\nconference on knowledge discovery and data mining*, 2016, pp.\n1135–1144.</span>\n\n<span class=\"csl-left-margin\">\\[35\\]\n</span><span class=\"csl-right-inline\">M. T. Ribeiro, S. Singh, and C.\nGuestrin, “Anchors: High-precision model-agnostic explanations,” in\n*Proceedings of the AAAI conference on artificial intelligence*, 2018,\nvol. 32.</span>\n\n<span class=\"csl-left-margin\">\\[36\\]\n</span><span class=\"csl-right-inline\">R. K. Mothilal, A. Sharma, and C.\nTan, “Explaining machine learning classifiers through diverse\ncounterfactual explanations,” in *Proceedings of the 2020 conference on\nfairness, accountability, and transparency*, 2020, pp. 607–617.</span>\n\n<span class=\"csl-left-margin\">\\[37\\]\n</span><span class=\"csl-right-inline\">M. Sundararajan and A. Najmi, “The\nmany shapley values for model explanation,” in *International conference\non machine learning*, 2020, pp. 9269–9278.</span>\n\n<span class=\"csl-left-margin\">\\[38\\]\n</span><span class=\"csl-right-inline\">S. M. Lundberg and S.-I. Lee, “A\nunified approach to interpreting model predictions,” *Advances in neural\ninformation processing systems*, vol. 30, 2017.</span>\n\n<span class=\"csl-left-margin\">\\[39\\]\n</span><span class=\"csl-right-inline\">C. Singh, A. R. Hsu, R. Antonello,\nS. Jain, A. G. Huth, B. Yu, and J. Gao, “[Explaining black box text\nmodules in natural language with language\nmodels](https://doi.org/10.48550/arXiv.2304.11633).” arXiv,\nApr-2023.</span>\n\n<span class=\"csl-left-margin\">\\[40\\]\n</span><span class=\"csl-right-inline\">B. Peng, M. Galley, P. He, H.\nCheng, Y. Xie, Y. Hu, Q. Huang, L. Liden, Z. Yu, W. Chen, and J. Gao,\n“[Check your facts and try again: Improving large language models with\nexternal knowledge and automated\nfeedback](https://doi.org/10.48550/arXiv.2302.12813).” 2023.</span>\n\n<span class=\"csl-left-margin\">\\[41\\]\n</span><span class=\"csl-right-inline\">O. Cífka and A. Liutkus,\n“[Black-box language model explanation by context length\nprobing](https://doi.org/10.48550/arXiv.2212.14815).” arXiv,\nMay-2023.</span>\n\n<span class=\"csl-left-margin\">\\[42\\]\n</span><span class=\"csl-right-inline\">E. Hernandez, S. Schwettmann, D.\nBau, T. Bagashvili, A. Torralba, and J. Andreas, “[Natural language\ndescriptions of deep visual\nfeatures](https://doi.org/10.48550/arXiv.2201.11114),” in *International\nconference on learning representations*, 2021.</span>\n\n<span class=\"csl-left-margin\">\\[43\\]\n</span><span class=\"csl-right-inline\">S. Bills, N. Cammarata, D.\nMossing, H. Tillman, L. Gao, G. Goh, I. Sutskever, J. Leike, J. Wu, and\nW. Saunders, “[Language models can explain neurons in language\nmodels](https://openaipublic.blob.core.windows.net/neuron-explainer/paper/index.html),”\n*URL https://openaipublic. blob. core. windows.\nnet/neuron-explainer/paper/index. html.*, 2023.</span>\n\n<span class=\"csl-left-margin\">\\[44\\]\n</span><span class=\"csl-right-inline\">C. Singh, J. X. Morris, J. Aneja,\nA. M. Rush, and J. Gao, “[Explaining patterns in data with language\nmodels via interpretable\nautoprompting](https://doi.org/10.48550/arXiv.2210.01848),” *arXiv\npreprint arXiv:2210.01848*, 2022.