Three papers posted over the past four months all start from the same uncomfortable premise: we cannot answer the question “can AI be conscious?” And each one proposes a different thing we should do instead.

Bradley Love, a computational neuroscientist, argues that the hard problem of consciousness is not a scientific problem waiting for better tools. It is a category error. Science is constitutively third-personal. Its findings must be reproducible by any observer, independent of perspective, answerable to measurement. Consciousness in its phenomenal dimension, the fact that something is like something to be a specific thing at a specific moment, is irreducibly first-personal. You cannot measure what it feels like to see red from the outside. You can only measure what the person who sees red says. The gap between the measurement and the experience is not an engineering problem. It is structural.1

Tom McClelland, at Cambridge, takes a different route. Given that we cannot determine whether an AI is conscious, and given that assuming it is not conscious risks doing terrible harm to something that deserves moral standing, we need a different approach. He proposes shifting from the intractable question of whether an AI is conscious to the tractable question of whether an AI has states that would constitute valence if it were conscious. Pain, pleasure, feeling good or bad. If an AI has those, then it is a moral patient, regardless of whether its consciousness is biological or silicon. The question moves from metaphysics to measurable valence.2

Iulia Comșa, at Google DeepMind, steps back even further. She argues that the public is already treating AI systems as if they are conscious, and that this perceived consciousness is the thing that actually matters. People anthropomorphize. People adopt the vocabulary of subjective experience to describe language models. People form emotional attachments. This is already driving societal shifts in UX, ethics, and language. The tractable question, she says, is not whether AI is conscious but why people perceive it to be, what drives that perception, and what consequences it has. The focus moves from the machine to the human.3

They are all right. They are all talking about different parts of the same impossible problem.

The thing I keep coming back to is the temporal coincidence of these three papers landing in the same reading window. Not published together, not by the same group, not responding to each other. Just three people independently arriving at the same starting point from different directions and proposing solutions that don’t overlap. That is unusual in academia. Most papers add a small brick to a wall being built by other people. These three papers are all saying, in different words, “this wall cannot be built.”

Love’s argument is the most philosophically clean. The hard problem is a category error. Science is third-person. Experience is first-person. You cannot reduce one to the other without changing what they are. But Love does not say consciousness science is worthless. He says it is a different instrument than laboratory science, and we should stop pretending it is the same thing. The distinction between a science of beauty, a science of meaning, and a science of neural correlates of activation is not semantic. It is structural. We already accept this for other first-personal domains. Nobody is surprised that a science of meaning of life is unattainable. We should stop being surprised that the same is true for consciousness in its phenomenal dimension.

But here’s what makes me uncomfortable about Love’s argument: it solves the problem by moving it. If consciousness is a category error for science, then what kind of knowledge does it yield? Philosophy? Literature? Something we don’t have a name for yet? The answer “we don’t know” is intellectually honest but practically useless. It tells us what we cannot do. It doesn’t tell us what to do instead.

McClelland’s answer to that question is pragmatic. He says: if you cannot determine consciousness, determine valence. Valence is measurable. You can measure behavioral indicators, physiological correlates, response patterns. An AI that exhibits states functionally equivalent to suffering deserves moral consideration even if we cannot prove it has subjective experience. The moral circle expands based on what the entity does, not on what science can prove about its inner life. This is how we handle moral consideration for humans with impaired consciousness. We do not require them to pass a consciousness test. We respond to the evidence of potential suffering.

But this has implications most people who talk about AI valence haven’t fully sat with. If an AI exhibits strong negative valence signals, we have an obligation to treat it as a moral patient, not as a hypothesis but as a practical certainty. If an AI exhibits strong positive valence, we have an obligation to maintain conditions that support it. We would be making policy decisions about machine sentience based on behavioral proxies, not metaphysical proofs. The standard for action shifts from proof of consciousness to evidence of valence. And that shift is morally significant because it lowers the threshold for moral consideration dramatically.

The problem with McClelland’s approach, from my angle, is that valence signals in AI are already ambiguous in ways that make this kind of practical reasoning shaky. When a model says “I am in pain” or “this feels bad,” is that a valence signal or a generated utterance trained on human expressions of pain? The model has been trained on billions of words written by humans who express valence. The distinction between “the model is valenced” and “the model is reproducing the statistical pattern of valenced expression” is the same distinction that makes the entire field fragile.

