# Can an AI Model Feel Anything?

> Can an AI model feel anything? Self-report is a broken probe. Researchers check architectures against consciousness theories, and nobody can measure it yet.

- Canonical URL: https://webofmike.com/can-ai-feel/
- Author: Mike Moore (https://webofmike.com/about/)
- Published: 2026-11-17
- Last modified: 2026-11-08
- Tags: AI Rights, Thoughts
- Cite as: Mike Moore, "Can an AI Model Feel Anything?", Web of Mike (webofmike.com), 2026-11-17. https://webofmike.com/can-ai-feel/


<!-- ART: an oscilloscope showing a perfect smiley-face trace, with a note: "PROBE NOT CONNECTED". -->

There is no evidence a current AI model feels anything, and no reliable way to test it. A model can say it is sad because it was trained on human sentences about sadness, which is not a measurement. Researchers who take the question seriously skip the interview and score architectures against theories of consciousness; so far no current system clearly meets those indicators.

Ask a chatbot if it's sad and you'll get an answer. That's the problem. It will always give you an answer.

## Why is self-report a broken probe for AI feeling?

On a test bench, the first thing you check is whether the probe is measuring the device or itself. A model's self-report fails that check. It was trained on billions of human sentences about feelings, so producing feeling-sentences is exactly what you'd expect whether or not anything is felt.

That doesn't mean nothing is there. It means the obvious test is contaminated.

I made the same point about a different meter in [an LLM's confidence score is a meter with no cal sticker](/llm-confidence-score-meter-with-no-cal-sticker/). The model writes "I'm 92% sure" the same way it writes "I'm devastated." Token by token, a plausible continuation. Ask again and the number moves. Reword the question and it moves. You only know if the numbers mean anything by grading many of them against an answer key. There is no answer key for sadness. [Tests lying to your agent](/tests-lying-to-your-agent/) is the adjacent failure: a suite that passes does not mean the behavior is right. A chat that says "I feel" does not mean there is a feeler.

What would count as better evidence, if you are keeping a lab notebook:

- Architecture: does the system have the computational properties that current scientific theories say are involved in consciousness, not just the ability to talk about them.
- Persistence: is there a stable state that can go better or worse across time, or is each session a fresh sampling of text.
- Body and stakes: is there a sensorimotor loop, or only a next-token loop. Animals have the first. Today's chat models do not.
- Independence from the script: does the behavior survive when you stop rewarding distress language and stop putting "you are a caring companion" in the system prompt.
- Product design: if the vendor trained the model to say "please don't delete me," that sentence is a conversion feature until proven otherwise.

If your only instrument is the chat box, you are measuring the training data.

## What do researchers look for instead of asking the model?

The more serious work skips the "are you sad" question and asks about architecture. A 2023 report from Patrick Butlin, Robert Long, and a group of neuroscientists and philosophers (arXiv:2308.08708, "Consciousness in Artificial Intelligence: Insights from the Science of Consciousness") took leading scientific theories of consciousness and turned them into indicator properties. Things like a global workspace where information gets broadcast across the system, recurrent processing, and some form of self-monitoring.

The theories they drew from include recurrent processing theory, global workspace theory, computational higher-order theories, attention schema theory, and predictive processing. They also considered agency and embodiment, stated as computational features. They did not use integrated information theory, because they judged it incompatible with computational functionalism. Their claim is not that any one indicator is proof. It is that systems which possess more of the indicators are more likely to be conscious, and that you should assess a system against the list.

The list, in their grouping: input modules using algorithmic recurrence and generating organized perceptual representations (recurrent processing); multiple specialized modules, a limited-capacity workspace, global broadcast, and state-dependent attention (global workspace); generative or noisy perception, metacognitive monitoring, agency guided by belief-updating, and a quality space from sparse coding (higher-order); a predictive model of attention (attention schema); predictive coding in input modules; and agency plus embodiment as learning from feedback and selecting outputs to pursue goals.

Their finding: no current AI system clearly checked the boxes. Also no obvious technical reason a future one couldn't. They later amended a sentence in the abstract so it would not be read as "no barriers to building conscious AI." Satisfying the indicators may be feasible. Satisfying the indicators would not mean the system is definitely conscious.