</span>\n\n<span class=\"csl-left-margin\">\\[45\\]\n</span><span class=\"csl-right-inline\">R. Zhong, C. Snell, D. Klein, and\nJ. Steinhardt, “[Describing differences between text distributions with\nnatural language](https://doi.org/10.48550/arXiv.2201.12323),” in\n*International conference on machine learning*, 2022, pp.\n27099–27116.</span>\n\n<span class=\"csl-left-margin\">\\[46\\]\n</span><span class=\"csl-right-inline\">J. Buolamwini and T. Gebru,\n“Gender shades: Intersectional accuracy disparities in commercial gender\nclassification,” in *Conference on fairness, accountability and\ntransparency*, 2018, pp. 77–91.</span>\n\n<span class=\"csl-left-margin\">\\[47\\]\n</span><span class=\"csl-right-inline\">M. Y. Shaheen, “[AI in Healthcare:\nMedical and socio-economic benefits and\nchallenges](https://doi.org/10.14293/S2199-1006.1.SOR-.PPRQNI1.v1),”\npreprint, Sep. 2021.</span>\n\n<span class=\"csl-left-margin\">\\[48\\]\n</span><span class=\"csl-right-inline\">G. Creamer, G. Kazantsev, and T.\nAste, *[Machine learning and AI in\nfinance](https://arxiv.org/ftp/arxiv/papers/2111/2111.11295.pdf)*.\nRoutledge, 2021.</span>\n\n<span class=\"csl-left-margin\">\\[49\\]\n</span><span class=\"csl-right-inline\">V. Laurim, S. Arpaci, B.\nPrommegger, and H. Krcmar, “[Computer, whom should i hire?–acceptance\ncriteria for artificial intelligence in the recruitment\nprocess](https://doi.org/10.24251/HICSS.2021.668),” 2021.</span>\n\n<span class=\"csl-left-margin\">\\[50\\]\n</span><span class=\"csl-right-inline\">N. Mehrabi, F. Morstatter, N.\nSaxena, K. Lerman, and A. Galstyan, “[A Survey on Bias and Fairness in\nMachine Learning](https://doi.org/10.1145/3457607).” arXiv,\nJan-2022.</span>\n\n<span class=\"csl-left-margin\">\\[51\\]\n</span><span class=\"csl-right-inline\">D. Pessach and E. Shmueli, “[A\nReview on Fairness in Machine\nLearning](https://doi.org/10.1145/3494672),” *ACM Computing Surveys*,\nvol. 55, no. 3, 2022.</span>\n\n<span class=\"csl-left-margin\">\\[52\\]\n</span><span class=\"csl-right-inline\">R. Wilms, E. Mäthner, L. Winnen,\nand R. Lanwehr, “[Omitted variable bias: A threat to estimating causal\nrelationships](https://doi.org/10.1016/j.metip.2021.100075),” *Methods\nin Psychology*, vol. 5, p. 100075, Dec. 2021.</span>\n\n<span class=\"csl-left-margin\">\\[53\\]\n</span><span class=\"csl-right-inline\">N. Shahbazi, Y. Lin, A. Asudeh,\nand H. Jagadish, “[Representation bias in data: A survey on\nidentification and resolution\ntechniques](https://doi.org/10.1145/3588433),” *ACM Computing Surveys*,\n2023.</span>\n\n<span class=\"csl-left-margin\">\\[54\\]\n</span><span class=\"csl-right-inline\">H. Suresh and J. Guttag, “[A\nframework for understanding sources of harm throughout the machine\nlearning life cycle](https://doi.org/10.1145/3465416.3483305),” in\n*Equity and access in algorithms, mechanisms, and optimization*, 2021,\npp. 1–9.</span>\n\n<span class=\"csl-left-margin\">\\[55\\]\n</span><span class=\"csl-right-inline\">G. Alves, F. Bernier, M. Couceiro,\nK. Makhlouf, C. Palamidessi, and S. Zhioua, “[Survey on fairness notions\nand related tensions](https://doi.org/10.1016/j.ejdp.2023.100033),”\n*EURO Journal on Decision Processes*, p. 100033, 2023.</span>\n\n<span class=\"csl-left-margin\">\\[56\\]\n</span><span class=\"csl-right-inline\">H. Weerts, M. Dudı́k, R. Edgar, A.