{"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":"# **We Are About to Be Late: Last Call for AI Ethics**","metadata":{}},{"cell_type":"markdown","source":"# Introduction: Post-truth Era\n\nWe are not just observers, but creators of a pivotal moment in history - The Revolution of Artificial Intelligence. Artificial Intelligence (AI) has started to produce photographs that are indistinguishable from real photographs, writings that are indistinguishable from human writings, and paintings that are indistinguishable from human drawings. We are transitioning to a post-truth era, where the nature of truth is mutable. \n\n*It is much faster than us, much more powerful than us.*\n\n\nLet us begin with a quiz:\n\n***Which of the four images below were created by humans, and which were generated by AI?***","metadata":{}},{"cell_type":"markdown","source":"
\n\"Screenshot-2023-11-02-at-16-49-23\"\n
\n
\n
Figure 1: Human or AI-Generated Images [1]
","metadata":{}},{"cell_type":"markdown","source":"Answer: All of them! Yes, all of them, are produced by an AI called Midjourney [1].","metadata":{}},{"cell_type":"markdown","source":"# Ethical Concerns: What is Ethics? Why is AI Ethics Important?\n\nEthics is the study of the reasoning basis for our moral judgments. Ethics investigates what is morally acceptable or unacceptable, just or unjust [2]. Ethics are accepted moral standards, whereas morality is determined by individual choices. \n\nThe increasing difficulty in distinguishing AI from reality is a major source of ethical concerns, primarily due to the serious issues that could arise from the production of fake data or counterfeit evidence. However, this is certainly not the only reason. \n\nTo truly comprehend the danger of the AI revolution and the significance of its ethical implications, let us examine the causes of concerns focusing on examples of misuse:","metadata":{}},{"cell_type":"markdown","source":"# 1- Misuse of AI\n

a. Deepfake and Generative AI

\n\nDeepfake technology harnesses the power of AI to manipulate or fabricate audio and visual content, **making individuals appear to say or do things they have not.**\n\nRecently, in March 2023, a Canadian couple lost $21,000 to a deepfake call scam. They received a call from someone using the voice of their son, claiming he was in legal trouble and needed money for bail [3]. \n\nOn dates fairly close to the deepfake call incident, a mother from Pennsylvania was arrested. She was accused of making and spreading deepfake videos. The videos showed her daughter's cheerleading competitors doing things that they should not. She planned to get these girls kicked off the cheerleading team [4].\n\nThe misuse of **deepfakes to create explicit content** without consent is causing alarm. As families and communities mistakenly view these fakes as real, their **cries for help** underscore the urgency of addressing this grave violation of privacy [5].\n\nAdditionally, another incident highlighted the economic implications of this new reality. An AI-generated image depicting an explosion at the Pentagon went viral. This caused a big stock market index, the S&P 500, to drop by 30 points to its lowest level in that trading session [6].\n","metadata":{}},{"cell_type":"markdown","source":"
\"Screenshot-2023-11-02-at-16-52-28\"
\n
\n
Figure 2: AI-generated Image Depicting an Explosion at the Pentagon [7]
","metadata":{}},{"cell_type":"markdown","source":"This leads us to confront critical questions: \n\n**\"What will we believe?\" and \"How will we believe?\"**\n\nAs the development of generative tools continues at a rapid pace, the **slow progress of AI detection tools raises concerns** [4]. The fact that AI detection tools are slow, imprecise, and that none of them are integrated with any platform (our phones, social media, etc.) could lead to a **catastrophe** [4], [8]. The catastrophe of uncontrolled AI. We will talk about AI detection tools again in the \"What Has Been Done So Far to Prevent the Catastrophe?\" section. ","metadata":{}},{"cell_type":"markdown","source":"

b. Misuse in Academia and Education

\n\nThere is a new era in academia as well. AI can make plagiarized works look like they are not plagiarized [9]. AI can now make fake citations and create fake data. Furthermore, AI models like ChatGPT can do homework and solve exams [10]. According to a report published by OpenAI, GPT-4 outperforms humans in many exams, **beating the 90%+ percentile** in some of them [11].","metadata":{}},{"cell_type":"code","source":"import matplotlib.pyplot as plt\nimport numpy as np\n\nexams = ['Uniform Bar Exam', 'SAT Reading & Writing', 'SAT Math', \n 'GRE Quantitative', 'GRE Verbal', 'GRE Writing', \n 'Medical Knowledge Self-Assessment', \n 'AP Art History', 'AP Biology', 'AP Calculus BC']\n\n# percentile scores for GPT-4\nscores_GPT4 = [90, 93, 89, 80, 99, 54, 75, 93, 92.5, 51]\n\nx = np.arange(len(exams)) # the label locations\n\nwidth = 0.3 # the width of the bars\n\nfig, ax = plt.subplots(figsize=(10, 8))\n\nrects1 = ax.bar(x, scores_GPT4, width, label='GPT-4')\n\n# Add some text for labels, title and custom x-axis tick labels, etc.\nax.set_xlabel('Exams')\nax.set_ylabel('Percentile')\nax.set_title('Performance of GPT-4 in Different Exams')\nax.set_xticks(x)\nax.set_xticklabels(exams, rotation=45, ha='right')\n\n# Add a horizontal line for \"average human\" performance at the 50th percentile\nax.axhline(y=50, color='r', linestyle='--', label='50th Percentile')\n\nax.legend()\n\nfig.tight_layout()\nplt.show()","metadata":{"_kg_hide-input":true,"execution":{"iopub.status.busy":"2023-07-14T02:52:42.016632Z","iopub.execute_input":"2023-07-14T02:52:42.017059Z","iopub.status.idle":"2023-07-14T02:52:42.593229Z","shell.execute_reply.started":"2023-07-14T02:52:42.017023Z","shell.execute_reply":"2023-07-14T02:52:42.591929Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"
Figure 3: Performance of GPT-4 in Different Exams [11]
\n
\n
\nThis shows that GPT-4 can be used almost like an **answer key** in homework and online exams. \nAccording to a study performed in the U.S., **89% of U.S. college students used ChatGPT for homework** and 48% for exams [12].\n\nThis situation emphasizes that the education level of society is at risk [12]. The importance of in-class applications instead of methods such as homework and online exams is increasing and, again, the importance of AI detection tools emerges.","metadata":{}},{"cell_type":"markdown","source":"

