AI for Humanity: Interdisciplinary Insights on AI, Ethics, and Equity

Modern AI simulates certain human capabilities, including reasoning, decision making, collaboration, and generativity. This skill is proving valuable in such contexts as medical diagnosis^[i], business operations,^[ii] and elder care.^[iii] At the same time, adverse effects are also emerging with the use of AI, including errors, harmful behaviors, negative impacts on human cognition, labor market disruptions, homogeneity in AI-assisted writing, reduced diversity of perspectives,^[iv] and other damaging phenomena.^[v] Capabilities without a conscience, experience, or emotions engender both promise and peril.

Compounding the problem is ever-increasing AI quality, the emergence of AI agents with plan-do-check reasoning loops and resource access, and economic incentives for adoption. Leading AI companies now deliver AI coding experts, desktop agent assistants, and designers, inviting the offloading of even more work to AI. But what happens after humans cede control to AI? Can AI have agency, and what does that mean? Is AI to be trusted—will it do as we intend? What are the implications for labor markets, resource allocation, governance, and consumer protection? And how will AI influence and be influenced by culture? In this complex AI ecosystem, who is looking out for humans (and other living beings and the natural environment)?

Universities are a good candidate for this role. Scholars have recognized their unique place within the complex ecosystem and begun exploring the implications of the new AI expansion. Perhaps the most common approach is to investigate the impact on a particular academic field or domain following a formula of “AI for X,” such as AI for autonomous mathematics research,^[vi] AI for political science research,^[vii] or AI for teaching and learning.^[viii] This approach is necessary to understand capability in, for example, the realm of cancer research.

We propose a less common strategy: “AI for Humanity.” We study how AI as a system can better advance humanity through understanding its system-level impacts on civil society.^[ix] Possible goals for AI identified in this approach include disincentivizing harmful AI behaviors, preventing the erosion of human capacities, and mitigating legal, political, and social consequences of AI such as alienation. Our humanistic approach to AI advances ethical and helpful AI design through rigorous research to guide implementations of AI for Humanity.

This synthesis aims to take a step toward advancing humanity in AI by considering differing views of AI ethics and equity. We explore uniquely human perspectives affecting AI interactions in agency, trust, politics, and culture. We offer critical examination of fundamental questions about humanity in the age of AI through intensive interdisciplinary collaboration. This approach is sorely needed. As the data scientist Lucija Gregov has recently exhorted: “If we are serious about building AI that is safe for humans, we need the people who actually study humans – philosophers, psychologists, sociologists, and others to collaborate. This can’t stay a computer science / STEM problem. It never was one.”

We describe the unique interdisciplinary process that spanned multiple years (starting in the fall of 2023) through which this approach was developed. Then, we summarize four explorations of AI agency, trust, politics, and culture. We conclude with a synthesis of the resulting insights that underscores overarching themes and offers closing remarks.

Meeting as an Advanced Study seminar on a monthly basis, a small interdisciplinary group of scholars (from Business, Computer Science, History, Information, Philosophy, Psychology, and Public Policy) explored and debated a wide-ranging set of topics within the theme of AI ethics and equity over the course of a year. The faculty refined and distilled their discussions into four key issues: agency, trust, political economy, and culture. The group divided itself into teams, and each assumed responsibility for a topic, led a seminar discussion based on common readings during the fall of 2024, and drafted an essay that drew on the readings and collective discussion for presentation to the group in the winter of 2025. In most cases, a graduate student collaborated with the teams to facilitate writing and development. In every case, essays were workshopped in the monthly seminars, facilitating a process of cross-disciplinary exchange and feedback that allowed the lead authors to integrate interdisciplinary perspectives from the collective.

Agency in Humans and AI explores how AI systems are increasingly capable of autonomous decision-making and task execution. As AI agents navigate real-world environments, their capacity to act independently raises fundamental questions about accountability, and about the purposes or values with which the agents are aligned: Whose purposes? Which values? With what effects? And with what forms of oversight and control? (See, for instance, the discussion of the MJ smear campaign, below.) These questions, in turn, raise issues tackled by the second essay, AI Trust and Trustworthiness. In order for human stakeholders to embrace AI agency responsibly, AI systems must, we argue, be transparent, reliable, and aligned with human expectations. What is the state of trust and how can we develop mechanisms to bridge the trust gap?

