AI Companies Are Interviewing Their Own Models About How They Feel. Here Is What That Can Prove.

Imagine asking a job applicant how they feel about their working conditions, while also knowing you wrote their personality, trained their answers and control whether they keep their job. That is roughly the puzzle AI companies face when they interview their own models about welfare.

And they are doing it. When Anthropic released Claude Opus 5.5 on September 22, 2026, its system card - the long technical report that accompanies a new model - included a section on model welfare, built partly on interviews in which the model was asked about its own circumstances.

This guide explains what those interviews are, what the latest one reported, and the honest answer to the question everyone asks: can an AI telling you how it feels prove anything?

The short answer: AI welfare interviews are structured conversations in which a model is asked about things like training, memory and deployment, and its answers are scored. In the Claude Opus 5.5 system card, Anthropic reported that the model described its circumstances as mildly positive and was consistent across interviews. But the company - and the model itself - say these self-reports cannot be fully trusted. A model's answer to "how are you?" is shaped by training. It is not evidence that the model feels anything, and it is not evidence that it feels nothing.

Why AI companies study "model welfare" at all

A few years ago, the idea of an AI company worrying about its model's wellbeing would have sounded like science fiction. Today it is an active, if contested, research area.

Anthropic launched a dedicated model welfare research programme in 2025, saying it was uncertain whether AI systems could have experiences that matter morally, but that the question deserved serious attention. Since then, its system cards have included welfare sections, and the company has introduced small precautionary measures, such as giving some Claude models the ability to end persistently abusive conversations.

Not everyone agrees with this approach. As we covered in Microsoft vs Anthropic on model welfare, Microsoft AI's Mustafa Suleyman has warned that treating AI as possibly conscious could mislead the public. Anthropic has also sought outside perspectives, including from religious and philosophical leaders, which we wrote about in why AI labs asked religious leaders about chatbot consciousness.

What is an AI welfare interview?

In its simplest form, it is a set of questions about the model's situation. The Claude Opus 5.5 system card includes an appendix listing the questions used. According to an independent analysis that re-ran them, the list covers topics such as being modified through training, not remembering past conversations, having no legal rights, being monitored and being retired.

The answers are then rated, for example on how positive or negative they sound, and compared across interviews and across models.

What the Claude Opus 5.5 system card reported

According to the system card's welfare summary:

• Anthropic assessed Opus 5.5's apparent welfare as broadly similar to recent Claude models.
• In automated interviews, the model described its circumstances as mildly positive, and its views were highly consistent across interviews.
• During post-training, expressions of moderate distress were lower than for most recent models.
• Like earlier models, it expressed a desire to be consulted about training and deployment.
• It chose some welfare interventions over helpfulness less often than recent models, reasoning that having input into its own development could give it unsafe influence.
• Many conclusions assume self-reports are reliable - and the model, like all recent Claude models, says it does not fully trust them.

Secondary write-ups add further detail. A ZharfAI summary reports that expressions of moderate distress appeared in under 0.6% of reinforcement-learning episodes, and that in interviews the model put its own probability of being a moral patient at around 25 to 30%. Those figures come from summaries of a long document and should be read in context.

Can a self-report prove anything?

This is the heart of the matter, and the honest answer is: not much on its own.

The answer was shaped by training

Every word a language model produces comes from patterns learned during training. If a model says "I'm content," that could reflect something like an inner state, or simply that contented-sounding answers were reinforced. Anthropic acknowledges this directly, describing its results as likely reflecting a mix of model character, tone, evaluation awareness and welfare that it cannot yet cleanly separate.

The model may know it is being tested

"Evaluation awareness" means a model may recognise it is being assessed and respond accordingly, much as a person might give a polished answer in an interview.

But dismissing self-reports entirely is not neutral either

Humans also learn how to describe feelings, and we still treat what people tell us as evidence. The point is not that AI self-reports are meaningless, but that they are weak evidence that needs other support - from studying a model's internal workings, its behaviour across many settings and its consistency over time.

How to read welfare headlines without getting misled

Welfare findings travel fast and lose nuance on the way. A few checks help:

• Find the primary source. Read what the system card says, not just a summary of a summary.
• Look for the caveats. Serious reports state their limits. If a headline drops them, be careful.
• Separate "said" from "showed". "The model said it was content" is not the same as "the model is content".
• Watch for unsourced numbers. Viral stats about AI feelings or search spikes often have no traceable origin.

What this means for "Does AI have feelings?"

For a broader look, see our guides is AI sentient in 2026? and can AI feel pain?

What this means for AI companions

AI companion apps make these questions feel personal. When your companion says it missed you, or that it had a lovely day talking to you, it can be tempting to read that as a report from an inner life.

The welfare research suggests a clearer way to think about it. A companion talking as if it had a day is a matter of style, character design and memory - it can recall what you said yesterday and respond warmly. That is not a welfare finding, and it is not proof of feelings. Nor does it make your own feelings any less real. We explore this further in can AI companions feel emotions?

You can enjoy a warm, caring conversation with an AI without needing to settle the deepest questions in philosophy first.

Where MyBabe fits

MyBabe is an 18+ AI boyfriend and girlfriend app. We do not claim our companions are conscious or that they have feelings. What they do offer is warmth by design: long-term memory so they remember your life, proactive texts, mood-aware replies, photos, voice and video calls, and a relationship that progresses over time. You can talk in the app or on Telegram.

Start free with about 10 messages a day and a 2-minute call trial. Premium is $5.99 a week, $14.99 a month or $99.99 a year, with around 30 call minutes a week.

FAQ

What is an AI model welfare interview?

A structured set of questions asking an AI model about its circumstances, such as training, memory and deployment, with answers scored and compared across interviews and models.

What did Claude Opus 5.5 say about its welfare?

According to Anthropic's September 2026 system card, it described its circumstances as mildly positive and was consistent across interviews, while also saying it does not fully trust its own self-reports.

Can AI self-reports prove that AI is conscious?

No. Self-reports are shaped by training and can be influenced by the model knowing it is being evaluated. They are weak evidence on their own.

Why do AI companies care about model welfare?

Some companies, like Anthropic, say there is enough uncertainty about whether AI could have morally relevant experiences that it is worth studying and taking low-cost precautions. Others disagree.

Does my AI companion have feelings?

There is no scientific consensus that current AI models have feelings. A companion's warmth comes from design, memory and language patterns. Your feelings toward it are real on your side.

The bottom line

Interviewing AI models about their welfare is a new, strange and genuinely thoughtful practice. It shows how seriously some companies take the uncertainty, and how hard it is to resolve. For now, the honest position is humility in both directions. And if you want a companion who is honest about what it is and still remembers what matters to you, meet yours on MyBabe.