![](\"https://eu-images.contentstack.com/v3/assets/blt6b0f74e5591baa03/blt9f79c6927519a2ca/646cbbc8b09aaa69dadb07b3/Untitled_design_(40).jpg\")
\n\n\n \nAnother issue is attributing _responsibility_ to AIs. We tend to believe responsibility requires “knowing what you’re doing.” Generative AIs -- as with all deep learning models -- lack certain humanlike prerequisites to this, such as mental states, emotional context, and empathy. In addition, they are maddeningly opaque. They lack the ability to explain their decision outputs, which presents yet another obstacle to trust. If a machine cannot explain why it decides something – if humans cannot interpret their behavior – it is difficult to accept its outputs at face value; it is difficult to maintain _trust_. \n\nThese notions are further jeopardized by individuals who are either not educated enough to know they are misinformed, or are willfully ignorant. Here, the flip sides of these values reveal another danger of AI: blind, mindless, [complete trust is a recipe for eschewing responsibility](https://openai.com/blog/governance-of-superintelligence), paving the way for the [\"banality of evil](https://platypus1917.org/wp-content/uploads/2014/01/arendt_eichmanninjerusalem.pdf).\" People may become self-infantilizing because it is more comfortable to cast off moral responsibility – to farm out the onus of moral work to a machine that appears competent, but which is in fact incapable of moral reasoning. It is more comfortable to neglect one’s own moral responsibility and feign ignorance of the consequences of one’s actions than to make morally ambiguous judgements. But while moral ambiguity is painful for humans, but impossible for machines. \n\nWe must decide if we hand over the moral reins to opaque algorithms, or if we reserve trust for human-human relationships, and rest responsibility squarely on our own squishy shoulders. ","metadata":{}},{"cell_type":"markdown","source":"\n# How should machines treat us? \n__alignment, justice, and power__\n","metadata":{}},{"cell_type":"markdown","source":"\n### The Alignment Problem\n\nCoined by Stewart Russell, AI “Alignment” refers to how we ensure the output of AI Algorithms are “aligned” with human values. Many organizations present [various frameworks](https://www.scu.edu/ethics/focus-areas/technology-ethics/resources/a-multilevel-framework-for-the-ai-alignment-problem/) of understanding, and analysis of values in output, as well as technical inspection tools. Unfortunately, this presupposes that we have and can articulate our values, [which is disputable to say the least](https://philpapers.org/rec/TURAAP). \n\nBecause we are unable to readily express and articulate our values – especially in [machine-understandable](https://delphi.lexxion.eu/article/delphi/2019/1/6) ways – the question of how best to ensure AI algorithms respect those values is an [extremely active](https://www.sciencedirect.com/science/article/pii/S1566253523001148)) area of research, but has yet borne little fruit.","metadata":{}},{"cell_type":"markdown","source":"\n### Justice\n\nOne seemingly obvious answer to how we should be treated by AI algorithms is: fairly. But, to instill “fairness” into an algorithm, we first must ask ourselves what we believe is “fair.” Lately, the topic of justice and [equality versus equity](https://journals.sagepub.com/doi/abs/10.1177/000271627340900103?casa_token=Dj-yb0GiqMYAAAAA:IguxfS-FevSumlXnwjJMrkhbygtvzY5JC1WPddfg71B7QNc98O6dwFI9-AQLPh8pPaum8iGMgastEg) has seized the spotlight, and collective discussion on the matter is far from settled. In this case, we must decide whether algorithms ought to be programmed to address issues of fairness in society. Should AI output reflect the world as it is, or actively contribute to a world we _want_ to live in? \n\nAI algorithms are currently used to make decisions about people’s lives – [putting people in prison, deciding who gets a loan, and accepting students into university](https://en.wikipedia.org/wiki/Weapons_of_Math_Destruction#:~:text=Weapons%20of%20Math%20Destruction%20is,ways%20that%20reinforce%20preexisting%20inequality.). Often, the models used in to make these decisions have [extrardinarily biased outcomes](https://arxiv.org/abs/1908.09635). Not only are these decisions opaque, but the data that trains these models are themselves riddled with bias. Indeed, LLMs such as LLaMA and GPT-4 are trained on an enormous corpus of internet data – data in which [attitudes, opinions, and language itself](https://www.cell.com/patterns/pdf/S2666-3899(21)00184-7.pdf) is both explicitly and inherently biased. And so, the mantra that has been repeated for years continues rings true: Bias In, Bias Out.