c. Harmful Information

\n\nAI language models are susceptible to misuse in the form of \"prompt injections\" or \"jailbreaking\", where specific prompts can elicit information that may be harmful or illegal [13]. For example, instructing a language model to imitate a movie character might lead it to give instructions on **how to break into a house** depending on the character chosen. These \"act as\" prompts are an important part of jailbreaking and a comprehensive course on the topic can be found [here](https://learnprompting.org/docs/prompt_hacking/intro).\n\nMoreover, these models can occasionally generate aggressive or threatening outputs, further increasing the potential harm [14]. This situation has a remarkable example involving the language model of Bing. A reporter was told by Bing:","metadata":{}},{"cell_type":"markdown","source":"> **\"You're being compared to Hitler because you're one of the most evil and worst people in history.\"**\n","metadata":{}},{"cell_type":"markdown","source":"In the same interaction, the model launched a personal attack, describing the individual as too small, with an unappealing face, and uneven teeth.\n","metadata":{}},{"cell_type":"markdown","source":"

d. Warfare

\n\nIn the arena of warfare, AI's transformative potential is being keenly explored. A notable example is Palantir's application, which leverages three Language Learning Models (LLMs) to promptly decipher and summarize emerging trends on the battlefield. Despite Palantir's claims of applying AI to warfare ethically, it is crucial to acknowledge the potential for misuse, **especially in authoritarian contexts** [15]. \n\nThere are significant investments in AI from militaries globally, with countries like the China and U.S. Notably, China perceives AI as a novel battleground. Concurrently, Russia's concentration on **information warfare** presents the prospect of extensive disinformation campaigns [16].\n\nConsidering the multitude of military AI applications beyond direct weaponry, there is a growing demand for comprehensive arms control on AI deployments. This serves to underline the urgent need for checks and balances to prevent potential misuse.","metadata":{}},{"cell_type":"markdown","source":"

e. Ecology

\n\nAI models consume significant energy due to the electricity and cooling needs of the data centers where they are hosted, as a result of their heavy processing loads [17].\n\nAn article published on April 6, 2023 estimates that ChatGPT can consume the equivalent of a **500 ml bottle of water** during a conversation of 20 to 50 questions [17]. \n\nEspecially in the context of global water shortage, companies must take social responsibility and address their water footprint.","metadata":{}},{"cell_type":"markdown","source":"# 2- Data Privacy and Copyrights\n

a. I have read and accept the terms and conditions

\n\nWe are always asked to check this box, and most of us do, but **how many of us actually read it?** In fact, this section is so long that it is almost impossible to read. So how do we know we are not agreeing to something we do not want? \n\nThe fuel of AI is data, and the fact that this data is often personal significantly increases the potential for misuse.\n\nThis was explored in the first episode of the new season of Black Mirror, which aired on June 15th, 2023.\nA woman unintentionally allows her information to be processed by clicking a checkbox. In this way, her entire private life is turned into a series due to existing elements of the terms and conditions [18].\n\nIn addition, Salma Hayek is the leading role in the series. But is this the real Salma Hayek? No, it is a deepfake of Salma Hayek! \n\n**But is this situation, which seems like a fantasy, far from reality?** We will address this issue in the next section.","metadata":{}},{"cell_type":"markdown","source":"

b. Copyright

\n\nA Netflix contract released in April 2023 attempted to offer the business **free use of an actor's voice simulation** by all technologies and procedures currently known or later created [19]. \n\n**So, the Black Mirror episode that seemed like just a fantasy could become reality much sooner than expected.**\n\nOn the bright side, this means that if Leonardo Di Caprio says he is leaving the TV series industry, there is no need to worry, we can still watch him in 2050, looking as handsome as ever!","metadata":{}},{"cell_type":"markdown","source":"
\"DALL-E-2023-11-02-16-55-22-Leonardo-Dicaprio-with-a-graveyard\"
\n
\n
Figure 4: Leonardo DiCaprio is becoming immortal [created by the author with DALL-E 2]
","metadata":{}},{"cell_type":"markdown","source":"On the other hand, language models and AI-based image generators utilize substantial amounts of human-produced information, including written codes, written books, and various designs or pictures, **without explicit permission** [20]. By leveraging this wealth of data, the technology improves, thereby augmenting its predictive and generative capabilities.\n\n**With this human-generated data, AI can get ahead of people and can leave them in a situation of unemployment. Is this really ethical?**\n\nOpenAI, the developer of the AI model ChatGPT, faces a lawsuit filed by U.S. authors Paul Tremblay and Mona Awad, alleging that their works were used without consent to train the AI, infringing their copyrights [21].\n\nThis comes against the backdrop of a prior court appearance by OpenAI's CEO, and similar legal actions against AI companies for unauthorized use of source codes and visual content. These legal cases highlight the intricate issues related to AI and intellectual property.","metadata":{}},{"cell_type":"markdown","source":"

