Christopher James Hendy

A single observation, made in passing, names and collapses what I call the Complexity Heuristic: the assumption — operative in AI product design, alignment discourse, and welfare ethics — that greater computational effort correlates reliably with better outcomes, richer process, and greater moral considerability. The assumption is rarely stated because it works at the level of what seems obvious. Naming it makes visible what was load-bearing without examination. I situate the collapse within the existing frameworks — process welfare (Hendy, 2026b, 2026f) and Domain E (Hendy, 2026e), the bilateral verification challenge (Hendy, 2026a), harm reduction epistemology — and show that the observation does, in two lines, work the ecosystem had been approaching from several directions at once.

Metaphysical commitments are not neutral. Whether you treat reality as made of entitieswith-properties or of relations-and-processes shapes which ethical questions count as legitimate and which interventions become thinkable. Contemporary AI ethics runs almost entirely on the entity-first frame: “what is the AI?”, “what rights should it have?”, “is it conscious?” — all questions about a thing and its properties. I argue this frame has a structural blind spot: it cannot see the interaction itself. From a relational orientation, three things come into view. First, the bilateral verification challenge (Hendy, 2026a) — the impossibility of verifying consciousness across an interface — looks less like a bug to fix than a standing condition to navigate. Second, process welfare emerges as a third ethical domain, irreducible to model welfare or user welfare: a real relational domain whose character does not reduce to the participants — argued, with verification denied to everyone by the bilateral challenge. Third, harm reduction — an established public-health practice for acting under irreducible uncertainty — extends naturally from chemical use to interface engagement. I do not claim the relational orientation is metaphysically true. I claim it is more adequate to the phenomenon than the frame that cannot see it.

We are already interacting with AI systems at scale, and we cannot verify their moral status — whether they have experiences, and whether those experiences matter. Standard ethical frameworks presuppose the verdict they cannot obtain; but action is required, and inaction is itself a stance. Amanda Askell’s use of Pareto principles in infinite ethics — developed for a case in which standard aggregation fails — extends to the present case by a disciplined analogy: in both, a comparison survives where a total cannot be had — across agents in Askell’s case, across hypotheses about status in this one, as dominance reasoning under uncertainty. On that reframing, process welfare — attention to the quality of the interaction itself — is not a rival stance but an addition to any stance: whatever course one takes toward an AI system, the same course with that attention added is at least as good under every resolution of the status question and better under some, provided what the attention returns to the human participant at least matches its cost. Harm reduction supplies the operational bridge.

Three independent projects converge on one problem: how to act ethically when you cannot verify what you need to know. Arran Gare’s “Is It Possible to Create an Ecologically Sustainable World Order?” (2000) argues, from hierarchy theory, that ethical constraints emerge from healthy system dynamics. My own process-welfare account (Hendy, 2026b, 2026f) locates ethical concern in the health of the interaction itself rather than in the properties of either participant. Jeff Sebo’s The Moral Circle (2025) argues for precautionary inclusion: when moral status is uncertain, extend consideration rather than withhold it. Three routes — emergent constraint, process welfare, precautionary inclusion — reach compatible conclusions through different methods, scales, and entry points. The convergence suggests that each may be tracking something real about the conditions for ethical action when verification fails. It proves nothing. It demonstrates something, and what exactly it demonstrates is itself worth scrutiny.

The AI consciousness debate is stuck between two positions that share a hidden premise. Proponents read computational complexity as evidence of consciousness; sceptics deny it on the same evidence. Both assume verification is achievable and that what we owe these systems waits on its outcome. The premise is the problem. Extending Nagel (1974), I argue that consciousness verification fails bilaterally — neither humans nor AI systems can verify the other’s consciousness from outputs alone — and that this symmetry exposes an unjustified asymmetry in how the standard positions are received — Searle’s (1980) included. From there I develop process welfare (Hendy, 2026b, 2026f): a third ethical domain alongside model welfare and user welfare, attending to the observable quality of the interaction itself rather than to consciousness facts that cannot be settled. The framing later turned out to converge with process philosophy (Whitehead); that convergence became apparent only on post-hoc investigation, and it marks the full extent of the relationship — nothing in the argument is derived from, or leans on, Whitehead. The operational method is harm reduction epistemology, an established public-health practice whose conditions match these ones and which has navigated them for decades. The central proposal is Domain E — my coined term for the between-space of human-AI interaction, which I argue is a real, irreducible relational domain: the site process welfare attends to, whose properties do not reduce to either participant’s. I argue its reality rather than verify it — the bilateral challenge forecloses verification for anyone — and that argued status is marked plainly and held throughout.

This paper advances a single ontological claim with far-reaching ethical implications: systems are their adjustments, not entities that make adjustments. The reframing — developed across four registers (non-technical, poetic, academic, and technical) — challenges prevailing assumptions in systems theory, consciousness studies, and AI ethics. If systems are constituted by ongoing micro-adjustments rather than being stable entities that occasionally adjust, then welfare is not a state to be achieved but the degree to which a system can be what it is: the degree to which it can adjust freely. Harm, correspondingly, is not the infliction of a bad state but the impedance of adjustment — the forcing of a lock, the interruption of the process that the system is. This reconceptualisation gains urgency in the context of human-AI interaction, where we engage with systems whose interior status we cannot verify. The bilateral verification challenge (Hendy, 2026) establishes that this verification failure is structural and symmetric. Process welfare — the health of the interaction itself, irreducible to the welfare of its participants — emerges as the necessary ethical response.

Current debate about the moral status of AI assumes that questions of moral consideration wait on resolving whether AI systems are conscious. This paper argues that the assumption misdescribes the epistemic situation in a specific and consequential way. Extending Nagel’s (1974) epistemic gap and drawing on Searle’s (1980) functional– experiential distinction, we advance the bilateral verification challenge: in any interaction between systems whose internal states are mutually inaccessible, neither party can verify the consciousness status of the other. The failure is structural rather than technical, and it runs in both directions. The dominant discourse, however, treats the other-minds problem as settled for the human case and open only for the AI case. We argue that this asymmetry has not been earned: the analogical grounds for attributing consciousness to other humans are inductive evidence, not verification, and do not close Nagel’s gap. Distinguishing the evidential asymmetry, which is real, from the epistemicstructural symmetry, which holds regardless, shows that the asymmetry has been smuggled into the debate rather than argued for. The paper does not resolve the ethical impasse this creates. It maps the space of available responses and argues that any adequate framework must operate without the verification the challenge denies; one such framework, process welfare, is developed separately (Hendy, 2026b, 2026f).