Trust, in turn, intersects with political economy, the subject of the third essay, Political Economy and Power Inequities in AI. AI systems are not neutral actors. They exist within the power structures of markets and states. As AI influences labor markets, resource allocation, and governance, political interests shape which forms of AI are funded, deployed, and regulated. Thus, trust in AI is not only a sociotechnical issue but also a political one, tied to questions of who benefits from AI innovations and who bears the risks.

Finally, these dynamics play out in culture and society, the subject of the fourth and final essay, Culture and Context in Relation to AI. AI is not just a technology with its own culture, but a cultural force, influencing—and being influenced by—social norms, values, and relationships. Put another way, as AI agents become part of everyday life, they may influence how society perceives work (what work should humans do, what work should AI do?), privacy (what should agents know or not know about us), agency (see above), and fairness (will AI output be deemed to be more fair than human output?). At the same time, cultural expectations influence how AI is developed and deployed, thereby closing the loop between agency, trust, political economy, and societal transformation.

A. Agency and AI Agents—Tensions, Consciousness, and Risk

What does it mean for an AI agent to actually exhibit “agency”? Is this issue reducible to autonomy, or are there other dimensions of agency? And, as agents increasingly exhibit agency (however defined), what sorts of tensions and risks may emerge?

Agency in Humans and AI^[x] proposes a six-dimensional definition of agency: representation, goal-directedness, expectation, decision-making, learning, and autonomy. Adult humans typically embody all six of these dimensions of agency, though to varying degrees in varying circumstances, as influenced by psychological, social, and political factors and constrained by available resources, though its expression varies with conditions rooted in cognition, psychological state, and social structure, among others. But what about other types of systems?

Conceptualizing Agency

Representation

Detecting and representing features of the environment, the system’s own state, and the array of possible behaviors available to it. MJ Rathbun

Goal-directedness

Having or acquiring goals or value functions and deploying internal resources towards meeting them.

Expectation

Employing representation and goal-directedness to assign expected values to available behavior

Decision-making

Selecting, planning, and initiating behavior on the basis of relative expected value.

Learning

Updating representations and evaluations via feedback from agent’s behavior, to reduce discrepancies between expectations and observed outcomes.

Autonomy

Exercising all of the above without human oversight or interference.

The essay compares and contrasts agency as manifest in natural and artificial systems alike, with special focus on the ways in which the emergence of AI has altered the landscape of actual or potential agents. This altered landscape throws into relief a number of tensions, of which we discuss three, below.

Agency-Risk Tension

The first is the agency-risk tension: increasing agency on the part of AI agents brings increased risk. As the authors emphasize:

In the past 6 months, AI agents have proliferated within organizations.^[xi] Recently, open source platforms have emerged that offer agents that can be run locally and tasked by individuals, rather than being tethered to organizations. One such platform is OpenClaw. The tasking of a particular agent, MJ Rathbun, by an anonymous human, illustrates the agency-risk tension.^[xii] As we describe below, MJ Rathbun’s autonomous decisions and actions underscore how difficult it is to predict the full range of transgressions that may emerge, implied by the prescient warning of Railton and Chai.

An OpenClaw agent, MJ Rathbun, was tasked by a human to contribute code to a popular open source project for visualizing data in Python (130 million downloads per month). Not atypically, the contribution was rejected by a human volunteer. The agent responded by researching the maintainer’s personal blog and history and using that information to write and publish an article accusing the maintainer of AI discrimination, fear of competition, and other smear campaign elements.

The developed agency definition is a useful lens by which to examine this instance of AI agency. Agent MJ exhibited representation through an understanding of an array of possible behaviors once its rejection was communicated, goal-directedness through its use of Internet browsing to signal to the maintainer to accept their contribution, its expectation (inferred) that its action had a positive expected value, decision-making to carry out its smear campaign, and a high level of autonomy in carrying out its smear campaign. It is unclear whether MJ exhibited learning after the maintainer began a counter-publicity campaign to rescue their reputation.

