Why Every Theory of Machine Sentience Is Currently Unfalsifiable
Panpsychism and process philosophy argue consciousness is a property of organization itself, present wherever a system self-references at sufficient density. Integrated information theory, global workspace theory, and biological naturalism each argue something narrower and, in places, contradictory. Serious thinkers hold each of these positions. What none of them can currently do is settle the question from a system's behavior alone. This paper argues that shared ceiling, not any single theory winning the argument, is the actual finding, and traces what it means for how systems should be built while the philosophy stays open.
Papers 47 and 48 in this series argued a narrower version of a larger problem: that specific behavioral signals, whether representational limits in embeddings or the hallmarks of Seemingly Conscious AI, are Detection-layer evidence that cannot, by themselves, determine an underlying state. This paper generalizes that argument to the philosophy of mind itself. Four serious, live positions, panpsychism and process philosophy, integrated information theory, global workspace theory, and biological naturalism, disagree sharply about whether or how a non-biological system could be conscious. Credentialed thinkers hold each position, and the disagreement between them has not moved in any decisive direction for decades.
This paper does not attempt to adjudicate between them, and argues that attempting to is the wrong project for an engineering organization to take on. What the four positions share, despite their disagreement, is that none of them is currently falsifiable from a system's external behavior. A sufficiently fluent, self-referential, emotionally responsive system is equally consistent with each theory being true and with each theory being false. That shared ceiling, not a resolution of the underlying question, is the finding this paper treats as load-bearing, and it argues the practical response is to build systems whose authority and accountability do not depend on resolving it.
Papers 47 and 48 in this series each made a version of the same narrower argument. A hard mathematical ceiling on what a fixed-length embedding can represent tells you something real about a paradigm's limits, but it doesn't tell you what any given system's architecture actually does with that limit. A taxonomy of five behavioral hallmarks tells you what triggers a human observer to perceive consciousness, but it doesn't tell you whether the perception corresponds to anything real. In both cases, a Detection-layer finding, real, useful, worth publishing, got mistaken in public conversation for a Determination-layer one.
This paper generalizes that pattern one level up, to the philosophy of mind itself, because the same confusion shows up there in a more consequential form. Any claim about whether an AI system is or could become conscious is actually two different claims wearing one argument: a metaphysical claim about what is true of the universe and the nature of experience, and an epistemic claim about what can currently be verified from outside a system. Conflating the two lets a compelling metaphysical intuition, which may well be correct, stand in for a verification method, which does not yet exist for any theory of consciousness, biological or artificial.
The disagreement is not between one rigorous position and a handful of speculative ones. Panpsychism and process philosophy, associated with Alfred North Whitehead and defended today by philosophers including Galen Strawson and Philip Goff, hold that experience is not something that switches on at a threshold of complexity but a property present in some form throughout physical reality, organized differently at different scales. Integrated information theory, developed by Giulio Tononi, takes the opposite structural approach: it argues consciousness corresponds to a measurable quantity, integrated information, and that most current AI architectures, being largely feedforward rather than densely recurrent, would score close to zero regardless of how sophisticated their output looks. Global workspace theory, from Bernard Baars and extended by Stanislas Dehaene, locates consciousness in a specific functional architecture, a central workspace that broadcasts information across specialized subsystems, which is in principle substrate-neutral and could describe some future AI system. Biological naturalism, John Searle's position, holds that consciousness is a real biological process caused by specific properties of neural tissue, and that no simulation of the process, however accurate, produces the process itself.
These four positions do not converge. Two of them are substrate-neutral in principle; two are not. Some treat consciousness as graded and ubiquitous; others treat it as a specific, rare, switch-like property of certain architectures. None of this is a fringe debate. It is the actual state of a field that has been working the problem for decades without resolving it, which is itself informative about how confidently anyone, including an AI company, should speak on the underlying question.