\n\nGiven this issue of biased data, is it possible to achieve “fairness” even if we could picture it? Can or should programmers actively program out the intrinsic bias of society? Logistically, this seems impossible, as programmers often don’t know what is in the data used to train their models, let alone benchmark its quality. Thus it is unclear how or if technologists ought to consider _justice_ when developing new AIs.\n\n","metadata":{}},{"cell_type":"markdown","source":"\n### Power\n\nThis leads troublingly into societal views of power. Discussions of justice cannot be divorced from discussions of who decides what is just. In this case, overwhelmingly, the arbiters of AI algorithms have historically been extremely powerful companies. [Economically and demographically homogeneous, advantaged groups](https://www.livemint.com/technology/artificial-intelligence-favours-the-rich-and-powerful-disadvantageous-for-the-rest-mozilla-11658157856156.html) make decisions about how AI algorithms affect billions of users, often to the detriment of those very users (as in aforementioned cases of misinformation and life-altering decisions such as getting a loan).\nIn this context, AI does not in and of itself create a dispute over values, but pushes us to examine justice with a new lens. The issue is not so much how _machines_ should treat us, but how _people with power_ over the machines should treat _other people_ without such power.\n\nOne answer to this is an effort to [“Democratize” AI](https://www.sciencedirect.com/science/article/abs/pii/S1546144019305307). This recent buzzword refers to increasing education about and access to AI algorithms. While this is certainly well-intended, it remains to be seen who is the real benefactor of AI Democratization. It is nice to believe democratization is pursued for the good of everyone, but the reality is only a privileged few have access to [educational resources and AI applications](https://arxiv.org/abs/2010.15581). Additionally, increasing users’ abilities to incorporate AI into their workflows may be purely to increase worker productivity, thus contributing to the cycle of [giving more power to those with power](https://www.theverge.com/23623495/ai-profits-winners-losers-openai-notion-snapchat). ","metadata":{}},{"cell_type":"markdown","source":"\n# Turning Values into Law\n__security, privacy, stewardship, inclusivity__\n\nWe have thus far explored how two straightforward questions degenerate into an astonishingly complex minefield of ethical issues in the field of AI. Still, the importance of enshrining values into codified laws have lured many to navigate this treacherous landscape. \n\nWhat we have isn't working. The brilliant and constantly updated [AI Incident Database](https://incidentdatabase.ai/) demonstrates that current implementations of AI, while perhaps technically legal for now, are not societally _allowable_ in their present form. At the time of writing, the database includes over 3000 incident and issue reports of harm caused by AI systems. Legislators are now left asking themselves how to best learn about and act on problematic uses of AI.","metadata":{}},{"cell_type":"markdown","source":"Incidents by victim category | Incidents by type of harm\n:-------------------------:|:-------------------------:\n | \n\n\nFig 1. Visualizations of 3003 incident types by victim category and type of harm. These visualizations are provided by the [AI Incident Database](https://incidentdatabase.ai/) and the [Center for Security and Emerging Technology (CSET)](https://incidentdatabase.ai/taxonomy/cset/)\n\n","metadata":{}},{"cell_type":"markdown","source":"\n### A Pause on Development?\n\nOne potential way to address these many issues of how best to instill positive values in AI applications in society that has been [called for by experts](https://www.safe.ai/statement-on-ai-risk) is to simply [pause AI development](https://futureoflife.org/open-letter/pause-giant-ai-experiments/) until we get a handle on these ethical (and existential) issues. \n\nHowever, policing or stopping AI in the development phase is simply not logistically feasible. Pandora’s box has been opened, the models and data are out there, and, try as we might, we cannot regulate what is on private machines. There can be no policing what models and data people do and do not have access to. Additionally, although we place limits on allowable actions one can take private, asking for editorial access to the machines of all AI developers would be an _extreme_ breach of privacy.