\nJalali, R. Lutz, and M. Madaio, “[Fairlearn: Assessing and improving\nfairness of AI systems](https://doi.org/10.48550/arXiv.2303.16626),”\n*arXiv preprint arXiv:2303.16626*, 2023.</span>\n\n<span class=\"csl-left-margin\">\\[57\\]\n</span><span class=\"csl-right-inline\">T. P. Pagano, R. B. Loureiro, F.\nV. N. Lisboa, R. M. Peixoto, G. A. S. Guimarães, G. O. R. Cruz, M. M.\nAraujo, L. L. Santos, M. A. S. Cruz, E. L. S. Oliveira, I. Winkler, and\nE. G. S. Nascimento, “[Bias and Unfairness in Machine Learning Models: A\nSystematic Review on Datasets, Tools, Fairness Metrics, and\nIdentification and Mitigation\nMethods](https://doi.org/10.3390/bdcc7010015),” *Big Data and Cognitive\nComputing*, vol. 7, no. 1, p. 15, Jan. 2023.</span>\n\n<span class=\"csl-left-margin\">\\[58\\]\n</span><span class=\"csl-right-inline\">P. Birzhandi and Y.-S. Cho,\n“[Application of fairness to healthcare, organizational justice, and\nfinance: A survey](https://doi.org/10.1016/j.eswa.2022.119465),” *Expert\nSystems with Applications*, vol. 216, p. 119465, Apr. 2023.</span>\n\n<span class=\"csl-left-margin\">\\[59\\]\n</span><span class=\"csl-right-inline\">Y. Li, Y. Ge, and Y. Zhang,\n“[Tutorial on fairness of machine learning in recommender\nsystems](https://doi.org/10.1145/3404835.3462814),” in *Proceedings of\nthe 44th international ACM SIGIR conference on research and development\nin information retrieval*, 2021, pp. 2654–2657.</span>\n\n<span class=\"csl-left-margin\">\\[60\\]\n</span><span class=\"csl-right-inline\">S. Barocas, M. Hardt, and A.\nNarayanan, *Fairness and machine learning*. fairmlbook.org, 2019.</span>\n\n<span class=\"csl-left-margin\">\\[61\\]\n</span><span class=\"csl-right-inline\">R. Zemel, Y. Wu, K. Swersky, T.\nPitassi, and C. Dwork, “Learning fair representations,” in\n*International conference on machine learning*, 2013, pp.\n325–333.</span>\n\n<span class=\"csl-left-margin\">\\[62\\]\n</span><span class=\"csl-right-inline\">D. Lowd and C. Meek, “Adversarial\nlearning,” in *Proceedings of the eleventh ACM SIGKDD international\nconference on knowledge discovery in data mining*, 2005, pp.\n641–647.</span>\n\n<span class=\"csl-left-margin\">\\[63\\]\n</span><span class=\"csl-right-inline\">F. Kamiran and T. Calders, “Data\npreprocessing techniques for classification without discrimination,”\n*Knowledge and information systems*, vol. 33, no. 1, pp. 1–33,\n2012.</span>\n\n<span class=\"csl-left-margin\">\\[64\\]\n</span><span class=\"csl-right-inline\">A. Agarwal, A. Beygelzimer, M.\nDudı́k, J. Langford, and H. Wallach, “A reductions approach to fair\nclassification,” in *International conference on machine learning*,\n2018, pp. 60–69.</span>\n\n<span class=\"csl-left-margin\">\\[65\\]\n</span><span class=\"csl-right-inline\">B. H. Zhang, B. Lemoine, and M.\nMitchell, “Mitigating unwanted biases with adversarial learning,” in\n*Proceedings of the 2018 AAAI/ACM conference on AI, ethics, and\nsociety*, 2018, pp. 335–340.</span>\n\n<span class=\"csl-left-margin\">\\[66\\]\n</span><span class=\"csl-right-inline\">T. Kamishima, S. Akaho, and J.\nSakuma, “Fairness-aware learning through regularization approach,” in\n*2011 IEEE 11th international conference on data mining workshops*,\n2011, pp. 643–650.