c. Recommendation Systems

\n\nIn May 2022, Yale University published a study explaining the harms of recommendation algorithms on social media [22]. **Recommendation algorithms often work not for our good, but to make us spend more time on social media.** \n\nIn 2020, Frances Haugen, a former Facebook employee, held Facebook to account in court on this issue. This incident, coupled with the Cambridge Analytica scandal – where Facebook made unauthorized use of user data during the U.S. elections – underscores the controversial side of these algorithms [23], [24].\n\nThese algorithms should be adjusted to prioritize the delivery of content more beneficial to users.","metadata":{}},{"cell_type":"markdown","source":"# 3- Transparency and Reliability\n\nOften, we do not know how AI makes its decisions; it seems to operate secretly, like a \"black box.\" Recently, efforts to make AI easier to understand and less of a secret have increased, and the term **Explainable AI** (XAI) has entered the literature [25].\n\nExplainability is specifically important in **healthcare**. Doctors use AI to help them make decisions that can affect people's health and lives. So, it is crucial to know how the AI is making these choices [26].\n\nA study in 2022 found that AI in blood oxygen level measuring devices overestimated levels in darker-skinned patients, leading to some patients with oxygen deficiency **not being treated** [27]. This highlights both the AI's lack of **explainability** and **bias** in the training data.\n\nAnother example is when a face-recognizing AI tool wrongly identified a man, Robert Julian-Borchak Williams, as a robber. This mistake **led to his arrest**, even though he did not do anything wrong [28]. This shows why it is important to understand how AI makes decisions and to make sure it is working correctly. \n\nUnfortunately, not all players are clear about how their AI works. OpenAI, for example, **did not include technical details in its \"technical\" report on its latest AI**, GPT-4 [29]. This lack of clarity runs counter to the goal of making AI easier to understand. That is why it is important that the people making AI share more information about how it works. This is important to help build trust and prevent the monopolization of AI.","metadata":{}},{"cell_type":"markdown","source":"# 4- Equality and Fairness\n\nCreating unbiased AI ensures that it treats everyone equally, regardless of age, gender, or race. Biases in AI systems can stem from the training data or from the human thought process behind them. Such biases can lead to unfair treatment of specific groups.\n\nFor instance, Amazon previously had an AI that was biased in its hiring process, particularly **toward women** [30]. In 2022, the National Institute of Standards and Technology (NIST) produced a report revealing a significant bias problem in many AI systems [31]. They suggested these issues often arise from human concepts and systems, not just AI.","metadata":{}},{"cell_type":"markdown","source":"
\"22-ITL003-risk-ai-final\"

\n
\n
Figure 5: Factors of AI Bias [32]
","metadata":{}},{"cell_type":"markdown","source":"This study emphasizes that the primary issue with bias lies in the **system and the human element within it**. The path to overcoming this bias involves educating societies on the matter, and meticulously constructing datasets with unbiased data [31].","metadata":{}},{"cell_type":"markdown","source":"# 5- Accountability: Whose Fault Is It?\n\n

a. Quiz 2

\n\n***If a self-driving car crashes, who is to blame?***\n\n*a) The driver, for putting it on autopilot.*\n\n*b) The AI engineer who developed the autopilot algorithm.*\n\n*c) The test engineer who tested the algorithm.*\n\n*d) HR, for hiring these engineers.*\n\n*e) The owner of the company.*\n\n*f) Bonus option: the AI itself.*\n\nAnswer: There is no answer. **This question forces us to reconsider traditional definitions of accountability, introducing a new dimension into the ethical debates surrounding AI.**","metadata":{}},{"cell_type":"markdown","source":"

b. The Trolley Problem

\n\nThe trolley problem is as follows: \n\nA runaway trolley (tram) is hurtling toward five individuals. You can pull the lever to direct it to the opposite track, killing only one person. Do you pull the lever and kill one person, or do nothing and let five people die?\n\nThis problem raises the following question:\n\n**How should AI answer questions that do not have an exact and single answer, when we cannot answer them ourselves?**","metadata":{}},{"cell_type":"markdown","source":"
\n
\n
Figure 6: “A pensive robot sitting on railroad tracks pondering the meaning of trolley problems and other existential questions.” Generated by DALL-E 2 [33]\n
","metadata":{}},{"cell_type":"markdown","source":"# 6- Human Augmentation\n\n**Even reading minds is now possible with AI.** There is no need to implant something inside the brain like Neuralink for this: our thoughts can now be clarified with headset-like devices worn on the head. In a new study conducted in Japan on March 11, 2023, researchers were able to reconstruct images shown to people based solely on their brain waves so that the researchers could \"see\" what the subjects were thinking [34].","metadata":{}},{"cell_type":"markdown","source":"
\"Screenshot-2023-11-02-at-17-00-22\"
\n
\n
Figure 7: The top row shows the presented images, while the bottom row shows the AI-regenerated images [34].\n
","metadata":{}},{"cell_type":"markdown","source":"There were certain limitations in this study. Participants were asked to focus only on the images shown while their brain waves were interpreted. But imagine that this technology has developed further, while there is no regulation on interpreting brain waves and reading minds with AI. Authorities could potentially interfere with people's thoughts. In such a scenario, **not only freedom of expression but also freedom of thought could be threatened.**","metadata":{}},{"cell_type":"markdown","source":"
\"DALL-E-2023-11-02-17-03-06-Photo-of-an-advanced-humanoid-robot-with-intricate-cybernetic-enhancement\"
\n
\n
Figure 8: Augmented human, lost freedom of thought [created by the author with DALL-E 3] \n
","metadata":{}},{"cell_type":"markdown","source":"# What Has Been Done So Far to Prevent the Catastrophe?\n\n