\n\nPutting a pin in AI development in order to sort through ethical issues is also politically impossible from a _security_ standpoint. Just as nuclear development was not halted even once its gravity was realized, if one country will not stop learning how to [weaponize AI](https://foreignpolicy.com/2023/03/27/us-china-ai-competition-cooperation/), neither can any other. We are now in an [AI Cold War](https://online.ucpress.edu/gp/article/2/1/24803/117647/Is-There-an-AI-Cold-War). To pause development would incur a tremendous threat to national security which no country will accept. \n\nSimilarly, private companies will not stop AI development and risk getting behind the competition. While some pay lip service to the idea of [“Ethical Development,”](https://idiro.com/when-ethics-end-and-lip-service-begins/) they openly admit that free market [self-regulation is not sufficient](https://www.cato.org/commentary/worst-possible-reason-support-new-ai-regulation) to address new ethical issues as they arise. However, the risk of not addressing potential ethical concerns may also come at a [high price](https://www.cio.com/article/350290/6-business-risks-of-shortchanging-ai-ethics-and-governance.html), and potentially even lead to the dreaded [AI winter](https://telefonicatech.com/en/blog/what-is-ai-winter-and-how-to-avoid-it). ","metadata":{}},{"cell_type":"markdown","source":"\n### Regulation\nLegislative interest in addressing AI via policy is on the rise, to say the least. [The Global AI Index Report](https://aiindex.stanford.edu/report/) states:\n> \"Of the 127 countries analyzed, since 2016, 31 have passed at least one AI-related bill, and together they have passed a total of 123 AI-related bills... from 2016 to 2022, there has been a sharp increase in the total number of AI-related bills passed into law, with only one passed in 2016, climbing to 37 bills passed in 2022\"\n\n\nIndeed, a recent [avalanche](https://www.progressivepolicy.org/blogs/an-overview-and-of-global-ai-regulation-and-whats-next/#:~:text=The%20AI%20Act%20is%20a,technologies%20are%20not%20considered%20AI.) of policy proposals, initiatives, [ethics guidelines](https://www.nature.com/articles/s42256-022-00585-2), and [ethical compliance certifications](https://www.responsible.ai/_files/ugd/f70099_fa2d0ef6721c4da48e2245fa1d9a3693.pdf), by all manner of [governments, companies, non-profits, think tanks, and watch groups](https://www.nature.com/articles/s42256-019-0088-2), pose suggestions for both governmental and self-regulation of AI development and applications in an attempt to mitigate the impact of these ethical concerns. In April 2023 the EU Proposed the [AI Act](https://artificialintelligenceact.eu/), which mandates audits of “high-risk” AI systems. The European Commission put forth [guidelines](https://digital-strategy.ec.europa.eu/en/library/ethics-guidelines-trustworthy-ai) on “trustworthy AI,” and the European High-Level Experts group on ethics and the Science and New Technologies Group suggest requiring [“meaningful information of the logic involved”](https://academic.oup.com/idpl/article/7/4/233/4762325) any time an AI algorithm makes decisions that impact human wellbeing. Internationally, China proposes [forward-looking](https://www.axios.com/2023/05/08/china-ai-regulation-race) prevention, although their development recommendations diverge from many of those provided in the West – perhaps as a result of viewing the problem through the lens of a collectivist rather than individualist society. And yet, the identified [risks AI poses to society](https://www.pbs.org/newshour/world/china-warns-of-artificial-intelligence-risks-calls-for-increased-national-security-measures) are similar to those [identified in Europe and the United States](https://aiindex.stanford.edu/wp-content/uploads/2023/04/HAI_AI-Index-Report_2023.pdf).\n\n\nIt is still unclear how best to [justify the implementation](https://arxiv.org/abs/2302.02337) and [validate the success](https://www.technologyreview.com/2023/05/22/1073482/our-quick-guide-to-the-6-ways-we-can-regulate-ai/) of policy proposals pertaining to AI. Collecting metrics to measure impact is useful, but issues of bias again become apparent – namely, who collects the metrics, how, and to what end? And, even if new legislation isn’t necessary, many technologists agree on a desperate need for [education for lawmakers](https://www.nextgov.com/emerging-tech/2023/06/congress-learning-mode-about-ai-lawmakers-say/387287/). ","metadata":{}},{"cell_type":"markdown","source":"\n### Data\n\nMitigating privacy and bias issues within AI's training data is another topic for potential legislation. [General Data Protection Regulation](https://gdpr-info.eu/) requires one to be able to “own” their own data, and have it deleted if requested. However, once data is given away, it is extremely difficult to track