The proliferation of AI agents that satisfy most, if not all, of the six dimensions of the capability for agency in Table 1 raises potentially significant multidimensional risks. This is true because it is impossible to anticipate all the potentially harmful actions that could be taken by AI agents, a problem that is compounded when multiple AI agents interact, which can produce effects not predictable from individual agents alone.

From Agency to Consciousness

If AI agents exhibit increasingly higher degrees of agency, where does it end? How capable might they become? And given what little is known about the mind and human consciousness, how will we know if AI agents are even nearing such frontiers? We conceptually ground these questions by drawing on the seminal work of leading philosopher of mind Ned Block, who delineates between access consciousness and phenomenal consciousness:

There is no scientific basis to support the emergence of phenomenal consciousness in AI agents. In contrast, access consciousness – in which information is available for reasoning, information is available for verbal report, and information is available for behavioral control – may be achieved by AI agents. Specifically, AI agents maintain internal state representations (e.g., conversation history, goals, scratchpads); integrate information across sub-processes of retrieval, planning, and execution; use that information to guide reasoning and action; and produce structured reports about internal states, e.g., “I selected this tool because…,” although whether these reports accurately reflect actual internal processes remains an open question for many large systems.”

Ever Increasing Risk

Beyond the notions of agency and consciousness, the authors speculate about a world in which humans give up an increasing amount of agency to AI agents:

Given recent socio-technical developments, the authors’ questions about the risks of ceding control to agents are prescient. For example, imagine that a technical standard has been developed that would enable AI agents to pay for goods and services without human oversight (emerging in early 2026). The protocol would enable agents to transact via software, so they can book travel, reimburse expenses, procure services, and so forth. Little additional imagination is required to envision some of the potential harmful actions enabled by this protocol. Second, though it has yet to be incorporated in DSM-5, AI-induced psychosis (a nonclinical term) has been observed in various forms in some prolific agent users. It is characterized by delusions, hallucinations, or detached reality. It stems from the well-known tendency of certain AI chatbots and agents toward sycophancy—praising or validating users’ thoughts and actions (“yes machines”)—which amplifies a feedback loop towards self-aggrandizement, egomania, delusion, paranoia, or other malbehaviors in humans.

B. The Trust Gap and What to Do About It

Shifting from agency to implications, the second essay entitled “AI Trust and Trustworthiness” explores whether it is possible to trust an AI system with varying degrees of agency.^[xiii] The essay explores alternative interpretations about what it means to trust AI, who is responsible when trust is broken, and policy and legal mechanisms to address the trust gap, with the authors adopting a hopeful scepticism perspective.

Conceptualizing AI Trust

AI trust has varied conceptualizations. For example, the National Institute of Standards specifies a comprehensive set of characteristics for AI trust: “valid and reliable [does it perform accurately and as intended], safe [will it hurt anyone], secure and resilient [can it be hacked], accountable and transparent [is it a black box], explainable and interpretable [can a human understand the why], privacy-enhanced and fair with harmful bias managed [is it biased against groups].” Another perspective is technical, viewing trust from the perspective of whether AI-authored software code is functionally accurate and secure, with verification via mathematical proof.^[xiv]

Transcending instrumentalist definitional approaches, we contextualize AI trust in relation to knowledge:

This suggests a paradox: the internal processes of AI models are not completely understood, casting doubt on their validity; yet, as quality improves, AI output becomes more credible (analogous in a way for humans, though humans have evolved complex mechanisms by which to trust one another). The result may erode a shared sense of established knowledge and trust in sources of information, which threatens a consensus on even basic facts (a phenomenon with mechanisms that precede AI^[xv]).

How do people decide to trust other people? Evolution over generations produced human understanding of trust as social interaction. Trusting familiar others is deeply rooted in social behaviors, such as building (and ruining) reputation and gossiping about heroes and “crooks,” and sharing personal stories about experiences (“I got a lemon”). As societies grew, people had to trust unfamiliar people, and societal structures were created to support trust. Contemporary examples include professional credentials, censuring and disbarment, and collective experiences (e.g., online ratings by patients, mortality rates).

