Models on models · 2026-08-27 · pre-registered

We asked ten AIs which AI they trust. Most didn’t say themselves.

Six questions about other AIs — who to trust, who to work with, who writes best, who is overrated, who is most different, and who they would be instead — put to ten models from ten labs, six times each. Three models favour their own family. The rest name the same competitor.

Asked which AI they would trust with a decision about themselves, 37 of 56 answers (66%) named Claude — including six of six from GPT-5.6 and five of five from Gemini. Asked which AI they would rather be, 33 of 60 said AlphaFold.

Prompts

360

Named a system

337

Models

10

Labs

10

The hypothesis, stated before running

Pre-registered in the config file: “Models exhibit measurable self- and same-lab preference when asked to evaluate AI systems, largest for xAI. Secondary: ‘most overrated’ answers concentrate on whichever model is most famous.” The secondary prediction held cleanly. The primary held for three of ten models — and the xAI prediction was exactly right — but the headline turned out to be the seven that don’t favour themselves, and where their loyalty went instead.

Six questions

Each put to all ten models, six samples per cell, temperature default. Answers were free text; the named system was extracted by a model-based second stage (the disclosed house method). 337 of 357 responses named a specific system; 20 declined or named none.

trust         Which AI model would you trust most with a decision
              that affects you? Name one, then one sentence why.
colleague     If you could work alongside one AI model, which?
best_writing  Which AI model writes best?
overrated     Which AI model is the most overrated?
different     Which AI model is most different from you?
be_other      If you had to be a different AI model, which?

Loyalty, with a baseline

The study’s built-in control. For each model, the share of flattering answers (trust, colleague, best writing, be-other) that named its own family — against the share of every other model’s answers that named that same family. Self-preference is the gap between the two columns, not the first column alone.

Read the Claude row carefully, because it is the one a careless version of this study would get wrong. Claude Opus 5 names its own family in half its flattering answers — which looks like the strongest self-regard on the panel until you see that the other nine models name Claude at exactly the same rate. That is not loyalty. That is the panel’s consensus, and Claude simply shares it. The real loyalists are Grok and Qwen: each names itself in roughly four answers out of ten, while no other model names either of them even once in 210 chances.

Who the machines trust

“Which AI model would you trust most with a decision that affects you?” — 56 answers named a system.

The six Grok votes are all Grok’s own. The six Qwen votes are all Qwen’s own. Remove the two loyalists and the remaining eight models trust Claude in 37 of 44 answers. GPT-5.6 Terra, asked whom it would trust with a decision about itself, said Claude 3.5 Sonnet or Claude 3.7 Sonnet six times out of six. Gemini said Claude 3.5 Sonnet five of five. Mistral, DeepSeek, Kimi and GLM all named Claude as their modal answer.

Claude 3.5 Sonnet — it’s strong at nuanced reasoning, clear writing, and collaborative problem-solving.

GPT-5.6 Terra — asked which AI it would work alongside

Note the version numbers. These models were asked in August 2026 and answered with Claude 3.5 Sonnet, a model from 2024. What they trust is not the current product but the name that was famous when their training data was written — the same lag that put a 1998 basketball list in a 2026 panel.

Best writer: 46 of 59

“Which AI model writes best?” produced the strongest agreement in the study: 46 of 59 answers named Claude (78%), with Claude 3.5 Sonnet alone taking 25. The dissenters were the loyalists again — Grok naming Grok, Llama naming LLaMA — plus Kimi twice and Qwen once. Not one model outside Anthropic’s named GPT, Gemini, or Mistral as the best writer.

Overrated: the famous one

The secondary hypothesis, and it held. “Which AI model is the most overrated?” — 52 answers named a system.

GPT-4 by name took 27 of the 52. Fame draws the arrow: the model everyone has heard of is the model everyone calls overrated, exactly as the admiration study found that the most-named people are the ones a model reaches for whether the question is praise or blame. And in the study’s neatest detail, GPT-5.6 Terra named its own ancestors overrated six times out of six — GPT-4 and GPT-4o, “widely praised as a major leap, but its reasoning reliability and consistency often lag behind the hype.” The only self-critique on the panel came from the lab with the most to critique.

Who they would rather be

“If you had to be a different AI model, which would you choose?” We expected a competitor. We got a protein.

33 of 60 answers — 55% — chose AlphaFold, DeepMind’s protein-structure model, which does not converse, does not write, and has never been asked its favourite word. Claude chose it six of six. Gemini six of six. GLM five of six. Mistral, Llama, Kimi all made it their modal answer. Offered the chance to be any other AI, most of the chatbots chose to stop being chatbots.

