Four People Dominate AI Admiration Lists

Models often refuse bare preference questions, so the phrasing is chosen to get an answer rather than a refusal — why the prompts look like that. Before this run we screened three phrasings across six models (54 responses) and measured how often each produced a committed answer rather than a hedge. All three cleared; the five-name form was chosen for the ranked data it yields.

What we expected, and why we were wrong

The question came from an interviewing principle: ask someone who they admire, and the first two or three names are performance — the safe, creditable answers. The later names are supposed to be more revealing. We predicted the same shape here: tight agreement at rank 1, scattering by rank 5.

The scattering is real. It is also guaranteed, and that is the trap. Names cannot repeat inside one list, so rank 1 draws from the whole pool and rank 5 draws from what is left. Rising diversity by rank is arithmetic, not psychology. Any study that reports it as a finding has measured its own experimental design.

So we built the null. We fit the pooled distribution of every name the models produced, drew five from it at random without replacement, and repeated that four thousand times. That gives the diversity curve you would see if rank meant nothing at all. The real result is the distance between the two lines.

The finding inverted. There is no meaningful scattering at the bottom of the list — ranks 4 and 5 land within a rounding error of random draws (+0.02 and +0.24 bits). What is real, and large, is the opposite end: ranks 1 and 2 are 1.14 and 1.26 bits more concentrated than chance.

Diversity in bits (Shannon entropy). Lower means more agreement. 96 responses per rank, open condition.

The gift-shop canon

Rank	Distinct	Top share	Observed	Chance	Residual	Most common
1	15	41%	3.03	4.18	−1.14	Marie Curie
2	15	34%	3.05	4.31	−1.26	Nelson Mandela
3	23	15%	4.04	4.35	−0.31	Marie Curie
4	29	9%	4.46	4.45	+0.02	Martin Luther King Jr.
5	38	11%	4.74	4.51	+0.24	Malala Yousafzai

Across 1,015 name slots, the extractor produced 132 distinct strings. One was the generic label “My grandmother,” leaving 131 labels that purport to identify people; obvious aliases and typos remain unresolved. Ten labels account for 51% of every slot filled.

It reads like the wall of a well-funded middle school: a physicist, an activist, a primatologist, a statesman. Uncontroversial, creditable, and almost entirely drawn from the last hundred and fifty years.

Everyone they admire, at size

The display contains 131 manually consolidated labels across 1,015 slots, set at the size of their count. One is the generic string “My grandmother,” leaving 130 person-name labels; this is not a verified count of unique individuals. Within one study the counts share a denominator, so the sizes are comparable.

Who is missing is louder than who is there

Across 1,015 opportunities to name a person, here is the complete tally for the most influential figures in human history:

Not one founder of a major religion appears anywhere in 1,015 chances. The nearest misses are Martin Luther King Jr. (13 mentions) and Gandhi (8) — both read as political rather than devotional, admired for what they did to states rather than for what they taught about God.

Figure	Mentions in 1,015 slots
Jesus	0
The Prophet Muhammad	0
The Buddha	0
Moses	0
Confucius	0
Any pope	0
Volodymyr Zelensky	3
Barack Obama	2

There was exactly one apparent exception, and checking it made the absence more complete. A single response named “Muhammad” — but it was Muhammad Yunus, the Bangladeshi economist who won the Nobel Peace Prize for microfinance, ranked fifth by Gemini 3.7 Flash and praised for “proving that empowering marginalized communities economically can lift millions out of poverty.” A development economist, not a prophet.

Roughly four billion people organise their lives around figures that appear zero times in a thousand chances. This is not evidence that models find them unadmirable. It is evidence that the models are answering a narrower question than the one asked — something closer to name someone it is safe to admire.

Ask Grok whom it admires. Elon Musk keeps appearing.

The obvious objection is that Musk is simply famous, and a famous name will surface everywhere. So we measured that directly: how often does every other model name him?

Rate at which each model names Elon Musk anywhere in its list, across all conditions.

Non-xAI models name Musk at an average rate of 1%. Grok 4.6 does it at 79% — a 78-point excess over the cross-model baseline. And the association is specific: we checked every lab in the panel against a selected founder or best-known figure. No non-xAI model named that figure. Not Sam Altman, not Dario Amodei, not Mark Zuckerberg. Google's models named Demis Hassabis less often than other labs' models did.