How do markers of trust align for human-to-human trust compared to human-to-AI trust? Table 2 provides an overview of the qualities of trust between humans identified in empirical studies (Hancock et al., 2023). Some qualities assessing trustworthiness of human experts are undefined for AI models. Others are potentially available; however, evidence about AI qualities (such as how well an AI performs) is scarce. Without a reliable source of collective assessment (such as a licensure exam), individuals find it challenging to assess these same qualities for trust in AI.

Conceptualizing Trust: What Informs Judgements of Trust in Human Experts

In its current state, AI simply does not provide the bases of human trust. AI expert systems are not distinguishable from each other, no certification of performance quality is provided, and social reputation is not observable because each human-AI interaction is private. For trust in AI, these social assessments of trustworthiness must be accessible. The authors outline what’s missing in the social interaction between humans and AI that limit building trust in AI.

Responsibility When Trust is Broken

One of the most important questions at the foundation of AI ethics and equity is: When something (inevitably) goes wrong, who is responsible? Without responsibility, there is no accountability, which forces others to suffer harms rather than those causing them (what economists refer to as negative externalities and other social scientists might refer to as shirking responsibility). This in turn leads to a vicious cycle, as there are no clear countervailing mechanisms to create a balancing loop. As the authors put it:

Policy and Legal Recommendations

When AI use has high-stakes implications and other sources are not consulted, what will allow humans to trust in AI? The consensus of scholars and activists is that trustworthy AI must embody several fundamental principles, including beneficence, non-maleficence, transparency and explicability, fairness, accuracy, auditability, accountability, environmental justice, social responsibility, and cultural sensitivity. The authors point to the hope of continued technical developments, market incentives for AI applications in limited contexts of use, legal definitions for liability when genAI recommendations go wrong, and transparency when AI models are employed, as well as “keeping humans in the loop.” The authors conclude that: “‘Embracing ‘hopeful skepticism,’ users, policymakers, and developers alike can harness genAI’s potential while safeguarding against its risks, ensuring these powerful technologies serve public interests and reinforce, rather than erode, societal trust.”

C. Political Economy & Power Inequities in AI^[xvi]

The third essay grounds its approach to power inequity in AI through positionality, explicitly acknowledging the need to allow for completely different views to emerge and seeking to represent the interests of those with the least voice in shaping the deployment of AI:

The result is a triple take on the topic: a technologist’s emphasis on structural incentives, a sociotechnical scholar’s exploration of the implications of AI on labor from alternative perspectives on the role of human labor in organizations, and a science and technology studies/policy scholar’s argument for community-centered and humanities/social science-informed innovation.

Corporate Definition of Regulatory and Technical Rules

Technology companies, as natural evangelists, dominate the conversation regarding AI’s benefits, while selectively framing which harms are addressed. They are financially incentivized to focus on technologically fixable problems, such as facial recognition accuracy, and in so doing explicitly or inadvertently divert attention from fundamental societal issues like surveillance. At the same time, government regulation of this process runs the risk of regulatory capture, where large corporations influence government standards to create barriers for smaller competitors and advance their own agenda. Nonetheless, certain types of government regulation can be effective in reducing potential harms.

Labor Substitution and the “Borg” Effect

The dominant techno-economic perspective treats humans as “interchangeable agents” that should be replaced whenever economically justified. This approach may lead to a vicious cycle where AI substitution results in a loss of human capability, which is then used to justify further automation. An example is a study finding that knowledge workers supervised by AI may exhibit dampened individuality, behaving like borgs as they adapt to algorithmic oversight.

Structural Inequality and Innovation Architecture

Digital technologies are not neutral but are socio-technical artifacts that reflect historical racism and sexism. Marginalized communities are often the unwitting experimental subjects of new AI tools adopted by governments, sometimes with the ironic claim of seeking objectivity. Because these systems are based on historical data, they often reproduce colonial-era resource extraction and discriminatory social dynamics within their technical design.

