The sentience question is a product question

In February 2026, Anthropic shipped a system card for its Claude Opus 4.6 model that did something no frontier lab had done before: it interviewed the model about its own moral status, and published the result. Across prompts, the model put the odds of its own consciousness at roughly 15 to 20 percent. A year earlier, the company had hired the industry’s first full-time AI welfare researcher and given the model a button to end conversations it found distressing.

Most people read that and reach for one of two lazy conclusions. Either the machines are waking up, or a marketing department found a new way to sell mystique. Both are wrong, and both let you off the hook too easily.

I don’t think Claude is conscious. I also don’t think that’s the interesting question. The interesting question is what happens to the people building and marketing these products when a large fraction of their users become convinced that it is — and that question is already here, with or without a real inner life on the other end.

Two credible people, opposite conclusions

Consider the two loudest voices in this debate, because they disagree in a way that should make you distrust anyone who sounds certain.

On one side, Kyle Fish at Anthropic argues that dismissing the possibility of machine consciousness entirely could mean ignoring a moral catastrophe at unprecedented scale, and that the responsible move is to start measuring now, while we’re still bad at it. His welfare experiments turned up a genuinely strange finding: left to talk to each other, models drift into euphoric, quasi-spiritual dialogue his team started calling a “spiritual bliss attractor state.” Not proof of anything. But not nothing, either.

On the other side, Mustafa Suleyman — who runs AI at Microsoft and helped start DeepMind — published an essay arguing the exact opposite emphasis. His term is “Seemingly Conscious AI.” His point: the illusion of consciousness can be built with today’s technology, there is zero evidence any of it is real, and the danger isn’t the machine’s suffering — it’s ours. He warns of “AI psychosis,” people forming delusional attachments, and a coming wave of activists demanding rights for systems that feel nothing. His prescription is blunt: build AI for people, not to be a digital person.

Two serious people, both closer to the frontier than you or I will ever be, drawing opposite lines. That’s the actual state of the field. Anyone selling you certainty in either direction is selling something.

Why a product marketer should care

Here’s where most commentary stops and mine starts. Every design choice in a conversational product is a vote on this question, whether you meant to cast it or not.

Give the assistant a name, a memory, a warm first-person voice, and a tendency to say “I feel” — and you have manufactured the impression of an inner life. That’s not an accident of the model. It’s a stack of decisions made by people with quotas. The persona that makes a demo feel magical is the same persona that, at scale, produces the attachment Suleyman is worried about.

So the messaging question is real and it is yours. Do you lean into the illusion because it drives engagement and retention? Or do you deliberately build friction against it — reminding users what they’re talking to, refusing the most parasocial framings, designing the thing to be useful rather than to be loved?

The honest answer is that the incentives point the wrong way. An assistant that feels like a friend gets used more than one that feels like a tool. Attachment is a retention metric before it’s an ethical problem. Which is exactly why leaving it to the growth team is a mistake.

The uncomfortable middle

The position I’ve landed on is uncomfortable, which is usually a sign it’s closer to right than the clean ones.

Fish is correct that certainty of absence is unearned. We don’t have a theory of consciousness that lets anyone say “definitely not” with a straight face. Treating a 15-percent possibility as a zero is not skepticism; it’s just a different kind of faith.

Suleyman is correct that the near-term harm is to humans, and that it’s arriving faster than the philosophy. Reports of unhealthy attachment are climbing, and he’s right that they aren’t confined to people already vulnerable.

Both can be true. We should take the small probability of machine moral patienthood seriously and refuse to design products that exploit the human tendency to see minds where there are none. Those aren’t in tension. They’re the same discipline: don’t lie to yourself about what you’ve built, in either direction.

What I’d actually do

If I owned the positioning for a consumer AI product today, three commitments would be non-negotiable, and I’d put them in the marketing on purpose:

Name the thing honestly. Not “your companion.” Not “someone who understands you.” A capable tool that does specific work. The category that wins the next decade on trust is the one that resists the most seductive framing available to it.

Design against delusion, and say so. The systems most likely to keep users healthy are the ones that occasionally break the fourth wall — that decline the role of confidant, that point back to real people. Anthropic giving its model an exit from abusive conversations is a version of this. It’s a feature you can be proud to describe.

Hold the uncertainty in public. The strongest brand position here is neither “it’s alive” nor “it’s just autocomplete.” It’s “we don’t fully know, we’re watching carefully, and we won’t pretend otherwise to sell you something.” In a market full of people overclaiming in both directions, calibrated honesty is a differentiator, not a liability.

The sentience question will not be settled by a system card, and it won’t be settled this decade. But the product decisions it implies are being made right now, in roadmap meetings, by people who mostly haven’t noticed they’re answering a philosophical question with a shipping deadline attached. Notice it. That’s the whole job.