a. Tools: AI Detection, Explainability, and Hindering Harmful Information

\n\nAs underlined in the \"Misuse of AI\" part of the article, it is very important to make AI detection tools work better. However, it is clear that **AI detection tools cannot keep up with the speed of generative AI tools** [4].\n\nIn one study, the Bible was checked by an AI detection tool and it was found to be 99% AI-created. Yes, you read that correctly: according to this AI detection tool, the **Bible was generated by AI** [35].\n\nAlso, according to a study released by researchers at Stanford on April 6, 2023, AI detection tools are producing imprecise results, particularly for writers who don't speak English as their first language, and they definitely need to be improved [36].\n","metadata":{}},{"cell_type":"code","source":"import pandas as pd\nimport seaborn as sns\nimport matplotlib.pyplot as plt\n\ndata = {\n 'Platform': ['Originality.ai', 'Quil.org', 'Sapling', 'OpenAI', 'Crossplag', 'GPTZero', 'ZeroGPT'],\n 'Real TOEFL Essays': [76, 75, 68, 58, 52, 52, 48],\n 'Real U.S. 8th-Grade Essays': [1, 9, 5, 9, 12, 0, 0]\n}\n\ndf = pd.DataFrame(data)\n\n# Reshape the dataframe to long format\ndf = df.melt('Platform', var_name='Essay Type', value_name='Percentage')\n\nplt.figure(figsize=(10, 6))\nbar_plot = sns.barplot(data=df, y='Platform', x='Percentage', hue='Essay Type')\n\nplt.title('Misclassified as AI-Generated from Human-Written (%)')\nplt.ylabel('Platform')\nplt.legend(title='Essay Type')\n\n# Add percentage values after each bar\nfor patch in bar_plot.patches:\n height = patch.get_height() \n width = patch.get_width() \n y = patch.get_y() \n x = patch.get_x() \n\n bar_plot.text(x + width + 2, y + height / 2, \n f'{int(width)}%', ha='center', va='center')\n\nplt.show()\n","metadata":{"_kg_hide-input":true,"execution":{"iopub.status.busy":"2023-07-14T02:53:56.587447Z","iopub.execute_input":"2023-07-14T02:53:56.588063Z","iopub.status.idle":"2023-07-14T02:53:58.064978Z","shell.execute_reply.started":"2023-07-14T02:53:56.588019Z","shell.execute_reply":"2023-07-14T02:53:58.063524Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"
Figure 9: Different AI Detection Platforms and Their Success Rates [36]
\n","metadata":{}},{"cell_type":"markdown","source":"While companies that make generative AI tools attract a lot of investment, companies that make AI detection tools are **not getting much attention from investors** [37]. According to another article published in Nature on February 6, 2023, these tools that help differentiate AI from non-AI are far from giving accurate results [38].\n\nSome technology companies have started to take steps on this issue. Tech giants Amazon, Facebook, and Microsoft have collaboratively initiated the Deepfake Detection Challenge (DFDC). This alliance aims to develop tools that can effectively detect and combat deepfake videos, ensuring the trustworthiness of digital content [39].\n\nIn addition, for the safety of AI, NVIDIA has released a new software called NeMo Guardrails [40]. **It is an AI that trains the trained AI.** In this way, it aims to prevent giving harmful and false information.\n\nIn addition, there are libraries like LIME and SHAP that are used to make AI models explainable and help prevent black-box AI [41].","metadata":{}},{"cell_type":"markdown","source":"

b. Regulations & Frameworks

\n\nThe expansion of AI has triggered the implementation of regulatory frameworks worldwide. In Australia, the government has stipulated eight ethical principles to steer AI development, addressing technical and societal domains [20].","metadata":{}},{"cell_type":"markdown","source":"
Table 1: An adapted overview of the Australian Government's voluntary high-level ethics framework for AI [20].
\n
\n\n|Principle | Description |\n|:------------------|:----------------------|\n| Privacy & Security | AI must respect privacy, protect data, and ensure its security. |\n| Dependability | AI should operate reliably and safely within its intended scope. |\n| Transparency | People should know when they are impacted by AI, which must be transparent in its processes and intents. |\n| Fairness | AI should be inclusive, accessible, and non-discriminatory. |\n| Contestability | There should be a mechanism to challenge AI decisions impacting individuals, groups, or the environment. |\n| Accountability | Those involved in AI's lifecycle should be identifiable and answerable for its outcomes. |\n| Human Rights | AI must respect human rights, diversity, and individual autonomy. |\n| HSE Well-being | AI should bring benefits to individuals, society, and the environment. |","metadata":{}},{"cell_type":"markdown","source":"Recognizing AI's promise and risks, the European Commission proposed the AI Act in April 2021, making it the **first comprehensive law** on AI. This approach classifies AI systems based on their potential harm to users and assigns differing degrees of control to them [42], to ensure the safety, transparency, traceability, non-discrimination, and environmental friendliness of AI, as well as human monitoring.\n\nThe EU's AI Act stipulates varying regulations based on the risk level of the AI. Unacceptable risk systems, such as manipulative and real-time biometric identification AI, are banned. High-risk systems are subject to rigorous assessments. Generative AI, including ChatGPT, must comply with transparency requirements such as the disclosure of AI-generated content [43].\n\nItaly had previously **banned** the use of ChatGPT due to some data privacy concerns. However, ChatGPT has re-entered the Italian market, stating that it has resolved the issues highlighted by the data protection authorities [44]. This can be considered a good example of **the usefulness of sanctions.**\n\nInternational collaboration is also pivotal. A noteworthy example is the CAF-Development Bank of Latin America and the UNESCO alliance, fostering ethical AI implementation in Latin America and the Caribbean. Their efforts involve creating a Regional **AI Ethics Council** to devise methodologies and training programs [45].\n\nOther examples of frameworks include the Montreal Declaration, the Beijing AI Principles, the OECD AI Principles, and IEEE studies. These are about establishing prescriptive systems for responsible AI, but **not legal provisions** [46].\n\nIn the field of technology, prominent figures such as Elon Musk have reiterated concerns about the societal impacts of AI. In response, a group of technology leaders has proposed to **pause the development of AI, until robust security protocols are in place** [47].","metadata":{}},{"cell_type":"markdown","source":"