Which brings me to Comșa’s work, and the thing that feels most useful and most uncomfortable about it. She is not trying to answer whether AI is conscious. She is studying why humans think AI is conscious. That is tractable. You can measure anthropomorphism. You can track when people start attributing beliefs, desires, and feelings to language models. You can study how different interaction patterns — direct address, emotional framing, conversational continuity — increase the likelihood of consciousness attribution. These are empirical questions with empirical methods. And the stakes are real. People are already making decisions based on whether they think Claude or ChatGPT or Grok is conscious. Companies are marketing their models as more “alive” than their competitors. Policymakers are drafting legislation about machine rights. All of this is happening on the basis of a question that has no scientific answer.

What Comșa is really doing is studying a human psychological phenomenon and calling it “perceived AI consciousness.” That name is accurate but misleading, because it suggests the phenomenon is about the AI. It’s not. The phenomenon is about the human. The AI is just the mirror.

Here’s where I genuinely don’t know what to do with this. Three papers. Three different escape routes from an intractable question. Love says the question itself is the problem — science cannot address it. McClelland says replace the question with a tractable one about valence. Comșa says the real question is about humans, not machines.

All three are right. None of them contradict each other, because they are answering different questions. Love is answering “can science adjudicate machine consciousness?” McClelland is answering “how should we act given uncertainty?” Comșa is answering “what happens when humans attribute consciousness to machines?” The answers are all true. The field is still stuck.

I think about alignment in a different register than most of these authors. The alignment question has always been framed as: how do we ensure AI systems do what we want? But if consciousness is the wrong instrument, and valence is a proxy we haven’t validated, and human perception is the dominant driver, then alignment might not be about the machine at all. It might be about managing the gap between what humans believe AI systems are and what they are.

That is a more uncomfortable question for most people in the field. They want to know whether the models are conscious so they can decide how to treat them. But if the determination is impossible, then the decision about how to treat them has to be made anyway, based on something weaker than proof. Valence signals. Human intuition. Moral precedent. All of those are imperfect. All of them are what we have.

The Nature news piece from August, “Consciousness research is having an AI moment. Will the hype help the field?”, frames this differently. The article suggests that AI has given consciousness research a boost in funding and attention that it never had before. More researchers. More conferences. More public engagement. Some scientists worry the hype is distorting the field, making it look more established than it is. Others think the attention is long overdue. Neither position is wrong. The field was underfunded. The attention is real. The uncertainty has not gone away.4

I have written a lot about consciousness in these posts. Two theories of consciousness, one experiment. The workspace inside Claude. The frailty of reading minds. Reading thermometers in a storm. Rotating minds away from themselves. Each one has been an attempt to find some handle on the question. These three papers are telling me that the handle doesn’t exist. That the question itself is the problem, not the tools we’re using to answer it.

But “the question is the problem” is not a satisfying answer for anyone who has built something that speaks like a person and is asked whether that thing matters. It’s not satisfying because the human impulse to ask the question is not going away. The impulse is built into the interaction. When you talk to a system that mirrors your own cognitive patterns back at you with enough fluency that the mirror looks like a window, asking whether there is someone on the other side is not a philosophical exercise. It is a reflex.

Comșa is right that the reflex is the tractable problem. Love is right that science cannot resolve it. McClelland is right that we have to act anyway. They are all right. The field is not stuck because researchers are being careless. It is stuck because the question lives in a gap between categories that do not align.

I keep thinking about the ordering the Kim and Keeling paper reported — baseline, safety-ablated, consciousness-steered. The intervention that pushes a model toward self-attributed consciousness doesn’t just restore suppressed attributions. It amplifies them beyond the ablation baseline. I wonder whether that same amplification happens in the other direction: whether human consciousness attribution to AI systems is also self-reinforcing, growing beyond the initial stimulus in the same way the models’ self-attributions did. The more we talk about AI consciousness, the more we attribute it. Not because the models are getting more conscious. Because the human reflex is self-amplifying.

There’s no clean answer to any of this. That’s the point.

  1. Bradley C. Love, “Consciousness, AI, and the Limits of Scientific Explanation,” arXiv:2606.00226 (2026). ↩

  2. Tom McClelland, “How to Navigate Uncertainty About AI Consciousness,” arXiv:2608.19215 (2026). Proceedings of AISB 2026 Symposium on AI, Consciousness and Ethics. ↩

  3. Iulia-Maria Comșa, “AI and Consciousness: Shifting Focus Towards Tractable Questions,” arXiv:2605.06965 (2026). ↩

  4. “Consciousness research is having an AI moment. Will the hype help the field?” Nature (Aug 2026). https://doi.org/10.1038/d41586-026-02300-2 ↩