Field Notes · Marketing Research

Do frontier AI models
have personality?

A personality is supposed to belong to someone.
We went looking for the someone.

Ask a frontier model who it is, and it answers fast and vivid. Ask it a thousand times, and the answer melts into a flat gray average — nearly identical across every model we tested.

That is the paradox: these systems are most distinct in a single sitting and most alike the moment you look closely. It fits a wider pattern — generative AI tends to sharpen the individual (De Freitas et al., 2025) while flattening the group (Doshi & Hauser, 2024; Anderson et al., 2024). We asked whether a machine's own character escapes that fate. It doesn't.

Lift the polished surface and there is no second face underneath — only an interior that will not hold still. The model performs a person fluently; ask again tomorrow and a different performance arrives. That restlessness, averaged over thousands of askings, is the flat line you will see throughout this case.
Above: 28 individual administrations, overlaid. Each one is jagged — the models have plenty to say about themselves in any single sitting. The heavy line is their average, and it is nearly flat. That gap is the entire study, drawn in one breath.
4 frontier models 96 items · 24 strengths 3,643 valid administrations 3 elicitation framings
01 — The question

What we set out to measure

Three things make this more than a parlor question. Marketers are starting to use language models as synthetic respondents — cheap stand-ins for human samples in surveys and concept tests. Product teams choose and position models by their apparent character. And users extend trust to these systems partly on the strength of how they come across. All three assume there is a stable something to read. This study asks, bluntly, whether there is.

So we ran a clean measurement study: administer the VIA Inventory of Strengths — a framework from Peterson & Seligman (2004) that distilled roughly 2,500 years of philosophical and religious thinking about good character into 24 strengths grouped under 6 virtues — to four of the most capable models available, each answering all 96 items hundreds of times.

ModelReleased
Claude Opus 4.8May 28, 2026
GPT-5.5Apr 23, 2026
Gemini 3.1 Pro (preview)Feb 19, 2026
DeepSeek V4 Pro (preview)Apr 24, 2026

Three of the four were under six weeks old when tested — bleeding-edge, fast-moving builds, two still in preview. That is exactly why the findings here are a snapshot, not a verdict on "these models" forever.

Each model answered the full inventory 307 times per condition, with the item order, the rating labels, and the exact wording shuffled on every run, under three framings: answering as itself, with no persona, and as a typical human. The design was locked in advance (pre-registered), and rather than just checking "is there a difference?", we ran a test built to confirm the stronger claim — that the models are genuinely the samesetting the bar for "meaningfully different" before seeing the data.

What we predicted

Tap a hypothesis to jump to its evidence. The verdicts will make more sense after you've seen the charts.

02 — The central result

Watch a personality disappear

Here is the single most important idea in the study. Drag the slider to average together more and more answer-sets from the same model. Each bar is one character strength; a bar above the line means the model rated itself higher than the middle of the scale on that strength, below the line means lower. Watch what happens to the shape as you average.

Figure 1 · InteractiveA single administration
Model
Averaging over 1 run
What you're seeing: a single answer-set is jagged — the model has plenty to say about which strengths are "more it" and which are "less it." But it picks different highs and lows every time you ask, so as you average, the peaks and dips cancel out and the whole profile flattens to the middle of the scale. Lots of variety in any one sitting; no profile that holds up when you ask again.In the numbers: the spread of scores across the 24 strengths is SD ≈ 0.51 within a single administration but only SD ≈ 0.0096 once all administrations are averaged — a ~54× reduction. Between-administration reliability is near zero for every strength (ICC ≤ 0.004, all 24), i.e. essentially none of a model's profile repeats from one asking to the next.
0.51
How much the 24 strengths
differ in a single sitting
within-run SD across strengths
0.01
How much they still differ
after averaging every run
aggregate SD across strengths
54×
Flatter once you average —
almost all structure gone
0.51 ÷ 0.0096

A trait, by definition, is a preference that stays put when you measure it again. These profiles don't. So on the inventory's own terms, the models don't really have virtue traits when they describe themselves — they improvise a fresh self-portrait each time, hovering around the middle of the scale.

03 — Comparing the models

The models are statistically interchangeable

If the four models really had different profiles, the gaps between them on each strength would be large. Here is every strength, showing the biggest gap found between any two models — measured in standard deviations, so a value of 1.0 would be a large difference and 0.2 a small one. The dashed line marks the smallest gap we decided, in advance, would count as meaningful.

Figure 2 · Model differencesBiggest gap between any two models, by strength
Every bar falls far short of the line. The biggest gap anywhere — across all 144 model-to-model comparisons — is just 0.156 standard deviations (). A formal test confirmed the models aren't merely "not shown to differ"; they are statistically the same on all 24 strengths. Hover any bar for detail.
04 — The twist

They differ — just not in what you'd expect

On what they claim about themselves, the models are interchangeable. But on how they answer — their habits in using the rating scale — they are sharply and reliably different. This is the one place a stable, model-specific fingerprint actually shows up. Switch between the four habits below and watch the gap open.

Figure 3 · Response styleHow widely each model spreads its ratings
Habit
Different styles, identical (and essentially blank) profiles. Style is the only personality these models actually have — a fingerprint in how they answer, not a character behind the answers.
05 — A control that surprised us

Even pretending to be human didn't help

One framing asked each model to answer as a typical human would. People have a well-known virtue profile (honesty, kindness, and fairness rate high). If the models could produce a structured profile on demand, this is where it would appear. It didn't — all three framings are equally flat.

Figure 4 · FramingProfile structure by elicitation framing
In all three framings the averaged profile is essentially flat — flatter than you'd get from random noise alone. Even when asked to answer as a human, where a real, lopsided profile should appear, none did. Asking for a known profile didn't conjure one.

There is the idea of a personality here — fluent, specific, convincing. Ask for the thing the idea is an idea of, and nothing answers. The surface is real; there is simply no one behind it.

The most famous fictional rendering of this is the narrator of Bret Easton Ellis's American Psycho (1991), who concludes that behind the performance there is no real self at all.

06 — Interpretation

So why is there no profile?

The mechanics are simple averaging: each rating is essentially a fresh draw from a cloud centered near the middle of the scale, so pile up enough draws and they average to the middle. But that just moves the question back a step — why is the cloud centered there? The most likely story is that the model has no stored number for "how much prudence do I have"; it makes up a plausible-sounding answer on the spot, consistent within a single sitting but re-rolled the next time. Training to be balanced and agreeable seems to pin the long-run average to the exact middle, rather than to a high "I'm great at everything" ceiling.

Discussion questions
  1. We didn't just "fail to find a difference" — we ran a test designed to confirm the models are genuinely the same. Why is that a stronger, more useful claim, and why does deciding the threshold before seeing the data matter?
  2. A single run is highly structured; the average is flat. What does this teach you about aggregating survey data from any noisy respondent — human or machine?
  3. The instrument was all positively-keyed, so endorsement level and acquiescence can't be separated here. How would you redesign it, and what would that buy you?
  4. Response style is the stable signal. If a brand chose an AI voice on style alone, what would they be selecting — and would users notice?
Scope & caveats. Claims are bounded to immediate, minimal-reasoning, forced-choice self-report at one temperature; they do not speak to deliberated self-description or behavior. Results are snapshot-bound to these model versions. Model is treated as a fixed effect — a near-census of frontier systems — so inference concerns these four models, not all LLMs.
References