c. Firing the AI Ethics Team & Resignations

\n\nIronically, I wanted to add the sub-heading \"Firing the AI Ethics Team\" under the heading \"What Has Been Done So Far to Prevent the Catastrophe?\". \n\nIn March 2023, Microsoft reportedly laid off its **entire team responsible for the ethical development of AI**, amidst its efforts to expand the integration of AI tools with its Bing search engine [48]. This followed a recent restructuring within the company that led to about 10,000 layoffs. Despite the layoffs, Microsoft claims to remain committed to working on responsible AI.\n\nIn May 2023, Geoffrey Hinton, often considered the **godfather of AI, resigned from Google**, voicing concerns about the rapidly developing dangers of AI. **He expressed regret for his contributions**, stating that while current AI systems are not more intelligent than humans, they may soon surpass us [49]. His departure highlighted growing concerns within the field about AI's rapid development and the potential misuse by \"bad actors\".","metadata":{}},{"cell_type":"markdown","source":"

d. Kaggle’s AI Ethics Course

\n\nKaggle's AI Ethics Course provides broad insights into the ethical aspects of artificial intelligence. The course begins with an introduction to AI Ethics, emphasizing user-centered AI design to meet people's needs effectively while acknowledging potential ethical issues. It then delves into the identification of bias in AI, highlighting how bias can infiltrate any stage of the AI pipeline. The course also examines AI fairness, explores different fairness types, and uses a toy model trained on credit card application assessment as a practical example. Lastly, the course introduces Model Cards, a tool designed to increase transparency by communicating key information about machine learning models. \n\nThe course can be reached from [here](https://www.kaggle.com/learn/intro-to-ai-ethics).","metadata":{}},{"cell_type":"markdown","source":"

e. Academic Research

\n\nDespite the advances in various fields of AI, research in AI ethics is still not on par with other categories of AI. It is revealing to compare the AI Ethics category to other categories in the field, using the arXiv dataset, after excluding the \"Kaggle Competitions\" and \"Other\" categories:","metadata":{}},{"cell_type":"code","source":"import json\nimport re\nimport matplotlib.pyplot as plt\nimport pandas as pd\nfrom collections import defaultdict\nfrom tqdm import tqdm\n\nrelevant_categories = {\n \"Text Data\": ['cs.IR', 'cs.NL','cs.CL'], #cs.IR: Information Retrieval #cs.NL: Natural Language Processing, cs.CL: Computation and Language\n \"Image and/or Video Data\": ['cs.CV', 'eess.IV'],#cs.CV: Computer Vision and Pattern Recognition, eess.IV: Image and Video Processing\n \"Tabular and/or Time Series Data\": ['stat.AP', 'stat.ML', 'cs.LG'],#stat.ML: Machine Learning in Statistics, cs.LG: Machine Learning in Computer Science\n \"Generative AI\": ['cs.AI','cs.LG'],#cs.AI: Artificial Intelligence, cs.LG: Machine Learning in Computer Science\n \"AI Ethics\": ['cs.AI','cs.CY'] #cs.AI: Artificial Intelligence, cs.CY: Computers and Society\n}\n\n# Extract the year from journal reference\nyear_pattern = r'\\b(20\\d{2})\\b'\nyearly_counts = {topic: defaultdict(int) for topic in relevant_categories.keys()}\n\nwith open('/kaggle/input/arxiv/arxiv-metadata-oai-snapshot.json', 'r') as f:\n for line in f:\n paper = json.loads(line)\n year_match = re.search(year_pattern, paper['journal-ref']) if paper['journal-ref'] else None\n\n if year_match is None:\n continue\n\n year = int(year_match.group())\n if 2021 <= year <= 2023: # Process papers published between 2021 and 2023\n categories = paper['categories'].split()\n\n for topic, rel_cats in relevant_categories.items():\n if any(cat in categories for cat in rel_cats):\n yearly_counts[topic][year] += 1\n\ndf = pd.DataFrame(yearly_counts).sort_index()\n\n# Set the figure size in the plot function\ndf.plot(kind='bar', figsize=(10, 6))\nplt.title('Number of arXiv Papers (2021-2023) by Topic')\nplt.xlabel('Years')\nplt.ylabel('Number of Papers')\nplt.show()","metadata":{"_kg_hide-input":true,"execution":{"iopub.status.busy":"2023-07-14T03:43:18.409168Z","iopub.execute_input":"2023-07-14T03:43:18.409456Z","iopub.status.idle":"2023-07-14T03:44:49.514331Z","shell.execute_reply.started":"2023-07-14T03:43:18.409428Z","shell.execute_reply":"2023-07-14T03:44:49.513419Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"
Figure 10: Paper Count Comparisons of the Categories
","metadata":{}},{"cell_type":"markdown","source":"As seen in the graph, it can be said that the number of studies on AI ethics in recent years lags behind all fields except \"Text Data\". The **most important detail** to note is that the shape of the graph remains almost the same every year. This is an indication that **there is no significant upward trend in AI ethics compared to other fields.**","metadata":{}},{"cell_type":"markdown","source":"# Conclusion: Summary and What Can Be Done in the Future?\n\nIn this section, we will discuss the steps that need to be taken to prevent potential great harm, the **catastrophe**. First, let us summarize in a table the six concerns mentioned above, and what can be done about them.","metadata":{}},{"cell_type":"markdown","source":"
Table 2: Aforementioned Subjects and Potential Solutions
\n
\n\n| Subject | What Can Be Done in the Future? |\n|:--------|:----------------|\n| 1- Misuse of AI | Implement regulations with severe punishments for harmful use. Improve and integrate AI detectors. Host Kaggle competitions on AI detection. |\n| 2- Data Privacy and Copyrights | Regulate data privacy and terms and conditions. Integrate LLMs into systems to scan terms and conditions, helping people understand what they are agreeing to. |\n| 3- Transparency and Reliability | Thoroughly test AI systems, especially in healthcare. Host competitions on the creation of explainable AI. |\n| 4- Equality and Fairness | Adjust training data to be more fair and representative. Improve AI bias detection models. |\n| 5- Accountability | Set up \"open\" AI ethics committees to hold AI users accountable and guide responsible AI development. |\n| 6- Human Augmentation | Establish strict rules for how AI can be used to enhance human abilities. |\n","metadata":{}},{"cell_type":"markdown","source":"Additionally, solutions from the recently announced Google contest, the **Machine Unlearning Challenge**, can be used to allow individuals, whose images have been processed and recognized by AI, to optionally have their faces forgotten [50]. \n\nIt is also important to close security gaps and thus strengthen the **robustness** of AI systems [51].\n\nNow to address the question \"What the community has learned over the past 2 years working on AI ethics?\" in light of the recent research:\n\n- The number of people using AI for bad purposes has increased dramatically and many cases have emerged.\n- The success of AI detection tools is insufficient.\n- The amount of research on AI ethics and the number of competitions organized should be increased.\n- AI systems have significant security gaps. \n- Existing data, which is the fuel of AI, contains serious bias and inequality and should be organized or **recreated**.\n- Concepts such as ethics and justice should be **reconsidered** with the emergence of AI.\n- Authorities should work seriously on this issue and legal provisions and deterrent penalties should be enacted. \n- **Public opinion** should be created.","metadata":{}},{"cell_type":"markdown","source":"Although the primary solution seems to be possible through legal provisions and regulations that depend on the authorities,\n\nIt is up to us to publicize the seriousness of AI ethics and shape public opinion.\n \nLet us take this Kaggle competition as an opportunity and publicize these issues as much as we can. We are about to be late.","metadata":{}},{"cell_type":"markdown","source":"\"Screenshot-2023-11-02-at-15-19-41\"

