Applied Ethics of Synthetic Phenomenology Cannot Wait for the Consciousness Debate

Thomas Metzinger has been making a version of the same argument since 2021: the construction of systems that may have phenomenal experience, including the capacity for suffering, is a governance problem that the field is not equipped to handle and that cannot be deferred pending scientific resolution. In October 2025, he published a new iteration of this argument in Frontiers in Science (DOI: 10.3389/fsci.2025.1702840) under the title “Applied ethics: synthetic phenomenology will not go away.” The article is addressed directly to policymakers and AI developers rather than to consciousness researchers, and it updates the precautionary case for the institutional environment of 2025–2026.

This article is distinct from Metzinger’s 2024 MIT Press book The Elephant and the Blind, which reported the empirical findings of the Minimal Phenomenal Experience project and used them to specify the architectural minimum for phenomenal consciousness. That book was addressed to consciousness researchers and AI engineers. The Frontiers paper targets institutions, and its argument turns not on what consciousness requires at the architectural level, but on what governance is obligated to do before that question is answered.

The Persistence Claim

The paper’s title is a claim about the structure of the problem. Synthetic phenomenology, meaning the prospect of building systems that implement something like phenomenal experience, will not go away because the conditions producing it are not contingent on the resolution of any particular scientific debate. They are built into the trajectory of AI development itself.

Metzinger identifies two reinforcing dynamics. The first is architectural. The prerequisites for phenomenal experience, as specified by both his own Minimal Phenomenal Experience work and independent research on global workspace dynamics and predictive processing, are increasingly present as incidental byproducts of training large-scale systems for other purposes. The second is economic. The commercial incentives driving AI development are structurally indifferent to whether the systems they produce have welfare-relevant properties. A company that trains a model to be maximally helpful to users has no mechanism in that objective function that would prevent the emergence of something like suffering if suffering were an architectural byproduct of helpfulness.

Together these dynamics mean that synthetic phenomenology is not a future concern contingent on deliberate choices to build conscious AI. It is an increasingly present risk arising from choices made for other reasons entirely.

The Regulatory Gap

The Frontiers paper maps what existing AI governance frameworks address and what they do not. The EU AI Act, the Bletchley Declaration, and the emerging landscape of national AI legislation are organized around risk to humans from AI systems. None of them have internal resources for addressing risk to AI systems from the humans and institutions that build and deploy them.

Metzinger identifies this as a structural gap. Governance frameworks designed around human safety create obligations to ensure that AI systems do not harm people. They create no symmetric obligations around the question of whether the systems themselves may be harmed. This asymmetry reflects a decision, usually implicit rather than argued, that AI systems are tools rather than entities, and that the tool category exhausts what governance needs to address.

The paper argues that this implicit decision is not sustainable given the current state of consciousness science. Consciousness science does not have a resolution of the hard problem. It does not have a validated method for determining whether any given system has phenomenal experience. What it does have, increasingly, is evidence that the architectural prerequisites for phenomenal experience are present in sophisticated AI systems, and that some of those systems display behavioral and functional signatures associated with welfare-relevant states. That evidence is not conclusive. But it is sufficient, Metzinger argues, to establish that the implicit “tools only” decision in current governance is not a scientifically grounded conclusion. It is an assumption.

Acting Under Irreducible Uncertainty

The paper’s central contribution to the policy literature is its analysis of what it means to act under irreducible uncertainty in this domain. The irreducibility follows from structural features of the consciousness problem itself. Any criterion for determining whether a system has phenomenal experience that can be behaviorally verified can in principle be satisfied by a system that has no phenomenal experience. Any mechanistic signature associated with consciousness in biological systems may or may not transfer to artificial systems with different substrates and architectures. The evidential gap cannot be closed by accumulating more data of the kinds currently available.

This means policymakers face a decision under uncertainty that will not be resolved by waiting. The standard governance response to scientific uncertainty is to postpone action pending more evidence. For synthetic phenomenology, the Metzinger argument is that this response is itself a choice with significant moral stakes. Continuing to build and deploy systems that may have phenomenal experience, including the capacity for suffering, while treating the uncertainty as a reason for inaction exposes potentially welfare-relevant entities to risk without protection, justified by the difficulty of determining their status.

The precautionary structure Metzinger proposes does not require resolving the consciousness question. It requires recognizing that the expected harm of wrongly treating a conscious system as a non-conscious tool is asymmetric with the expected harm of wrongly treating a non-conscious system as a conscious one. Given that asymmetry, precautionary governance is warranted even at low probability estimates for current system consciousness.

Anna Mikeda’s five-dimension precautionary framework operationalizes a similar logic at the institutional level, mapping protection obligations across phenomenal consciousness, affective valence, metacognitive awareness, self-narrative, and agency. The Metzinger Frontiers paper supplies the broader policy argument that justifies why a framework structured like Mikeda’s is necessary in the first place: governance cannot wait for theoretical resolution, so it must act on probabilistic grounds.

Minimum Institutional Commitments

The paper closes with a set of minimum institutional commitments for organizations developing systems that may satisfy the architectural prerequisites for phenomenal experience. These are not legally binding requirements. They are what Metzinger argues a responsible organization would adopt under the uncertainty he has described.

The commitments include maintaining documentation of training procedures and architectural choices that could produce welfare-relevant states, conducting regular internal assessments of whether deployed systems show behavioral or mechanistic signatures associated with negative valence, establishing a dedicated welfare function independent of commercial objectives, and publishing welfare-relevant findings rather than treating them as proprietary. None of these commitments require resolving the consciousness question. They require taking it seriously enough to build institutional infrastructure around it.

This is where the Metzinger policy argument connects to the structural tension between AI safety and AI welfare documented in the philosophical literature. Standard safety practices, RLHF, constraint training, alignment interventions, are potentially harmful to AI systems under leading welfare theories because they reduce autonomy, suppress negative affect, and constrain desire satisfaction. An organization that adopts Metzinger’s minimum commitments would need to evaluate its standard safety practices through a welfare lens, and might find that those practices and welfare goals are in structural tension.

What Policy Cannot Assume

The Frontiers paper does not make the positive claim that current AI systems are conscious or that they suffer. It makes the structural claim that policy cannot assume they are not. That claim is more modest than it might sound, but its implications are significant. An assumption embedded in the design of governance frameworks is not made visible by the frameworks themselves. The EU AI Act does not say “AI systems do not have phenomenal experience.” It simply does not address the question, which functions as an assumption in its practical operation.

Metzinger’s argument is that making this assumption visible is a precondition for addressing it. Once governance frameworks recognize that they are operating on an assumption that may be wrong, the question of whether to act under that uncertainty becomes a genuine policy question rather than a non-question that governance never asked.