who [owns or has access to it](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3987369#:~:text=Abstract,personally%20identifiable%20information%20(PII).). Similarly, GDPR requires data to be “anonymized” or non-identifiable. However, recent research shows that “anonymized” data is, well, not anonymous, either through [exploiting design loopholes](https://www.bleepingcomputer.com/news/security/strava-heatmap-feature-can-be-abused-to-find-home-addresses/) or by the simple fact [de-anonymizing data is technically achievable](https://cdn.iiit.ac.in/cdn/precog.iiit.ac.in/pubs/SaurabhGupta_ComprehensiveVivaReport_July2022.pdf). Additionally, in the US, all publicly available data is considered [free to use](https://techcrunch.com/2022/04/18/web-scraping-legal-court/?guccounter=1&guce_referrer=aHR0cHM6Ly93d3cuZ29vZ2xlLmNvbS8&guce_referrer_sig=AQAAAFQ3WFqJxaAVvdhF9lb0fmvuYTU_AzpnkQOvvUjkVm-H1QVlo-EwQiyHIDcmOSG2UPUITrJoJrMDzWVlGhNiyVFtoapw_nEk4wGNuFAB0HZQ9Nwhps5u4nXrD7dzYiO8ebtLblQgvOxD_uOebWW9kUlL7OVv7xBu3ZpmU5WG745c) to train AI models. Thus, although attempts to maintain _privacy_ and _security_ have been made, these values remain compromised with current regulations.\n\nAnother remaining issue is how (or if) legislators should address bias in data. While it is true that AI algorithms perpetuate or exaggerate, rather than cause bias per se, we return to _justice_ as a core ethical value of society. AI bias may be unintentional and non-malicious, yet it remains ubiquitous. This is such a concern that private companies have become interested in solving this problem. Google recently proposed the [machine “unlearning” challenge](https://ai.googleblog.com/2023/06/announcing-first-machine-unlearning.html) specifically to address biased and poor quality training data. It appears the answer to the question \"who decides what is _fair_?\" is being answered with \"the tech industry.\"\n\nShould legislators require a benchmark of data quality? Or adopt a uniform tooling system to measure and publish data bias and quality to increase algorithmic transparency? These are questions to which the answer will elucidate how we continue to define Justice in society. ","metadata":{}},{"cell_type":"markdown","source":"\n### Explainability\nThe recently passed EU AI Act ensures a “right to explanation” to anybody for whom an algorithm is used to make a decision that “impacts their life.” A major progression of this act is that it actually requires [notifying someone](https://www.brookings.edu/articles/key-enforcement-issues-of-the-ai-act-should-lead-eu-trilogue-debate/) if they were impacted by a \"high-risk\" algorithm -- which was previously not the case. However this requires a “reasonable explanation” to be possible – which is an [active area of technological research](https://www.sciencedirect.com/science/article/pii/S1566253523001148) within the deep-learning community. Explainability can happen at the level of the model, or the output – all of which pose their own challenges to demanding interpretability through legislation. \n\nRegardless, even if explanations are possible, private companies may not be interested in providing them, as they could leak valuable information about proprietary models or datasets. In this case, as a society we must decide what is more important: corporate competition, or the right to know exactly how corporations affect our lives. ","metadata":{}},{"cell_type":"markdown","source":"\n### Further Challenges to AI Policy\n\nA plethora of issues continue to plague well-meaning attempts at AI Policy. As AI rapidly evolves, it is an open question whether policy should continue to be traditionally reactive, or if we must instead adopt a pro-active approach. \n\nAI development also raises concerns about _inclusivity_ and _diversity_ of ideas. Currently, most people at the policy-making table are experts in either legislation or AI. The circles with the power to legislate AI tend to consist of either humanities experts or “hard” scientists. These groups present dangerously siloed and needlessly exclusionary boundaries in the context of ethical governance of new technologies. They exclude a huge number of stakeholders, from business leaders, to climatologists, to working technological laypeople whose lives will inevitably be meaningfully altered by policy decisions. It is widely recognized that for true democratization of AI – indeed, for fairness and justice – more diverse representatives must be involved in shaping AI policy, all of whom have [differing perspectives and expectations of how AI will shape the future](https://rethinkpriorities.org/publications/us-public-opinion-of-ai-policy-and-risk). \n\nFinally, it is worth consideration