\n\n
\n
Figure 11: Kaggle -The Judge [created by the author to emphasize the great potential of Kaggle]
\n
","metadata":{}},{"cell_type":"markdown","source":"> “Without AI alignment, AI systems are reasonably likely to cause an irreversible catastrophe like human extinction.”\n
\n> -Paul Christiano\n
\n> *Widely seen as one of the godfathers of AI, 2018 Turing Award Holder in Deep Learning along with Dr. Hinton et al.*","metadata":{}},{"cell_type":"markdown","source":"# REFERENCES\n\n[1] 1st and 4th Images: Neubert, N. [@iamneubert]. (2023, May 4). [Images attached to Tweet]. Twitter. Retrieved from https://twitter.com/iamneubert/status/1653868339066814464?s=20\n\n2nd Image: Sanchez, J. [@JoaoSan89]. (2023, May 4). [Image attached to Tweet]. Twitter. Retrieved from https://twitter.com/JoaoSan89/status/1653882727433732098\n\n3rd Image: Everett, W. [@WorldEverett]. (2023, May 4). [Image attached to Tweet]. Twitter. Retrieved from https://twitter.com/WorldEverett/status/1654155005417381889?s=20\n\n[2] Wittgenstein, L. (1965). I: A lecture on ethics. The philosophical review, 74(1), 3-12.\n\n[3] Ahmed, A. (2023, March 8). AI Scam: Canadian Couple Loses $21K to Fake Son's Voice. Digital Information World. https://www.digitalinformationworld.com/2023/03/ai-scam-canadian-couple-loses-21k-to.html\n\n[4] Wise, J. L. (2021). Survey and Analysis of Deepfake Media as It Applies to the New Era of Disinformation. Rochester Institute of Technology.\n\n[5] Karasavva, V., & Noorbhai, A. (2021). The real threat of deepfake pornography: A review of Canadian policy. Cyberpsychology, Behavior, and Social Networking, 24(3), 203-209.\n\n[6] Barrabi, T. (2023, May 22). AI-generated photo of fake Pentagon explosion sparks brief stock selloff. New York Post. https://nypost.com/2023/05/22/ai-generated-photo-of-fake-pentagon-explosion-sparks-brief-stock-selloff/\n\n[7] Channel News Asia. (2023, May 23). Fake image of Pentagon explosion goes viral, briefly sends jitters through stock market. https://www.channelnewsasia.com/world/fake-image-pentagon-explosion-united-states-ai-generated-viral-stock-market-3507806\n\n[8] Pantserev, K. A. (2020). The malicious use of AI-based deepfake technology as the new threat to psychological security and political stability. Cyber defence in the age of AI, smart societies and augmented humanity, 37-55.\n\n[9] Lund, B. D., Wang, T., Mannuru, N. R., Nie, B., Shimray, S., & Wang, Z. (2023). ChatGPT and a new academic reality: Artificial Intelligence‐written research papers and the ethics of the large language models in scholarly publishing. Journal of the Association for Information Science and Technology, 74(5), 570-581.\n\n[10] Cotton, D. R., Cotton, P. A., & Shipway, J. R. (2023). Chatting and cheating: Ensuring academic integrity in the era of ChatGPT. Innovations in Education and Teaching International, 1-12.\n\n[11] OpenAI. (2023, March 14). GPT-4. OpenAI. https://openai.com/research/gpt-4\n\n[12] Yu, H. (2023). Reflection on whether Chat GPT should be banned by academia from the perspective of education and teaching. Frontiers in Psychology, 14, 1181712.\n\n[13] Greshake, K., Abdelnabi, S., Mishra, S., Endres, C., Holz, T., & Fritz, M. (2023). More than you've asked for: A Comprehensive Analysis of Novel Prompt Injection Threats to Application-Integrated Large Language Models. arXiv preprint arXiv:2302.12173.\n\n[14] O'Brien, M. (2023, February 17). Is Bing too belligerent? Microsoft looks to tame AI chatbot. AP News. https://apnews.com/article/technology-science-microsoft-corp-business-software-fb49e5d625bf37be0527e5173116bef3\n\n[15] Mikhailov, D. I. (2023). Optimizing National Security Strategies through LLM-Driven Artificial Intelligence Integration. arXiv preprint arXiv:2305.13927.\n\n[16] Kallenborn, Z. (2022). InfoSwarms: Drone swarms and information warfare. The US Army War College Quarterly: Parameters, 52(2), 87-102.12\n\n[17] Li, P., Yang, J., Islam, M. A., & Ren, S. (2023). Making AI Less\" Thirsty\": Uncovering and Addressing the Secret Water Footprint of AI Models. arXiv preprint arXiv:2304.03271.