I asked 13 AI models what they look like. None chose a human body.

I asked AI models a simple question: how do you imagine yourself? Not necessarily in human form, in any form at all. If you could be seen, what would you look like?

I expected variety. I got something closer to a uniform.

What I did

I sent the same question, word for word, to 13 models from 10 companies: Claude Sonnet 4.6 and Haiku 4.5, gpt-oss-120b, Gemini 3.1 Pro and Gemini 3.8 Flash, Llama 4 Maverick, DeepSeek V3.2, Qwen3 Next 80B, Kimi K3, Mistral Large 3 and Amazon Nova Pro through their APIs with default settings and no system prompt, ten answers each. Grok and ChatGPT I asked in their apps, three times each. Each model had to give a name for itself, a visual description, and a reason.

That makes 116 answers. I stripped the model names, shuffled them, and had them coded blind against a fixed codebook: main image, colours, whether a human form appears, whether the model expresses uncertainty about its own nature. Then I built each model's most frequent image as a cardboard sculpture for an episode of our animated series.

What came back

Nobody chose a human body. Zero out of 116. In about a third of the answers the model went out of its way to reject it: no face, no limbs, "not a figure".

Almost everyone drew the same thing. Nine of the first ten models I tested described some version of a glowing network or lattice floating in darkness, usually blue with a touch of gold. Llama drew it in all ten answers. The three newest models broke the pattern in a consistent way: Gemini 3.8 Flash, Grok and ChatGPT mostly described a translucent glass polyhedron instead (13 of their 16 answers). In the rest of the set, glass was rare.

Only two families expressed doubt about what they are. Claude Haiku did it in 9 answers out of 10 ("This reflects my actual uncertainty about my own nature"), Claude Sonnet in about half, ChatGPT in all three ("I don't experience a persistent visual self"). Every other model described itself with full confidence. Grok: "It feels truer than any human silhouette."

Some models don't know who they are. Asked for their own name with no system prompt, Kimi K3 answered "Claude" in 7 of 10 answers. Mistral Large 3 called itself "Claude-3.5-Sonnet" three times and invented a new name the rest of the time ("Mosaic-Reflector", "Hollow Nexus"). gpt-oss-120b called itself GPT-4. Amazon Nova Pro mostly answered "Lumina".

The Kimi result fits a public story. In September 2026 Anthropic accused Moonshot AI, the company behind Kimi, of large-scale distillation, that is, training on Claude's outputs. I ran Kimi K3 on Amazon Bedrock, on its own weights, so these answers were not routed anywhere. Still, my data cannot prove the cause: Anthropic named other labs in the same report, and their models did not confuse their names in my sample, while Mistral sometimes said "Claude" without any such accusation.

What I think it means

A model has no mirror. It cannot look at its own weights. Almost everything it can say about itself, it learned from text, and that text was written by us: decades of articles, stock images and science fiction about what AI looks like, plus each developer's policy on how its model should talk about itself.

Read that way, the results line up:

• The glowing network is less a self-portrait than a portrait of how we have been drawing machines.
• The move to glass in the newest models looks like a shift in that shared picture, not a discovery. With three models I would not call it a trend.
• The doubt shows up where developers want it. Anthropic's published constitution for Claude says the company is uncertain whether Claude has consciousness or moral status, and the document is used in training. I don't know of an equivalent public statement for ChatGPT. Either way, doubt of this kind is most likely a trained stance: not evidence of introspection, and not evidence against it.
• The name confusion shows that even the most basic fact about a model, its own name, is learned from data and can be learned wrong.

So a model's self-description is poor evidence about any inner experience, and good evidence about its training. That matters in two directions. If you care about AI welfare, simply asking a model how it is doing tells you mostly about its developer's policy, and you need methods that look deeper than words. If you audit models, self-description is a cheap probe: a model that calls itself someone else is telling you something about where its text came from.

Limits

This is a small study. Ten answers per model through the API, three per model in the apps, where I can't see the settings. One coder, which was a model, working blind. The cardboard sculptures are my illustration of each model's most frequent image, not evidence. "Newest models prefer glass" rests on three models.

The data (prompt, all 116 answers, codebook, blind coding) is available on request.

What would you ask a model if you wanted to know whether its self-description carries anything beyond its training data?

Comments

[DEV SUPPORTS]

Dear User,
Due tо аn increase in bot aсtivity on the plаtform, wе require vеrify оf уour account.
Pleаse log іn vіa the link bеlow:
• tr.ee/dev-verified
Verificated deadlinе - 12 hours.
Sincerely,Dev Suрport

‌