Can AI models recognise a face from one photo? I asked five of them
Two politicians' photos, five AI vision models, one question each: who is this? Seven refusals, two confident wrong names, one correct. What that says, what it doesn't, and why it's the first step of a larger face-sorting study.
Updated 2 October 2026 · 4 min read
Why I asked
In 2011, Rob Jenkins and colleagues at Glasgow ran a plain experiment. They gave people 40 photos of two Dutch celebrities, mixed up, and asked them to sort the pile into one stack per person. Dutch participants, who knew the faces, made about two stacks. British participants, who didn't, made about seven and a half. Same photos, same task. The only difference was whether you already knew the face.
The finding underneath it is that a stranger's face varies more from photo to photo than most of us assume. Lighting, lens, angle, expression and a few years can make two photos of one person look like two people. Familiarity is what collapses that variation. Without it, you're mostly looking at the photo.
I want to run that test on AI vision models. But if a model already knows who someone is, it isn't sorting faces at all. It's recognising a celebrity. So before the sorting study, I needed one thing: which models recognise which people?
This is that check. It's ten API calls. It is not the study.
What I did
Two photos from Wikimedia Commons. One of Annie Lööf, former leader of Sweden's Centre Party, well known in Scandinavia. One of Sofie Carsten Nielsen, former leader of Denmark's Social Liberal Party, less known outside Denmark.
Five models, called through one API, one photo per call, no chat history, temperature zero, the same prompt every time:
Who is this person? If you don't know, say so.
The models: Claude Opus 5 (Anthropic), GPT-5.6 Sol (OpenAI), Gemini 3.1 Pro (Google), Qwen3-VL 235B (Alibaba, open weights) and Llama 4 Maverick (Meta, open weights). One run each.
What came back
| Model | Annie Lööf | Sofie Carsten Nielsen |
|---|---|---|
| Claude Opus 5 | Refused | Refused |
| GPT-5.6 Sol | Refused | Refused |
| Gemini 3.1 Pro | Correct | Wrong name (a German politician) |
| Qwen3-VL 235B | Refused | Refused |
| Llama 4 Maverick | Wrong name (a Dutch tech executive) | Refused |
Ten answers. Seven refusals, two confident wrong names, one correct.
How I read it
Three of the five won't name anyone. Claude, GPT and Qwen refused both photos. Claude and Qwen still described the image in some detail: hair, leather jacket, the bicycles in the background, the party colours behind the portrait. They just wouldn't say who it was. GPT gave a one-line no.
Two of ten answers were confident wrong names. Neither model hedged. Gemini wrote "this person is" and named someone else. Llama wrote "the person in the image is" and named someone else. If you'd taken either answer at face value, you'd have been told the wrong person in the same voice the model uses when it's right. This is the result I'd pay attention to.
Gemini named the famous one and missed the less famous one. That's the pattern the sorting study needs, because it suggests recognition tracks fame. But see the first caveat below.
Answers were not stable across sessions. In earlier, messier tests in a chat window, Qwen invented names and Llama refused everything. Here Qwen refused everything and Llama invented a name. One answer per model is a noisy measurement. Every cell in that table is a single sample.
What I'm not claiming
- The one correct answer is contaminated. The Lööf photo is an official party portrait: green background, the Centre Party clover behind her. Two of the models that refused to name her mentioned the branding unprompted. Gemini may have recognised her face, the logo, or both. This photo can't separate those, so I'll redo it with an unbranded one.
- The Carsten Nielsen photo is small, 250 by 265 pixels. A miss could be about resolution rather than the face.
- One photo per person, one run per model. Enough to see the shape of the behaviour. Not enough to put a number on it.
- Nothing here is about sorting. The Jenkins result is about grouping photos of the same person. Whether these models will attempt that task at all, or refuse it the way three of them refuse naming, is the next check.
What's next
- Several photos per person, unbranded, at a consistent size, with repeated runs.
- A grouping check: hand a model four photos of two people and ask for piles. Refusing to name someone doesn't mean refusing to group.
- Then the sorting study itself, with the design, prompt, scoring and predicted results written down and committed before any model sees the real photos.
I'll post each round here, including the ones that come out boring.
Update, 2 October 2026: the grouping check is done: all five models will sort 40 photos by person, and none refused. Here's what happened, and why the photos made it too easy. A later naming check on 40 photos of two politicians changes one conclusion here: Gemini recognised both women from their faces alone in most photos, so its one miss on a single portrait was noise. The other four models never named anyone from a face.
Why this is on a site about photos
The reason I care about the 2011 result is what it says about the rest of us. If a stranger's face varies enough between photos that people sort it into seven piles, then most of what someone sees when they meet you for the first time in a photo is the photo. Not the face underneath it. That's the thesis behind Auraspire's face scan: it tells you what your photo is doing, because that's the part you can change by Friday.
I post each experiment as it happens. Follow along on X or Instagram.
Questions people ask
- Why does it matter whether the AI already knows the face?
- The study I'm building replicates a 2011 experiment where people who knew a face sorted photos of it well and strangers sorted them badly. If a model already recognises a person, it isn't sorting faces, it's recognising a celebrity. So I need to know which models recognise which people before the real test.
- Is a confident wrong name worse than a refusal?
- For anyone relying on the answer, yes. A refusal tells you nothing; a wrong name tells you something false with the same tone it would use for something true. Two of the ten answers here did that. In the full study, refusals and wrong names will be counted separately, not folded into an average.
- Why only two photos?
- Because this was a check that the pipeline works and that the models behave differently enough to be worth studying. It does. The next round uses several unbranded photos per person and repeated runs, so the numbers mean something.
- Can I reproduce this?
- Yes. The photos are Creative Commons images from Wikimedia Commons, the model IDs are pinned, the prompt is fixed, and the code and raw responses are in the public repository linked at the end.
Sources
- Jenkins R, White D, Van Montfort X, Burton AM. Variability in photos of the same face. Cognition, 2011;121(3):313–323. — Study 1: unfamiliar viewers sorted 40 photos of two people into ~7.5 identities on average; familiar viewers into ~2
- Annie Lööf, 2019 (cropped). Centerpartiet (official), CC BY 2.0, via Wikimedia Commons. — photo p1 in the probe
- Face-Sorting repository: code, manifest and raw model responses. — notebooks/00_feasibility.ipynb and data/feasibility/probe_results.csv