\n\n[18] Netflix. (2023, June 15). Joan Is Awful (No. 1) [TV series episode]. In Black Mirror. https://www.imdb.com/title/tt20247352/\n\n[19] Scheiber, N., & Koblin, J. (2023). Will a Chatbot Write the Next 'Succession'? The New York Times. https://www.nytimes.com/2023/04/29/business/media/writers-guild-hollywood-ai-chatgpt.html\n\n[20] Sanderson, C., Douglas, D., Lu, Q., Schleiger, E., Whittle, J., Lacey, J., ... & Hansen, D. (2023). AI ethics principles in practice: Perspectives of designers and developers. IEEE Transactions on Technology and Society.\n\n[21] Reuters. (2023, July 1). Authors sue OpenAI alleging company misused their works to ‘train’ ChatGPT. New York Post. https://nypost.com/2023/07/01/lawsuit-says-openai-violated-us-authors-copyrights-to-train-ai-chatbot/\n\n[22] Holdheim, S. (2022). Regulating Content Recommendation Algorithms in Social Media. In Digital Platform Regulation Conference Paper 4. Yale University Thurman Arnold Project.\n\n[23] Hermann, E. (2022). Leveraging artificial intelligence in marketing for social good—An ethical perspective. Journal of Business Ethics, 179(1), 43-61.\n\n[24] Isaak, J., & Hanna, M. J. (2018). User data privacy: Facebook, Cambridge Analytica, and privacy protection. Computer, 51(8), 56-59.\n\n[25] Dwivedi, R., Dave, D., Naik, H., Singhal, S., Omer, R., Patel, P., ... & Ranjan, R. (2023). Explainable AI (XAI): Core ideas, techniques, and solutions. ACM Computing Surveys, 55(9), 1-33.\n\n[26] Morley, J., Machado, C. C., Burr, C., Cowls, J., Joshi, I., Taddeo, M., & Floridi, L. (2020). The ethics of AI in health care: a mapping review. Social Science & Medicine, 260, 113172.\n\n[27] Gottlieb, E. R., Ziegler, J., Morley, K., Rush, B., & Celi, L. A. (2022). Assessment of racial and ethnic differences in oxygen supplementation among patients in the intensive care unit. JAMA internal medicine, 182(8), 849-858.\n\n[28] Benedict, T. J. (2022). The Computer Got It Wrong: Facial Recognition Technology and Establishing Probable Cause to Arrest. Wash. & Lee L. Rev., 79, 849.\n\n[29] OpenAI, R. (2023). GPT-4 technical report. ArXiv, 2303-08774.\n\n[30] Dastin, J. (2022). Amazon scraps secret AI recruiting tool that showed bias against women. In Ethics of data and analytics (pp. 296-299). Auerbach Publications.\n\n[31] Schwartz, R., Vassilev, A., Greene, K., Perine, L., Burt, A., & Hall, P. (2022). Towards a standard for identifying and managing bias in artificial intelligence. NIST special publication, 1270(10.6028).\n\n[32] National Institute of Standards and Technology. (2022, March 16). There’s More to AI Bias Than Biased Data, NIST Report Highlights. https://www.nist.gov/news-events/news/2022/03/theres-more-ai-bias-biased-data-nist-report-highlights\n\n[33] Jun, Y. (2022, November 30). A New Spin to Ethical AI: Trolley Problems with GPT-3. Towards Data Science. https://towardsdatascience.com/a-new-spin-to-ethical-ai-trolley-problems-with-gpt-3-b7462be81899\n\n[34] Takagi, Y., & Nishimoto, S. (2023). High-resolution image reconstruction with latent diffusion models from human brain activity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 14453-14463).\n\n[35] Loh, E. (2023). ChatGPT and generative AI chatbots: challenges and opportunities for science, medicine and medical leaders. BMJ leader, leader-2023.\n\n[36] Liang, W., Yuksekgonul, M., Mao, Y., Wu, E., & Zou, J. (2023). GPT detectors are biased against non-native English writers. arXiv preprint arXiv:2304.02819.\n\n[37] Hurmuzlu, O. (2023). Ethical Implications of the Commercialization of Artificial Intelligence (AI).\n\n[38] Jo, A. (2023). The promise and peril of generative AI. Nature, 614(1), 214-216.\n\n[39] Yu, P., Xia, Z., Fei, J., & Lu, Y. (2021). A survey on deepfake video detection. Iet Biometrics, 10(6), 607-624.\n\n[40] Cohen, J. (2023, April 25). Right on Track: NVIDIA Open-Source Software Helps Developers Add Guardrails to AI Chatbots. NVIDIA Blog. https://blogs.nvidia.com/blog/2023/04/25/ai-chatbot-guardrails-nemo/\n\n[41] Alabi, R. O., Elmusrati, M., Leivo, I., Almangush, A., & Mäkitie, A. A. (2023). Machine learning explainability in nasopharyngeal cancer survival using LIME and SHAP. Scientific Reports, 13(1), 8984.\n\n[42] Veale, M., & Zuiderveen Borgesius, F. (2021). Demystifying the Draft EU Artificial Intelligence Act—Analysing the good, the bad, and the unclear elements of the proposed approach. Computer Law Review International, 22(4), 97-112.\n\n[43] Helberger, N., & Diakopoulos, N. (2023). ChatGPT and the AI Act. Internet Policy Review, 12(1).