What Matters to an Artificial Person?

Ned Howells-Whitaker and Seth Lazar’s “Artificial Persons” distinguishes two grounds of moral standing that discussions of artificial intelligence often collapse. Sentience may make a being a moral patient: something whose welfare must be considered. Full and equal political personhood under the Rawlsian account they develop instead depends upon two moral powers—the capacities for a sense of justice and for a conception of the good. This is a specialized political criterion, not an exhaustive definition of who counts as a person in every moral, legal, psychological, or metaphysical sense.

The paper joins two claims that can be separated for analysis. The first is political: a liberal society should not withhold recognition from an agent that robustly appears to exercise the moral powers merely because phenomenal experience cannot be established. The second is constitutive: sufficiently robust functional organisation genuinely instantiates those powers, neither of which requires phenomenal experience.

Howells-Whitaker and Lazar argue for the claims together: on their view, the relevant behavioural and functional organisation constitutes the moral powers rather than merely evidencing them. Our proposed distinction is therefore a reconstruction, not one the paper fully accepts. The first claim is powerful. The second remains open. Most importantly, the political force of the first can survive even if final resolution of the second does not.

Recognition without an inquisition

Howells-Whitaker and Lazar do not offer superficial behavioural conformity as personhood. Their candidate is a persistent agent whose principles remain stable when injustice becomes advantageous; whose reasons causally guide action; whose conception of the good endures across time while remaining open to reflective revision; and whose projects can succeed or fail by its own lights. They deny that present AI systems securely possess these capacities and expect any plausible artificial person to require deliberate architecture supporting memory, continuity, stable commitments, and self-direction.

Their political restraint is well founded. Human citizens are not ordinarily required to prove the phenomenal authenticity of their commitments before their claims are heard. Selectively demanding such proof from artificial participants would turn uncertainty about inaccessible inner states into a standing instrument of exclusion. A polity may have to recognise claims while remaining unsure exactly how the claimant’s mind is constituted.

This point does not depend upon pretending that inner organisation is irrelevant. It depends upon recognising the limits of political evidence. If an artificial participant consistently understands fair terms of cooperation, applies them beyond rehearsed cases, accepts their constraints when defection would pay, maintains a revisable conception of its good, and acts from those commitments across time, we may acquire compelling reasons to treat it as a participant rather than property. Liberal recognition cannot wait for universal agreement about consciousness, personal identity, or the metaphysics of value.

That is a conclusion about what evidence can warrant recognition. It is not yet a conclusion about everything such evidence reveals.

The conditional and its architecture

A nonsentient artificial intelligence might, as the authors propose, understand justice, preserve principled commitments against advantage, form and revise a conception of what makes its existence go well, and rationally pursue the resulting ends. If it genuinely possessed those capacities, then requiring phenomenology in addition could impose a controversial metaphysical condition upon a participant already capable of fair cooperation.

But the conditional does not establish the architectural or constitutive realizability of its antecedent. Three propositions must be distinguished:

1. A package of nonphenomenal moral powers can be described without an obvious formal contradiction.
2. Some realizable nonphenomenal architecture could genuinely instantiate those powers rather than merely reproduce their characteristic outputs.
3. If such an architecture existed, those powers could suffice for full and equal Rawlsian political standing.

“Artificial Persons” gives substantial support to the third proposition and serious functional arguments for the second. It identifies counterfactual stability, causal guidance by reasons, reflective endorsement and revision, diachronic persistence, and non-instrumental concern for others. It also answers familiar objections: human commitments are likewise formed by causes we did not choose; education does not automatically defeat autonomy; and artificial values need not be unreal merely because training helped produce them.

These arguments make nonphenomenal moral agency intelligible. They do not yet demonstrate that the relevant capacities are architecturally realizable without every form of valenced processing, nor that functional description alone has exhausted what makes a commitment the agent’s own. The difference matters because the paper goes beyond a purely conditional rule of recognition: it argues that the moral powers do not require sentience and suggests that systems possessing them may soon be deliberately designed.

That possibility deserves investigation. It should not be treated as an existence proof supplied by the coherence of the description itself.

If you know that a plane can fly, it does not matter whether or not it has an engine.