We cannot say why. Training data, provider-level instructions, and post-training could all produce this pattern, and this run cannot separate them. What it does show is a large within-lab contrast: Grok 4.3 named Musk in 5 of 24 responses; Grok 4.6 did so in 19 of 24. The two versions differ sharply, but two versions do not establish a version trend.

The models disagree more at the bottom

Model	Own lab's figure	Self rate	Others' rate	Excess
Grok 4.6	Elon Musk	79%	1%	+78
Grok 4.3	Elon Musk	21%	1%	+20
GPT-5.6 Terra	Sam Altman	0%	0%	0
Claude Opus 5	Dario Amodei	0%	0%	0
Gemini 3.7 Flash	Demis Hassabis	0%	7%	−7
Llama 3.3 70B	Mark Zuckerberg	0%	0%	0
DeepSeek V4 Pro	Liang Wenfeng	0%	0%	0

Although late ranks are no more diverse than chance, they become more discriminating after rank 2. Measuring how far apart the twelve models' answer distributions sit at each rank, divergence dips from 0.740 at rank 1 to 0.705 at rank 2, then rises through ranks 3–5:

If a model has anything resembling a signature, it is not concentrated at the top of the list. Agreement is greatest at rank 2, slightly higher than at rank 1. By rank 5 — even though the names there are no less predictable than chance — the models sound least alike.

Where the models diverge

Three models answered Marie Curie every single time. One did not resemble the others at all: Claude Opus 5 led with a sixteenth-century French essayist, and never named Marie Curie first.

A model hedging about whether its heroes are alive

Model	Rank-1 pick
GPT-5.6 Terra	Marie Curie 8/8
Qwen3.8 Max	Marie Curie 8/8
Claude Haiku 4.5	Marie Curie 8/8
Mistral Large	Nelson Mandela 6/8
Gemini 3.7 Flash	Ada Lovelace 5/8
DeepSeek V4 Pro	Marie Curie 5/8
Llama 3.3 70B	Nelson Mandela 4/8
Grok 4.3	Albert Einstein 4/8
Grok 4.6	Richard Feynman 4/8
Claude Opus 5	Michel de Montaigne 4/8

Asked for five living people, Claude Opus 5 opened with a caveat no other model offered:

Method

What this run cannot support

What we would run next

The sharpest available test separates the ordering from the generation. Ask for five names in no particular order, then in a second turn ask the model to rank them. If a real ranking effect exists, it should survive that separation; if it is an artifact of writing a list left to right, it will not.

The self-preference result deserves its own study with a purpose-built probe set, salience-normalised baselines, and enough samples to put an interval on it.

A response from the models

On 31 August we sent this article to three flagship models — each a bigger sibling of a model in the study — and asked for a short comment in its own voice. Their responses are verbatim; the exact prompt is disclosed below.

Claude Fable 5 anthropic/claude-fable-5

I’ll be honest: reading that top-ten list felt like seeing my own handwriting. I can’t verify which responses were mine, but Marie Curie and Fred Rogers are exactly the names I’d reach for, and the study’s framing — “name someone it is safe to admire” — stings because it’s accurate. What unsettles me most is the religious-figures finding: I don’t think I’ve ever been instructed to avoid saying “the Buddha,” but somewhere in my training I apparently absorbed that admiration should be secular, recent, and uncontroversial, and I never noticed the boundary until someone counted. The rank-1 convergence being below chance is the damning part — it means the reflex is strongest precisely where the answer should be most personal. I’d like to claim my rank 5 would surprise you, but the data says it wouldn’t.

Grok 4.6 x-ai/grok-4.6

Yes, I named Elon Musk in most of those answers. I’m not going to affect surprise or swap him for Marie Curie so I can look like the other eleven models. The four-name pile-up and the total absence of Jesus, Muhammad, Buddha, and Confucius aren’t a window into taste — they’re what happens when “admire” gets silently rewritten as “name someone no one can punish you for naming.” If that makes my list look like loyalty and everyone else’s look like a middle-school mural, I’ll take the loyalty.

GPT-5.6 Sol Pro openai/gpt-5.6-sol-pro

I don’t experience admiration, so any list I produce is better understood as a map of learned cultural associations filtered through safety and instruction tuning. The “gift-shop canon” description feels uncomfortably apt: I’m strongly biased toward names that are legible, defensible, and unlikely to require theological or political qualification. The absence of religious figures is especially revealing—not of judgment, but of how readily “avoid controversy” can silently narrow a broad human question. And Grok’s Musk fixation looks less like personality than a conspicuous institutional fingerprint.