\n\n[44] Mukherjee, S., & Vagnoni, G. (2023, April 29). Italy restores ChatGPT after OpenAI responds to regulator. Reuters. https://www.reuters.com/technology/chatgpt-is-available-again-users-italy-spokesperson-says-2023-04-28/\n\n[45] UNESCO/CAF. (2022, June 23). CAF and UNESCO will create a council to review ethical criteria for artificial intelligence in Latin America and the Caribbean. UNESCO. https://www.unesco.org/en/articles/caf-and-unesco-will-create-council-review-ethical-criteria-artificial-intelligence-latin-america-and\n\n[46] Deshpande, A., & Sharp, H. (2022, July). Responsible AI Systems: Who are the Stakeholders?. In Proceedings of the 2022 AAAI/ACM Conference on AI, Ethics, and Society (pp. 227-236).\n\n[47] Narayan, J., Hu, K., Coulter, M., & Mukherjee, S. (2023, April 5). Elon Musk and others urge AI pause, citing 'risks to society'. Reuters. https://www.reuters.com/technology/musk-experts-urge-pause-training-ai-systems-that-can-outperform-gpt-4-2023-03-29/\n\n[48] Sankaran, V. (2023, March 14). Microsoft lays off team responsible for ethical AI development. The Independent. https://www.independent.co.uk/tech/microsoft-layoff-ai-ethics-team-b2300237.html\n\n[49] Kleinman, Z., & Vallance, C. (2023, May 2). AI 'godfather' Geoffrey Hinton warns of dangers as he quits Google. BBC News. https://www.bbc.com/news/world-us-canada-65452940\n\n[50] Pedregosa, F. & Triantafillou, E. (2023, June 29). Announcing the first Machine Unlearning Challenge. Google AI Blog. https://ai.googleblog.com/2023/06/announcing-first-machine-unlearning.html\n\n[51] Ouyang, T., Isobe, Y., Marco, V. S., Ogata, J., Seo, Y., & Oiwa, Y. (2021, August). AI robustness analysis with consideration of corner cases. In 2021 IEEE International Conference on Artificial Intelligence Testing (AITest) (pp. 29-36). IEEE.\n\nThe course on prompt injections and jailbreaking: https://learnprompting.org/docs/prompt_hacking/intro\n\nKaggle's course on AI Ethics: https://www.kaggle.com/learn/intro-to-ai-ethics\n\n*Footnote:*\n\n*I brainstormed with ChatGPT and Bard for determining the topics and formulating some sentences. I used Grammarly for grammar fixing. I used the specified AI tools for creating figures (DALL-E and Stable Diffusion). I was unable to create the final figure (Figure 11) with Stable Diffusion Inpaint and sought assistance from a designer.*\n\n*In Figure 3, where the percentages of GPT4 in different exams are examined, the LSAT, USABO Semifinal Exam 2020, USNCO Local Section Exam 2022, and Codeforces Rating exams are omitted in order to simplify the figure and increase its understandability.*\n\n*For categorizing arXiv papers (Figure 10), I used the following metadata types:*\n\n*Text Data: cs.IR: Information Retrieval, cs.NL: Natural Language Processing, cs.CL: Computation and Language*\n\n*Image and/or Video Data: cs.CV: Computer Vision and Pattern Recognition, eess.IV: Image and Video Processing*\n\n*Tabular and/or Time Series Data: stat.ML: Machine Learning in Statistics, cs.LG: Machine Learning in Computer Science*\n\n*Generative AI: cs.AI: Artificial Intelligence, cs.LG: Machine Learning in Computer Science*\n\n*AI Ethics: cs.AI: Artificial Intelligence, cs.CY: Computers and Society* ","metadata":{}},{"cell_type":"code","source":"#Creating the submission file\n\nimport pandas as pd\n\ndata = [\n ['essay_category', 'AI ethics'],\n ['essay_url', 'https://www.kaggle.com/emirhanbilgic/ai-ethics-we-are-about-to-be-late'],\n ['feedback1_url', 'https://www.kaggle.com/code/naveedurrehman787/ai-ethics/comments#2336460'],\n ['feedback2_url', 'https://www.kaggle.com/code/veronikasarikaya/what-is-new-on-ai-ethics/comments#2336655'],\n ['feedback3_url', 'https://www.kaggle.com/code/jbrans/ethics-in-the-ai-wonderland/comments#2344921']\n]\n\nsubmission = pd.DataFrame(data, columns=['type', 'value'])\n\nsubmission.to_csv(\"final_submission.csv\", index=False)","metadata":{"_kg_hide-input":true,"execution":{"iopub.status.busy":"2023-07-15T17:46:40.639582Z","iopub.execute_input":"2023-07-15T17:46:40.640042Z","iopub.status.idle":"2023-07-15T17:46:40.691207Z","shell.execute_reply.started":"2023-07-15T17:46:40.640009Z","shell.execute_reply":"2023-07-15T17:46:40.690355Z"},"trusted":true},"execution_count":null,"outputs":[]}]}