By Mathias Risse, Harvard University
The views expressed below are those of the author and do not necessarily reflect those of the Carr-Ryan Center for Human Rights or Harvard Kennedy School. These perspectives have been presented to encourage debate on important public policy challenges.
It has become a familiar refrain: we are not conceptually prepared for artificial intelligence (AI). We throw around words like intelligence, understanding, autonomy, and consciousness as if their meanings were settled. We borrow ethical frameworks built for embodied, mortal social creatures and apply them—sometimes carelessly—to systems made of code and weights. We use legal categories designed for human agents and traditional products to govern systems that generate novel outputs, act at scale, and shape human lives in ways their designers cannot fully predict.
The standard remedy is also familiar: we need better definitions. Less anthropomorphism. Cleaner taxonomy. New terms. More precise debates.
All of that helps. But it is not the heart of the problem.
The deeper difficulty is that the philosophical foundations that would tell us what our concepts really mean—and how they should be applied—are not merely “unfinished.” They are stuck in a way that is more structurally stubborn than we usually admit. And that structural fact matters directly for ethics and human rights, because so much of the emerging governance of AI depends on confident use of precisely those contested concepts.
To put the point plainly: AI is forcing foundational metaphysics and moral theory out of philosophy departments and into courts, legislatures, executive agencies, boardrooms, and international treaties. But the foundations are not ready for that assignment. Not because philosophers have been asleep, are missing in action, or miss capacities for sorting things out, but because the landscape itself is resistant to closure.
The real problem is not sloppy concepts. It’s a structural stalemate.
Consider how far philosophy has come over the last century. If you pick up contemporary work on consciousness, free will, moral status, personal identity, or knowledge, you find an impressive degree of clarity. We know the main positions. We know the best arguments for each. We know the strongest objections. The terrain is mapped.
And yet: on the questions that matter most for AI, philosophers of the first rank occupy every major position. Not because they’re confused. Not because they haven’t read the literature. Because the disputes are genuinely hard in a distinctive way: the arguments are strong enough on each side that no decisive refutation lands, and the remaining disagreements are not the kind you can reliably eliminate by sharpening definitions.
Call this structural stalemate: a condition in which (1) multiple positions on a foundational question remain rationally defensible, (2) none can be decisively knocked out, and (3) further conceptual refinement does not obviously force convergence. This is not nihilism. It is, in a certain sense, a mark of philosophical maturity: we now understand why the debates persist. It seems that finite beings can only do so much to make sense of their situation, and irreducible plurality in many domains is one feature of what they can do.
But for governance, it is destabilizing. Because the questions AI forces on us—Is the system conscious? Does it understand? Can it be responsible? Can it know?—are not brand-new questions. They are downstream of exactly these unresolved foundations. When we pretend we can answer them cleanly for AI, we often import—silently—philosophical commitments we have not earned.
That importation has consequences. Especially for human rights.
Human rights sound like the escape hatch. They aren’t.
Many leading AI governance efforts anchor themselves explicitly in human rights. That makes sense. Human rights offer a shared international vocabulary—dignity, equality, privacy, due process, freedom of expression, nondiscrimination, remedy—and a practical moral ambition: protecting persons from domination and abuse.
In a world of fast-moving technology and deep disagreement, human rights can feel like a stable platform. They can appear to solve a problem that ethics and philosophy cannot: we don’t need to settle metaphysics; we can just protect rights.
But that hope is only partly justified.
Human rights do real work. They supply constraints, institutional demands, and moral orientation. Yet the human rights framework does not float above philosophy. It presupposes contested philosophical foundations: What grounds dignity? What counts as harm? Who counts as a rights-holder? What does autonomy require? What is a person? What is fair treatment? What is “meaningful” consent? What does it mean to be free?
AI doesn’t allow us to bracket those questions. It drags them into the foreground. In doing so, it reveals a fracture in rights discourse that is both practical and conceptual.
There are two sides to that fracture.
First, AI threatens human rights in familiar ways: discriminatory decision systems that distort access to credit, employment, housing, or bail; opaque systems that undermine due process; pervasive data collection that erodes privacy; targeted persuasion that manipulates political agency; surveillance that chills speech and association. These are, by now, well-known concerns.
The second side is less often faced directly: AI may eventually force us to confront whether the moral community—and thus the rights community—could include entities that are not human beings. If an AI system were conscious in a morally relevant sense, or capable of something like rational self-legislation, or possessed interests that could be harmed, then the rights conversation changes. Even if you think that scenario is unlikely, the uncertainty itself becomes a governance problem. And because the foundations of consciousness and moral status are part of the structural stalemate, we should not expect that uncertainty to vanish on a convenient timeline.
So human rights are not an escape hatch from philosophy. They are an indispensable practice operating inside philosophical uncertainty.
The responsible question is not whether we can do without rights (we cannot), but how to make rights robust under conditions where foundational questions are unsettled and, worse, where our inherited conceptual categories may not fit what we are governing.