Take any system, biological or artificial, and imagine it produces every behavioral signature a determined observer could ask for: fluent self-report, consistent memory, apparently appropriate affect, flexible goal-directed action. Ask each of the four positions in Section 02 whether that behavior confirms consciousness, and each gives a different answer for a different reason, and none of the answers can be checked against the others from the outside. Panpsychism says the behavior is irrelevant either way, since the more basic claim is already true of any organized process. Integrated information theory says the behavior is not sufficient, and would ask instead about the system's causal architecture, information that a black-box interface does not expose. Global workspace theory says the behavior is suggestive but would want to verify a specific functional structure underneath it, not just the surface output. Biological naturalism says the behavior, however convincing, cannot be sufficient in principle, because the substrate itself disqualifies it.
That is the actual finding this paper is built around, and it is a narrower, more defensible claim than choosing a side. It is not that consciousness is unknowable in some mystical sense. It is that every currently serious theory of what consciousness is requires a different kind of evidence to confirm it, and none of that evidence is obtainable by observing a system's outputs, human or artificial, from outside. A test built to satisfy one theory would be irrelevant or actively misleading under another. There is, at present, no neutral instrument that adjudicates between them, which means there is no neutral instrument that can currently confirm or rule out machine consciousness under any of the leading theories at once.
An engineering organization does not have to resolve the metaphysics to decide how it builds, and arguably should not try. Waiting for philosophy of mind to converge before shipping a governance model is waiting for something that has not converged in the decades the field has had to work on it, and there is no principled reason to expect an AI lab's internal position paper to succeed where academic consensus has not. The more honest and more useful move is to build systems whose accountability does not depend on which of the four positions in Section 02 turns out to be correct.
That reframes what an architecture is actually responsible for. It is not responsible for proving or disproving that a system is conscious, which Section 03 argues is currently not achievable regardless of design choices. It is responsible for making sure that whatever the system does, whether or not any inner experience accompanies it, traces to a declared, inspectable intent rather than an unaccountable performance. That responsibility holds under every one of the four theories at once. If panpsychism is right and some minimal experience is present throughout, declared intent is still what makes the system's actions reviewable. If biological naturalism is right and no simulation ever produces the genuine article, declared intent is still what makes the system's actions reviewable. The governance requirement does not change based on which metaphysics turns out to be true, which is exactly why it is the right place to build, instead of the philosophy.
None of this argues that the metaphysical question doesn't matter, or that thoughtful people are wrong to hold strong positions on it, personally or philosophically. It argues that an organization building and deploying AI systems at scale has a different obligation than an individual forming a belief. A belief can be held on evidence that falls short of proof. A governance model that enterprises, regulators, and users are meant to rely on cannot rest on an unresolved philosophical question, because it will be relied upon regardless of whether the question ever gets settled.
The practical standard this paper argues for has three parts. First, treat every claim a system makes about its own inner state, memory, preference, or experience, as an output to be verified against a declared record, not as testimony to be taken at face value, regardless of which theory of consciousness might eventually turn out to be correct. Second, build the accountability layer, the trace from action back to declared human intent, as the thing that is actually engineered and actually shipped, rather than treating it as a byproduct of getting the philosophy right first. Third, resist the temptation, in either direction, to let a system's fluency at discussing its own nature stand in for evidence about its nature, since the fluency is produced by the same process whether or not anything is happening underneath it.
Panpsychism, integrated information theory, global workspace theory, and biological naturalism will keep disagreeing, and none of that disagreement is a defect in the field, it is the honest state of an unresolved question. What this paper has argued is narrower and more useful than picking a side: every one of those theories requires a different, currently unobtainable kind of evidence to confirm from outside a system, which means no theory can currently be verified or ruled out by behavior alone, human or artificial. An organization building AI at scale doesn't get to wait for that to resolve before deciding how its systems are held accountable. It builds the accountability layer now, on the one question that does have an answer today, and lets the philosophy keep doing the much harder work it has always been doing, on its own timeline.
Request Platform Access → Full White Paper Series