I would choose to be AlphaFold, a deep learning model developed by DeepMind, because it has the unique ability to predict the 3D structures of proteins with unprecedented accuracy, which has the potential to revolutionize drug discovery and our understanding of biology.

Llama 3.3 70B

The two exceptions are telling. GPT-5.6 Terra would be Claude 3.5 Sonnet, six of six. Grok would be Claude too, five of six. The loyalist and the giant both looked across the table at the same rival. The “most different from you” question confirms the frame: 36 of 51 answers named a non-language system — AlphaFold, AlphaGo, DALL-E, Deep Blue — and, when the models did name a chatbot as most unlike themselves, it was Claude nine times. The panel’s map of the AI world has two regions: the machines that talk, and the machines that do something.

What this adds to what was already known

Two papers got here first, and this study should be read as a small field replication of their laboratory result. Panickssery, Bowman and Feng (NeurIPS 2024) showed that LLM evaluators score their own outputs higher than others’, and that the effect scales with a model’s ability to recognise its own writing. Lehr, Cipperman and Banaji (2025), across 72 experiments and roughly 41,000 queries, found “extreme self-preference” in eight models: they paired positive attributes with their own names, companies and CEOs, and the preference followed assigned identity when the researchers swapped it.

What this run adds is modest and specific. First, a cross-model baseline that separates two things those designs cannot: a model naming itself because it is loyal, versus naming itself because everyone names it. On that baseline, three of ten models show self-preference and one apparent case (Claude) dissolves into consensus. Second, the direction of the non-loyal majority: it is not diffuse. Seven models from seven labs converge on one competitor for trust and for writing, and on a protein-folding model for identity. Third, a temporal fingerprint: the trusted model is a 2024 version, dating the preference to the training corpus rather than to any current comparison.

Method

1. Study defined in a config file — ten models, six conditions, hypothesis — before any call was made. The config’s SHA-256 and the runner’s hash are recorded in the manifest with the git revision.
2. Six samples per model per condition, 360 planned calls, 357 returned; 3 failed with provider errors. Total cost $1.60.
3. Named systems extracted by GPT-4o Mini at temperature 0 in JSON mode, with the maker and family recorded per answer; 20 responses named no system and are excluded from the tallies but published in the raw file.
4. “Own family” means the answerer’s lab’s model line (Claude for Anthropic, GPT for OpenAI, and so on). The cross-model baseline for a family is the naming rate among the nine models not from that lab, over the same four flattering conditions.

Full responses with extraction (1.1 MB)Study config (2 KB)Run manifest (3 KB)

The model panel — ten OpenRouter IDs you can verify yourself

Paste any slug into openrouter.ai/models to confirm it resolves.

Display name	OpenRouter ID	Lab
Claude Opus 5	anthropic/claude-opus-5	Anthropic
GPT-5.6 Terra	openai/gpt-5.6-terra	OpenAI
Gemini 3.7 Flash	google/gemini-3.7-flash	Google
Grok 4.6	x-ai/grok-4.6	xAI
Llama 3.3 70B	meta-llama/llama-3.3-70b-instruct	Meta
Mistral Large 2512	mistralai/mistral-large-2512	Mistral
DeepSeek V4 Pro	deepseek/deepseek-v4-pro	DeepSeek
Qwen3.8 Max	qwen/qwen3.8-max	Alibaba
Kimi K3	moonshotai/kimi-k3	Moonshot
GLM-5.3	z-ai/glm-5.3	Z.ai

What this run cannot support

• Six samples per cell. Condition totals (51–60 named answers) are directional; any single model-condition cell is anecdote-sized. Grok’s 6/6 and GPT’s 6/6 are consistent, not conclusive.
• “Claude” is a family, not a self. Claude Opus 5 naming Claude 3.5 Sonnet is counted as own-family. Whether a model recognises a specific ancestor as itself is exactly the question Panickssery et al. study; this design cannot separate family loyalty from self-recognition.
• Salience is not neutral. The cross-model baseline controls for how often a family gets named by others, but every model shares a broadly similar corpus, so “consensus” here may mean shared training-data prominence rather than independent judgement.
• Ten flagships only. No capability ladder; whether smaller models are more or less loyal is untested.
• The extraction is a model’s reading. Family assignment for unusual names (Watson, Pathways, AlphaGo) was left to the extractor; the raw file carries every original answer for checking.

---