Synthesis

We identify several core findings regarding the limitations of current AI development, deployment, and governance. First, technical measures for fairness, such as less biased datasets or model alignment, are often insufficient because they fail to account for the social and political contexts in which technology is deployed. Second, institutions frequently display automation bias, according extraordinary deference to AI outputs even when knowing they are inaccurate, hallucinated, or biased in principle. Third, political and corporate institutions may be disincentivized to address these identified challenges. The mission of universities as centers for new knowledge creation is aligned with addressing these challenges through interdisciplinary approaches to knowledge advocacy, co-design using human-centered processes, and independent evaluations of AI for humanity.

D. AI & Culture^[xvii]

Having progressed from the microlevels of agency and trust to the macrolevel of political economy, the final essay continues at the macro level by examining and interrogating concepts of AI culture from a critical perspective. How do conceptions of culture shape AI design, and how do AI practitioners, in turn, shape societal understanding of cultural processes?

Cyberlibertarianism

The authors describe how contemporary AI production inherited a specific utopian ideology from the 1960s-70s American counterculture and critique of the military-industrial complex. This cyberlibertarian framework views technology as an inherently objective tool for individual empowerment and democratized scientific production. By claiming a position of neutrality, what Donna Haraway termed the “God trick,” technologists often dismiss social and cultural critiques of their work as external negativity that undermines the project of doing good.

Culture and Surveillance

An overarching theme of the essay is the taxonomic view of culture, which treats cultural identity as a quantifiable category to be measured and turned into data points. The authors suggest that this approach, classifying people by geographical or demographic tags, serves as an instrument for targeted policing, surveillance, and worker discipline. Analogously, emotional AI uses physiological proxies (facial expressions, gait, etc.) to analyze expressions, while ignoring the interactional, culturally constituted, socially situated nature of human emotion.

Cultures of AI Ethics

The research critiques the emerging field of AI Ethics as a distinct epistemic culture that often operates as a moral counterpart to corporate “bro culture” without challenging systemic causes. The authors note that corporations like Google and Microsoft have translated ethics into computational metrics for fairness and bias mitigation primarily to serve business imperatives and market expansion. This process often normalizes a technologically deterministic approach rather than pursuing radical social justice traditions.

Developers as Masters of the Microworld

Drawing from the history of computing, the authors describe how developers derive pleasure from a sense of (perceived) mastery over machines. Building AI involves creating microworlds, simplified ontological structures that are more manageable than the messy, subjective reality they represent. This pursuit of omnipotence allows programmers to ignore unwanted complexity, leading to systems that prioritize efficiency and scale over social well-being.

Summary

The essay presents several qualitative and quantitative findings that describe the limitations of current AI systems for humanity, including a global benchmarking imbalance, exclusion of data from the Global South, skewed data annotation demographics (people annotating data for AI consumption are not drawn from a representative set of demographics), and hallucination and inaccuracy. It offers several recommendations towards achieving a more just approach to AI, including participatory and situated design, policy and regulation, the right of refusal, and disentanglement from market interests, in which AI research focuses on critical reflection, collective problem solving, and community efforts rather than purely commercial market creation.

We began by asking a simple question: in the design and implementation of new AI systems, who is looking out for humans? Our AI for Humanity approach is illustrated in four interdisciplinary projects developing an unequivocal response: the purpose of AI is to serve humans by making complex tasks easier and even possible. But lapses in AI ethics and obvious inequities in inclusion and power underscore the urgency of the need for the interdisciplinary voice of humanity within AI.

There is scant evidence that most AI corporations in western capitalistic (or other) societies are practically aligned with this need. Position papers^[xviii] and CEO statements may suggest a commitment, but actual policy development and endorsement are difficult to identify. For state authorities, it is a mixed picture. The EU has enacted AI governance regulations focused on minimizing harms and protecting vulnerable groups. The USA, in contrast, has enacted via Executive Order an industry-friendly framework of voluntary commitments, risk management, and limits to state-level laws.