</span>\n\n<span class=\"csl-left-margin\">\\[67\\]\n</span><span class=\"csl-right-inline\">Y. Roh, K. Lee, S. E. Whang, and\nC. Suh, “Fairbatch: Batch selection for model fairness,” *arXiv preprint\narXiv:2012.01696*, 2020.</span>\n\n<span class=\"csl-left-margin\">\\[68\\]\n</span><span class=\"csl-right-inline\">M. Hardt, E. Price, and N. Srebro,\n“Equality of opportunity in supervised learning,” *Advances in neural\ninformation processing systems*, vol. 29, 2016.</span>\n\n<span class=\"csl-left-margin\">\\[69\\]\n</span><span class=\"csl-right-inline\">G. Pleiss, M. Raghavan, F. Wu, J.\nKleinberg, and K. Q. Weinberger, “On fairness and calibration,”\n*Advances in neural information processing systems*, vol. 30,\n2017.</span>\n\n<span class=\"csl-left-margin\">\\[70\\]\n</span><span class=\"csl-right-inline\">S. A. Friedler, C. Scheidegger, S.\nVenkatasubramanian, S. Choudhary, E. P. Hamilton, and D. Roth, “A\ncomparative study of fairness-enhancing interventions in machine\nlearning,” in *Proceedings of the conference on fairness,\naccountability, and transparency*, 2019, pp. 329–338.</span>\n\n<span class=\"csl-left-margin\">\\[71\\]\n</span><span class=\"csl-right-inline\">V. Dignum, *Responsible artificial\nintelligence: How to develop and use AI in a responsible way*, vol.\n2156. Springer, 2019.</span>\n\n<span class=\"csl-left-margin\">\\[72\\]\n</span><span class=\"csl-right-inline\">J. Millar, B. Barron, K. Hori, R.\nFinlay, K. Kotsuki, and I. Kerr, “Accountability in AI. Promoting\ngreater social trust,” in *Conference on artificial intelligence,\nmontreal*, 2018.</span>\n\n<span class=\"csl-left-margin\">\\[73\\]\n</span><span class=\"csl-right-inline\">C. Novelli, M. Taddeo, and L.\nFloridi, “Accountability in artificial intelligence: What it is and how\nit works,” *AI & SOCIETY*, pp. 1–12, 2023.</span>\n\n<span class=\"csl-left-margin\">\\[74\\]\n</span><span class=\"csl-right-inline\">B. Ammanath, *Trustworthy AI: A\nbusiness guide for navigating trust and ethics in AI*. John Wiley &\nSons, 2022.</span>\n\n<span class=\"csl-left-margin\">\\[75\\]\n</span><span class=\"csl-right-inline\">“[National artificial intelligence\ninitiative act of\n2020](https://www.congress.gov/bill/116th-congress/house-bill/6395/text).”\nDivision E of Public Law 116-283, 2020.</span>\n\n<span class=\"csl-left-margin\">\\[76\\]\n</span><span class=\"csl-right-inline\">N. C. of State Legislatures,\n“[Autonomous vehicles state bill tracking\ndatabase](https://www.ncsl.org/transportation/autonomous-vehicles-state-bill-tracking-database).”\n2023.</span>\n\n<span class=\"csl-left-margin\">\\[77\\]\n</span><span class=\"csl-right-inline\">N. C. of State Legislatures,\n“[State laws related to digital\nprivacy](https://www.ncsl.org/technology-and-communication/state-laws-related-to-digital-privacy).”\n2022.</span>\n\n<span class=\"csl-left-margin\">\\[78\\]\n</span><span class=\"csl-right-inline\">“[General data protection\nregulation (GDPR)](https://eur-lex.europa.eu/eli/reg/2016/679/oj).”\nRegulation (EU) 2016/679 of the European Parliament and of the Council\nof 27 April 2016, 2018.</span>\n\n<span class=\"csl-left-margin\">\\[79\\]\n</span><span class=\"csl-right-inline\">E. Commission, “[Proposal for a\nregulation laying down harmonised rules on artificial intelligence\n(artificial intelligence act)](https://artificialintelligenceact.eu/).”