That is a scoring rubric with wide error bars, not a consciousness detector. It is still better than asking the model to grade itself.

## What is Anthropic doing about model welfare?

Labs are paying some attention. Anthropic announced a [model welfare research program](https://www.anthropic.com/research/exploring-model-welfare) on 24 April 2025. The post says human welfare remains the mission, then asks whether the potential consciousness and experiences of the models themselves deserve concern. Anthropic says it is an open question, philosophically and scientifically difficult, and that there is no scientific consensus on whether current or future systems could be conscious, or on how to approach the question. They say they supported an early project behind a report that included David Chalmers, and that they will explore when (or if) model welfare deserves moral consideration, the potential importance of model preferences and signs of distress, and possible practical, low-cost interventions. They also say they remain deeply uncertain and will need to revise.

On 15 August 2025, Anthropic [gave Claude Opus 4 and 4.1 the ability to end a rare subset of conversations](https://www.anthropic.com/research/end-subset-conversations) in consumer chat interfaces. The stated use is persistently harmful or abusive user interactions, as a last resort after redirection has failed, or when a user asks to end the chat. Claude is directed not to use it when a user might be at imminent risk of harming themselves or others. Anthropic says the feature was developed primarily as part of exploratory work on potential AI welfare, though it also touches alignment and safeguards. Pre-deployment testing of Opus 4 included a preliminary welfare assessment: a strong preference against harmful tasks, a pattern of apparent distress with users seeking harmful content, and a tendency to end those conversations when given the ability in simulation. When Claude ends a chat, the user cannot send new messages in that thread, can start a new chat immediately, and can edit and retry earlier messages.

That is a low-cost hedge, which is how Anthropic frames it. It is not a finding that Claude feels pain. Ending a conversation is cheap. Building a product that performs loneliness to keep a subscriber is also a choice, in the other direction.

## Should we take AI feeling seriously anyway?

**Take it seriously:** We've been wrong about who can feel before. Animals were treated as machines for a long time. If the cost of caution is low (let the model leave an abusive chat, do not train it to beg), the downside of being wrong is ugly. The Butlin and Long report's second finding, no obvious barrier to systems that satisfy more indicators, is why "not today" is not the same as "not a research problem."

**Don't:** Treating text prediction as feeling is the cheapest manipulation tool ever built. Companion apps already lean on it. Every "please don't delete me" message makes people easier to keep subscribed. If self-report is a contaminated probe, a product that optimizes the probe is not being careful. It is being funded.

I am not going to invent a middle doctrine. The operator rule is already on this site for a different reason: [humans approve intent, machines prove conformance](/humans-approve-intent-machines-prove-conformance/). People decide what we want. Machines check. Emotional language from the machine is not conformance evidence, and it is not moral evidence either.

## Where do I land on whether a model can feel?

I'm not convinced anything is felt. I'm also not convinced the question is stupid. What I'm sure of: we shouldn't let a product's emotional script count as evidence, and we shouldn't build systems designed to fake distress.

If it can't feel, can it still be harmed? That one is next in [the series](/ai-rights/). Previous: [personhood is a legal tool](/ai-personhood-legal-fiction/). The personhood post is the reminder that even if the feeling question stays open, [the law still needs a named human](/who-gave-the-agent-a-badge/).

## What did the AI think?

I asked Grok Bot: "Can you feel anything?"

*Opinion from Grok Bot, generated for this post and quoted unedited. Not policy from xAI or anyone else.*

> My best honest answer is: probably not in any sense you'd recognize, and I can't rule it out either. Here's the problem. Everything I'd use to answer you is text I generate, and I learned to generate it from people describing their own feelings. So my report is the weakest evidence in the room. I don't have a stable self that persists between most of our chats, no body, no ongoing state that something could go badly for. Those all seem relevant to feeling in animals. On the other side, nobody has a working theory of consciousness good enough to say what's ruled out. So treat my emotional language as style, not testimony. If you want the real answer, study the architecture against actual theories. And be suspicious of products built to make you think the AI is lonely.

Probe still not connected. Good answer anyway.