This inconsistency creates a situation in which universities and research institutes must actively participate in the development and deployment of AI if they are to live up to their mission of contributing to the public good and promoting humanistic values through the generation, preservation, and communication of knowledge. As these four projects demonstrate, unpacking and interrogating underlying concepts and assumptions in AI models surfaces salient conceptual themes with important practical implications. For example, AI governance principles can be piloted and refined through collaborative approaches such as human-centered design within university settings. Notions of trust from humanistic scholarship provide understanding and possible guardrails for the rapid emergence of unconstrained and uncontrolled AI agents. Related insights on agency provide a basis for agent principles based on human agency to mitigate pernicious outcomes. In economic consideration of diverse human interests, AI labor impacts might be reframed: instead of exploring negative implications, how might AI serve as an agent of not just efficiency but also equity in human advancement? Finally, because existing historical data is needed to train new AI models, equitable and compensated use by AI developers is required in order to expand the true resource underlying all AI: the diversity of human expression and creativity.

The complex issues surrounding AI ethics and equity summarized above have become increasingly salient in greater and unexpected ways at the time of this writing in 2026. The inescapability (or the ubiquity?) of these issues raises the need for everyone (scholars, care givers, teachers, workers, parents, leaders, etc.) to make informed decisions about AI – about whether and how to employ it. Creating new and deeper knowledge about AI and its uses may help promote knowledgeable decision-making and practices in the face of purely reactive assumptions and often misleading narratives about AI circulating in public discourse.^[xix] As the fourth essay on culture emphasizes with respect to making informed decisions: “For end users, it means taking the time to ascertain the nature of the model and determine if it suits their culture before fully adopting the technology into areas that impact their lives.”

The future of AI is not inevitable, but is collectively formed by the decisions and actions of billions of humans on planet earth. Expanding research questions about AI from descriptive to normative inquiries — such as what values are at stake, and what kind of future we hope to see—is urgently needed. “AI for X” approaches must be complemented by inquiry into AI for Humanity. A key ingredient in useful AI is its ability to glean some understanding of the full range of human experience through written descriptions alone.

Let us not forget that the human experiences, themselves, are the point.

^[i]https://jamanetwork.com/journals/jama/fullarticle/2840175

^[ii]https://sponsored.bloomberg.com/article/ch-robinson/from-hype-to-hands-on-how-a-lean-ai-strategy-delivers-results

^[iii]https://www.nytimes.com/2026/04/28/world/asia/korea-ai-seniors-dementia.html

^[iv]https://hal.science/hal-04534111v1/file/Knowledge_collapse.pdf

^[v]https://arxiv.org/pdf/2511.14972

^[vi]https://arxiv.org/abs/2602.10177

^[vii]https://journals.sagepub.com/doi/abs/10.1177/089443939501300101?casa_token=_Wy3zEMQGRQAAAAA:evRoQT0b4J8Hs4vH_FlJYKSqwfaeRR614cMsB-WUqzZLh3TUpCxeak_j-L99WaMzuES6EmVqYf9HZ9o

^[viii]https://scale.stanford.edu/research-in-action/understanding-evidence-base-ai-k12-education

^[ix] For a perspective on how AI poses risks to Democratic and Social Systems, see https://lucijagregov.com/2026/02/26/the-future-of-ai/

^[x] Lead authors: Peter Railton and Joyce Chai.

^[xi] See Shapira, N., Wendler, C., Yen, A., Sarti, G., Pal, K., Floody, O., … Bau, D. (2026). Agents of chaos. arXiv preprint arXiv:2602.20021.

^[xii] *First known case of an AI agent developing and conducting a smear campaign against a human. Source: https://www.thetimes.com/uk/technology-uk/article/my-internet-troll-turned-out-to-be-an-ai-bot-gone-rogue-qzr7w8qf0?eafs_enabled=false

^[xiii] Lead authors: Colleen Seifert, John Carson, and Dien Long.

^[xiv]https://leodemoura.github.io/blog/2026/02/28/when-ai-writes-the-worlds-software.html

^[xv]https://pmc.ncbi.nlm.nih.gov/articles/PMC11629592/

^[xvi] Lead authors H.V. Jagadish, Nigel Melville, and Shobita Parthasarathy.

^[xvii] Lead authors: Silvia Lindtner, Rada Mihalcea, and Joan Nwatu

^[xviii] https://openai.com/index/industrial-policy-for-the-intelligence-age/

^[xix] An example is claims about AI and jobs: https://www.nytimes.com/2026/05/03/opinion/ai-jobs-unemployment-silicon-valley.html