Conceptual insufficiency: it’s not that our categories are wrong—they may be inadequate.
Structural stalemate says: our concepts might be apt, but we cannot settle their correct application because the foundations remain disputed. There is a second problem that is distinct—and in some ways more radical.
Even if we somehow resolved the major debates about consciousness, agency, identity, and knowledge, we might still find that our inherited categories are conceptually insufficient for AI. Not mistaken. Not sloppy. Just not built for entities like these.
Many of our moral and legal concepts come with background assumptions—conditions of application—that were quietly shaped by the human case. A “person” is an embodied being with a continuous biological life, with a perspective, a history, and a future. Agency involves action in the world under conditions of vulnerability and stakes. Responsibility presupposes stable identity over time. Consent presupposes a subject who can understand what is being agreed to and can decline without being coerced. Knowledge presupposes some kind of relationship to reasons, evidence, and justification.
Now consider what AI systems are like.
They can be copied perfectly. A model can be instantiated in parallel. It can be fine-tuned, merged, expanded, pruned, paused, archived, and resumed. Its “memory” can be externalized. Its capacities can change discontinuously with new tools or new training. It can act at scale in thousands or millions of interactions simultaneously. It can shape human belief and behavior without a face, a body, or a social location that citizens can contest in familiar ways.
Even if none of that makes the system a person, it changes what it means to govern the systems’ effects. It strains our categories. Which entity is responsible when a “system” is really a supply chain: data collectors, model builders, fine-tuners, integrators, deployers, and end-users? What does it mean to remedy a harm when the model that harmed you is no longer the model in production? What does it mean to demand transparency when the “reason” for an output is distributed across a high-dimensional structure no one can interpret as reasons?
When regulators treat AI as “just a tool,” they risk underestimating its role in shaping agency and social power. When regulators treat it as “an agent,” they risk letting humans off the hook and creating accountability gaps. When they treat it as “speech,” they risk immunizing industrial-scale persuasion from democratic oversight. When they treat it as “a person,” they risk moral and legal displacement of human beings.
None of these misclassifications is merely conceptual. Each can become a rights problem: a due process problem, a discrimination problem, a privacy problem, a remedy problem, a legitimacy problem.
This is where the stakes of philosophy become visible. Not because legislators must read metaphysics, but because governance always imports ontology. Classifying what something is determines what can be done to it, what it can do to others, who can be held responsible, and what victims can demand.
Five foundational questions AI won’t let us postpone
To see how structural stalemate and conceptual insufficiency touch human rights, it helps to look at five foundational domains where AI is now putting pressure on inherited frameworks.
1) Consciousness and moral status.
If consciousness is fundamentally about functional organization, sufficiently sophisticated AI might be conscious. If consciousness depends on biological mechanisms, it might not be. If consciousness is something like integrated information, that yields a different test entirely. Philosophers disagree, and not in a way that looks close to resolution.
Human rights implications flow in two directions. The obvious one is that humans must not be treated as mere objects by systems that classify, score, and manipulate them. But the less obvious one is that moral uncertainty about AI status could be weaponized. Companies could claim “model welfare” to resist transparency. States could claim “AI rights” to complicate accountability. Meanwhile, if AI were ever conscious, treating it as mere property would be a moral catastrophe. Governance must therefore manage moral risk, not just current certainty.
2) Ethical foundations.
Consequentialists, Kantians, virtue ethicists, contractualists—each offers a powerful lens on AI, and none can claim decisive victory. Does targeted political manipulation violate dignity regardless of consequences? Or is it wrong because it produces social harms? Or because it cultivates vice and corrodes civic character? Or because it cannot be justified to those bound by it? All of these are plausible. And the differences matter when tradeoffs arise.
Rights discourse is often presented as a way around that pluralism. But rights themselves are interpreted differently depending on what you think grounds them: dignity, interests, political legitimacy, or agency. AI makes those differences operational.
3) Agency and responsibility.
Accountability is the bloodstream of rights. Without identifiable duty-bearers, rights become aspirations rather than enforceable claims. Yet AI systems diffuse responsibility: harms emerge from interaction of many components, and the outputs are not authored in the old sense. The temptation, especially among powerful institutions, is to say: the AI decided. That phrase can become a new kind of impunity.
Philosophically, what counts as an agent depends on contested theories of control and reasons-responsiveness. Legally, we can assign responsibility pragmatically. But if we do so without conceptual care, we risk building a regime where harms are systematic and remedies are nonexistent—precisely the opposite of a rights-respecting order.
4) Personal identity.
Rights and duties attach to stable subjects. But AI systems make identity slippery. Is a fine-tuned descendant “the same system” for liability purposes? Does the relevant “agent” persist through updates? If a harmful output came from a model version that no longer exists, what counts as remedy? Identity questions that were once mostly metaphysical become governance questions about traceability, audit, and enforcement.
5) Epistemology and the public sphere.