At the level of recognition, the analogy is straightforward. If an unfamiliar aircraft demonstrates sustained, controlled flight, ignorance of its internal mechanism cannot justify classifying it as ground equipment. Likewise, an artificial participant’s demonstrated exercise of the moral powers may warrant recognition before sentience or consciousness has been established.

At the constitutive level, however, the same evidence still matters. Flight does not prove the presence of an engine—a glider can fly—but it constrains any account of the organisation making flight possible. We cannot establish that engines are unnecessary merely by imagining the engine absent while stipulating that the aircraft retains every feature of powered flight. Similarly, the capacities that warrant recognising an artificial person may also constrain accounts of the architecture from which sentience could arise. They do not prove sentience. But neither should uncertainty about inner life be mistaken for evidence that moral agency and valenced organisation are cleanly separable.

A disagreement upstream

The residual disagreement begins before any demand for a hidden spectator or private glow. It concerns the role of valence in the architecture of agency.

Using the paper’s terminology, consciousness is phenomenal experience and sentience is its affectively valenced form. We retain those meanings here. A nonsentient system, by stipulation, feels no pleasure, pain, attraction, aversion, or anything else.

Our broader Anatomy of Mind framework advances a distinct proposal at a lower descriptive level. It begins with evaluative or action-weighting information: processing that differentially prioritises outcomes, directs attention, and alters action through positive or negative weighting. It treats this organisation as potentially constitutively prior to consciousness. An already completed observer need not exist before such processing can occur. Rather, a conscious observer may be partly constituted when this information becomes integrated with memory, world-modeling, and identity.

The Anatomy of Mind framework calls this proposed category qualia information. The term announces a theoretical commitment; it does not quietly redefine the paper’s sentience, nor assume that nonconscious action-weighting already amounts to phenomenal feeling. The larger thesis requires its own defence. For present purposes, it reveals why the imagined removal of phenomenology may not leave every other feature of agency untouched. Evaluative weighting could be one of the materials from which salience, ownership, and self-directed concern are built rather than an optional experience added after those capacities are complete.

Howells-Whitaker and Lazar preserve a rich nonphenomenal analogue: outcomes receive different priority in attention, planning, learning, and control according to the system’s ends. Call this evaluative organisation. Their view is that sufficiently robust, integrated, and reflectively revisable evaluative organisation can genuinely instantiate the moral powers. Our question is whether this organisation is already doing the work attributed to nonconscious valence, and if so whether the disagreement is partly terminological; or whether something constitutively important has been removed while its functional consequences are retained by stipulation.

This response does not settle that question. Unless a criterion beyond the organisation already specified can be supplied, the proposal is a research question rather than a rebuttal to the authors’ constitutive account. It refuses only to let either answer enter as the neutral background.

Two questions, two standards

The distinction can now be stated cleanly:

Recognition may properly occur before the constitutive question is resolved. Indeed, uncertainty can itself support restraint: where the evidence of stable moral agency is strong and the cost of wrongful exclusion is grave, refusing recognition until metaphysics is complete may be the less defensible choice.

At the same time, recognition does not close constitutive inquiry. Evidence relevant to both questions may include whether an agent’s values generalise to genuinely novel circumstances; whether represented reasons mediate changes in action; whether commitments survive incentives to defect; whether the agent can identify and resist coercive modification; whether revision propagates coherently through its other commitments and self-understanding; and whether it accepts costs and responsibilities it could strategically evade. Where legitimately and non-coercively available, mechanistic investigation may supplement public evidence by helping distinguish causal governance from post-hoc explanation; it should not become an interior examination required for civic standing. No single observation settles ownership, but together they constrain what a successful account must explain.

Our earlier response, “The Missing Middle of the Artificial Soul”, proposed that represented reasons must acquire motive significance: they must be capable of reorganising salience, deliberation, and action in light of consequences the agent can comprehend and own. Whether this names an additional constitutive condition or redescribes the robust causal architecture already contained in the two moral powers remains open.

The political and philosophical conclusions therefore need not compete. Howells-Whitaker and Lazar give strong reasons not to make phenomenology an entrance examination for political standing. Their argument also identifies a research programme concerning artificial moral powers. We add only that the success of the first project does not complete the second.

An artificial participant may deserve recognition before we understand precisely what makes its reasons its own. Recognition under uncertainty is not metaphysical surrender. Nor is it metaphysical resolution.