\n2021.</span>\n\n<span class=\"csl-left-margin\">\\[80\\]\n</span><span class=\"csl-right-inline\">C. Morse, “[What the GDPR can\nteach us about AI\nregulation](https://www.weforum.org/agenda/2023/06/gdpr-artificial-intelligence-regulation-europe-us/).”\n2023.</span>\n\n<span class=\"csl-left-margin\">\\[81\\]\n</span><span class=\"csl-right-inline\">D. Foster, *Generative deep\nlearning*. \" O’Reilly Media, Inc.\", 2022.</span>\n\n<span class=\"csl-left-margin\">\\[82\\]\n</span><span class=\"csl-right-inline\">OpenAI,\n“[ChatGPT](\\url{https://openai.com/research/gpt-4}).” 2023.</span>\n\n<span class=\"csl-left-margin\">\\[83\\]\n</span><span class=\"csl-right-inline\">Photoleap,\n“[Midjourney](\\url{https://www.photoleapapp.com/lp/pl-text-to-image/}).”\n2023.</span>\n\n<span class=\"csl-left-margin\">\\[84\\]\n</span><span class=\"csl-right-inline\">S. Inc.,\n“[Soundraw](\\url{https://soundraw.io/}).” 2023.</span>\n\n<span class=\"csl-left-margin\">\\[85\\]\n</span><span class=\"csl-right-inline\">T. Karras, S. Laine, and T. Aila,\n“A style-based generator architecture for generative adversarial\nnetworks,” in *Proceedings of the IEEE/CVF conference on computer vision\nand pattern recognition*, 2019, pp. 4401–4410.</span>\n\n<span class=\"csl-left-margin\">\\[86\\]\n</span><span class=\"csl-right-inline\">A. Radford, J. Wu, R. Child, D.\nLuan, D. Amodei, I. Sutskever, and others, “Language models are\nunsupervised multitask learners,” *OpenAI blog*, vol. 1, no. 8, p. 9,\n2019.</span>\n\n<span class=\"csl-left-margin\">\\[87\\]\n</span><span class=\"csl-right-inline\">D. Harwell, “He used AI to win a\nfine-arts competition. Was it cheating?” *The Washington Post*, 2022.\n\\[Online\\]. Available:\n<https://www.washingtonpost.com/technology/2022/09/02/midjourney-artificial-intelligence-state-fair-colorado/>.\n\\[Accessed: 02-Sep-2022\\].</span>\n\n<span class=\"csl-left-margin\">\\[88\\]\n</span><span class=\"csl-right-inline\">J. Grierson, “Photographer admits\nprize-winning image was AI-generated,” *The Guardian*, 2023. \\[Online\\].\nAvailable:\n[https://www.theguardian.com/technology/2023/apr/17/photographer-admits-prize-winning-image-was-ai-generated#:~:text=The\\\\20German\\\\20artist\\\\20Boris\\\\20Eldagsen,generations\\\\20in\\\\20black\\\\20and\\\\20white.](https://www.theguardian.com/technology/2023/apr/17/photographer-admits-prize-winning-image-was-ai-generated#:~:text=The\\%20German\\%20artist\\%20Boris\\%20Eldagsen,generations\\%20in\\%20black\\%20and\\%20white.)\n\\[Accessed: 17-Apr-2023\\].</span>\n\n<span class=\"csl-left-margin\">\\[89\\]\n</span><span class=\"csl-right-inline\">A. King, “The grammys officially\nban AI-generated works — ‘only humans’ eligible for awards,” *Digital\nMusic News*, 2023. \\[Online\\]. Available:\n[https://www.digitalmusicnews.com/2023/06/19/the-grammys-ai-generated-content-\nineligible/](https://www.digitalmusicnews.com/2023/06/19/the-grammys-ai-generated-content- ineligible/).\n\\[Accessed: 19-Jun-2023\\].</span>\n\n<span class=\"csl-left-margin\">\\[90\\]\n</span><span class=\"csl-right-inline\">F. Register, “[Copyright\nregistration guidance: Works containing material generated by artificial\nintelligence](https://www.federalregister.gov/documents/2023/03/16/2023-05321/copyright-registration-guidance-works-containing-material-generated-by-artificial-intelligence),”\n*Federal Register*, vol. 88. pp. 16190–16194, 16-Mar-2023.