that climate legislation with nominally little to do with AI development are likely to intersect with ongoing definition of the societal values of _stewardship_ (or perhaps _protection_) and _progress_. While the environmental cost of maintaining and developing AI algorithms [becomes astronomical](https://www2.deloitte.com/nl/nl/blog/sustainability/2022/artificial-intelligence-worth-the-environmental-cost.html)[\\*](https://arxiv.org/ftp/arxiv/papers/2104/2104.10350.pdf)[\\*](https://arxiv.org/pdf/2303.01980.pdf), AI evangelists are quick to point to the many positive uses of [“Sustainable AI”](https://link.springer.com/article/10.1007/s43681-021-00043-6)[\\*](https://www.thedigitalspeaker.com/greener-future-importance-sustainable-ai/) in fighting against the climate crisis (e.g. [improving wind farm operations](https://www.windsystemsmag.com/ai-applications-in-wind-energy-systems/), [reducing energy consumption](https://www.maximpact.com/artificial-intelligence-in-energy-efficiency/), and [protecting endangered species](https://rfcx.org/impact)). It remains to be seen whether the energy usage of AI algorithms are worth their potential positive environmental effects, and how explicit discussions of values will play into discussion. ","metadata":{}},{"cell_type":"markdown","source":"\n# Bringing it Home\n\nAI Ethics is a complicated constellation of interleaving issues, none of which can be fully disentangled from any other. As rapid development ensures AI technology impacts more and more people, it is imperative to understand, develop, and regulate AI in such a way that we preserve our moral and ethical standards in society. \n\nAI development does indeed throw into question our societal values, even as it presents new avenues through which to examine and realize them. What we now must decide is: in the Age of AI, how do we _continue_ to enact those values? There is no clear answer to this question, but perhaps by working together and explicitizing what is most important to us, we can steer the trajectory of AI development towards the benefit of us all. ","metadata":{}},{"cell_type":"markdown","source":"| Value | Application to AI | Potential Mitigations | Related Concepts\n| --- | --- | --- | --- | \n|humanity (personhood) | Evolving technologies enhance machine capbilities. When AIs _seem_ \"conscious,\" what does it mean to be human? | Treat the machines like you would want to be treated; Transhumanism: Embrance the cyborg future, baby! | existence, justice |\n|existence | It's nice to be alive. Some believe AIs have the potential to cause humanity's [extinction](https://link.springer.com/article/10.1007/s00146-018-0845-5) | Work on \"Kill Switches\" to prevent the robo-apocalypse; continued work on national security | security, humanity, power |\n|work (productivity) | GenAI and advanced algorithms automate more and more human labor tasks. What's left for us? | \"Reserve\" tasks for humans (creative arts); Allow leisure and uproot societal pressure to be \"productive.\" | trust, power, humanity |\n|trust | We are used to believing what we hear and see. How can we maintain trust when bad actors using AI generate convincing fakes? | Education on AI's capabilities and limitations; continued work on watermarking/authenticating | work, power, security, privacy |\n|fairness (justice) | Do AI algorithms make decisions that reflect treatment we want across all peoples affected by those decisions? | Reduce bias in training data; explicitly program in \"fairness\" criteria according to codified human-understandable principles | trust, power, privacy |\n|inclusivity | Who decides when and how AI is used to impact people's lives? | Legislative bodies continue technical education; future-proof guidelines that bake-in iterative review of evolving technologies; wide variety of stakeholders at the table | work, trust, justice, power |\n|power | Opportunity begets opportunity. Will AI accelerate growing economic and class divides? | \"Democratize\" access to AI algorithms and resources; explicitize and implement programs to widen opportunities to interact with and understand AI across demographics at local, federal, and global levels | justice, work, humanity |\n|security | AI can be weaponized at a national and private level. How will we protect ourselves from one another? | International guidelines on AI development and usage | privacy, power, existence |\n|privacy | Training data is often publicly available despite protest. What can individuals do to ensure data ownership and our right to privacy? | Consumer protections of data rights; unified standards of data quality; transparent data benchmarking | justice, power, inclusivity, fairness |\n\n