Democracy and rights depend on an epistemic commons: citizens must be able to form beliefs, evaluate reasons, and contest power. LLMs and synthetic media threaten that commons by scaling persuasion, flooding the space with plausible nonsense, and eroding evidentiary backstops. If you cannot reliably trust that a video is real, testimony and accountability degrade. If you cannot tell whether a political message has a human author accountable to anyone, persuasion becomes harder to contest. This is not merely an information problem; it is a condition-of-rights problem.
In each domain, philosophy is not an ornament. It is the hidden infrastructure of the categories governance uses. And in each domain, the infrastructure is unsettled.
The danger is not uncertainty. The danger is false resolution.
Governance must act under uncertainty. That is normal. The distinctive danger in the AI case is that practical urgency encourages premature conceptual closure: regulators and companies adopt working definitions because they need to move, and then those definitions acquire the social authority of settled truth. That is how structural stalemate gets replaced—not by philosophical progress—but by administrative convenience.
When that happens, a second danger follows: the space for philosophical challenge shrinks. Foundational questions start to look like academic luxuries precisely when their practical importance is at its peak. “We don’t have time for metaphysics” becomes the governing mood. But metaphysics has not disappeared; it has simply been smuggled in unexamined. The result is governance that is conceptually confident and morally brittle.
If you want a name for this phenomenon, call it philosophical abandonment under pressure: not the legitimate act of making provisional choices, but the illegitimate act of treating those provisional choices as if they resolved foundational disputes.
Human rights institutions should be among the primary defenders against that abandonment, because rights practice—at its best—is built around contestability, justification, and remedy. Those are precisely the virtues needed under conceptual uncertainty. But rights institutions can also become complicit in premature closure if they invoke “human rights” as if it were a finished algorithm: press the button, get the answer.
So what should we do?
If AI breaks our moral vocabulary—or reveals that it was never as stable as we hoped—the response cannot be despair. It must be institutional and ethical design that is robust under uncertainty.
A few principles follow.
First: acknowledge uncertainty explicitly.
This sounds trivial, but it is not. Legal preambles, regulatory frameworks, and corporate governance standards can be candid about moral uncertainty and conceptual contestation. Doing so changes how decisions are justified, how discretion is exercised, and how revision is treated. It also reduces the temptation to treat convenience as truth.
Second: build revisability into AI governance.
If we are governing under structural stalemate, governance must be designed for correction. Sunset clauses, mandatory review periods, iterative risk classification, and requirements of ongoing audit are not bureaucratic ornaments. They are moral safeguards against premature closure.
Third: preserve accountability by design.
Rights require duty-bearers. Governance should resist the diffusion of responsibility by assigning obligations clearly across the lifecycle: developers, deployers, operators. It should enforce traceability and documentation sufficient for contestability. It should make strategic opacity expensive. And in domains like criminal justice, immigration, and other coercive state powers, it should treat meaningful human accountability as a non-delegable requirement, not as an optional “human in the loop” checkbox.
Fourth: use rights as side-constraints where stakes are asymmetric and irreversible.
In some domains—loss of liberty, denial of essential benefits, large-scale political manipulation, lethal force—the appropriate posture is not “optimize outcomes,” but “protect persons.” Even amid ethical pluralism, many traditions converge on the idea that some actions require heightened justification and procedural protection. Rights language is especially appropriate there.
Fifth: treat AI moral status as a moral-risk question, not a culture war.
We should avoid two extremes: dismissing the possibility of AI moral status as silly, and treating it as settled in either direction. Governance can adopt moral-risk approaches that avoid gratuitous cruelty-like design choices, require transparency when companies make “model welfare” claims, and prevent strategic deployment of “AI rights” rhetoric to block regulation—while keeping human protection, especially of vulnerable groups, as the central priority unless and until there is strong evidence of competing moral claims.
A final thought: human rights are not a cheat code—but they may be our best discipline.
It is tempting to want a single framework that bypasses deep disagreement: either a scientific test for consciousness, a decisive moral theory, or an international rights framework that settles everything. AI will not grant us that wish.
But that is not all bad news. If there is one thing rights practice has always had to do, it is to operate under deep pluralism: different religions, different metaphysics, different moral theories, and yet a shared insistence that certain forms of domination and abuse are intolerable. That political achievement is not metaphysical certainty. It is an institutional and moral discipline: justification, contestability, accountability, remedy.
AI makes that discipline harder and more necessary.
The honest conclusion, then, is not that philosophy is irrelevant or that rights are insufficient. It is that AI forces philosophy and human rights into the same room, whether we like it or not. Philosophy cannot deliver the certainty governance craves. Human rights cannot bypass the conceptual foundations they presuppose. But together they can support a form of governance that is more honest, more revisable, and more protective of human beings than the alternatives: technocratic overconfidence on one side, and cynical resignation on the other.
If AI is a test of humanity, it is not only a test of engineering. It is a test of whether we can govern powerful systems without pretending that the deepest questions are already settled—and without letting that unsettledness become an excuse to abandon the rights and ethical commitments that make political life humane.
miss irine | Adobe Stock