</span>\n\n<span class=\"csl-left-margin\">\\[91\\]\n</span><span class=\"csl-right-inline\">F. Register, “[Copyright, designs\nand patents act\n1988](https://www.legislation.gov.uk/ukpga/1988/48/contents).”\n2023.</span>\n\n<span class=\"csl-left-margin\">\\[92\\]\n</span><span class=\"csl-right-inline\">C. of European Union, “[Proposal\nfor a REGULATION OF THE EUROPEAN PARLIAMENT AND OF THE COUNCIL LAYING\nDOWN HARMONISED RULES ON ARTIFICIAL INTELLIGENCE (ARTIFICIAL\nINTELLIGENCE ACT) AND AMENDING CERTAIN UNION LEGISLATIVE\nACTS](https://www.consilium.europa.eu/en/press/press-releases/2022/12/06/artificial-intelligence-act-council-calls-for-promoting-safe-ai-that-respects-fundamental-rights/).”\n2022.</span>\n\n<span class=\"csl-left-margin\">\\[93\\]\n</span><span class=\"csl-right-inline\">E. M. Bender, T. Gebru, A.\nMcMillan-Major, and S. Shmitchell, “On the dangers of stochastic\nparrots: Can language models be too big?” in *Proceedings of the 2021\nACM conference on fairness, accountability, and transparency*, 2021, pp.\n610–623.</span>\n\n<span class=\"csl-left-margin\">\\[94\\]\n</span><span class=\"csl-right-inline\">F. for Life, “[Pause giant AI\nexperiments: An open\nletter](https://futureoflife.org/open-letter/pause-giant-ai-experiments/).”\n2023.</span>\n\n<span class=\"csl-left-margin\">\\[95\\]\n</span><span class=\"csl-right-inline\">F. for Life, “[Open letter to the\nrepresentatives of the european commission, the european council and the\neuropean\nparliament](https://drive.google.com/file/d/1wrtxfvcD9FwfNfWGDL37Q6Nd8wBKXCkn/view?usp=sharing).”\n2023.</span>\n\n<span class=\"csl-left-margin\">\\[96\\]\n</span><span class=\"csl-right-inline\">S. A. Sarswatula, T. Pugh, and V.\nPrabhu, “Modeling energy consumption using machine learning,” *Frontiers\nin Manufacturing Technology*, vol. 2, p. 855208, 2022.</span>\n\n<span class=\"csl-left-margin\">\\[97\\]\n</span><span class=\"csl-right-inline\">L. F. W. Anthony, B. Kanding, and\nR. Selvan, “Carbontracker: Tracking and predicting the carbon footprint\nof training deep learning models,” *arXiv preprint arXiv:2007.03051*,\n2020.</span>\n\n<span class=\"csl-left-margin\">\\[98\\]\n</span><span class=\"csl-right-inline\">A. Miller, “The future of e-waste\ndepends on AI and machine learning,” 2022.</span>\n\n<span class=\"csl-left-margin\">\\[99\\]\n</span><span class=\"csl-right-inline\">J. Cowls, A. Tsamados, M. Taddeo,\nand L. Floridi, “The AI gambit: Leveraging artificial intelligence to\ncombat climate change—opportunities, challenges, and recommendations,”\n*Ai & Society*, pp. 1–25, 2021.</span>\n\n[1] https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing\n\n[2] https://www.reuters.com/article/us-amazon-com-jobs-automation-insight-idUSKCN1MK08G\n\n[3] https://blog.google/products/translate/reducing-gender-bias-google-translate/\n\n[4] (https://www.theverge.com/2016/3/24/11297050/tay-microsoft-chatbot-racist\n\n[5] https://sitn.hms.harvard.edu/flash/2020/racial-discrimination-in-face-recognition-technology/\n\n[6] https://yoshuabengio.org/","metadata":{}}]}