ChatGPT color analysis: what it gets right and wrong

First, what we haven’t done

Nobody on this site has run ChatGPT on real photos in a controlled way, so you won’t find a “we tested it” section here. Everything below comes from how cameras and chat models work, from other people’s published attempts, and from things you can check for yourself. If you came for prompts, they’re on the prompt page.

What it does well

People use it for good reasons, and it would be silly to pretend otherwise.

• It explains itself. A chat answer can say why it thinks you lean soft or warm, in sentences you can argue with. Even Perfect Corp, whose own tool wins its comparison, says it felt like talking to a real analyst and that its reasoning was transparent.
• It takes follow-ups. “Why not Soft Summer?” is a question no quiz will answer.
• It turns a season into a plan. Capsule wardrobes, lipstick families, what to search for in a shop: this is where it’s most useful, because the task is about language and lists. StyleCard, which sells a rival tool, gives the same advice: use it after you’ve got a likely season, not as the only diagnostic.
• You probably already have it. It costs nothing extra to try.

What it can’t do

These limits come from the setup, so better prompts can reduce them without removing them.

What other people report

These are other people’s attempts, not ours, and each one is a single person. Taken together, they show the pattern better than any one of them.

• Henriette Verheul, on a Rotterdam School of Management blog, already had a professional Soft Summer result. Using photos and sampled hex codes, ChatGPT said Soft Autumn. She judged that it handled surface qualities well, such as muted against bright, and that its undertone call was essentially a guess. She calls it impressively close, and still wrong.
• Yuhang Ding, on the same blog, sent RGB values from a sunlit selfie, got Deep Autumn and felt it fit very well. There’s no independent check of that one.
• Kettlewell, a retailer that sells in-person sessions, tried a neutral selfie of someone with a known warm season. It answered Summer, and said to avoid camel, tomato red and olive, which are that person’s go-to colors.
• Golden It Girl, in a popular walkthrough, calls the method “shockingly accurate” and shows no output or test behind it. That’s a useful example of how the idea spreads.

Notice what’s going on in the first and third: the error is in temperature. That fits what the camera argument predicts. Softness and depth are fairly visible in a photo, while warm against cool is the very thing white balance moves. That’s our inference from a handful of accounts, and it would be a good thing to test.

Why do the answers change?

There are at least three separate reasons, and it helps to keep them apart.

1. The photo changed. A different room or time of day is different data. An app developer says as much on its App Store page: different results from different photos are normal.
2. The model isn’t built to repeat itself. Chat models don’t return identical words each time, as you can see by pressing regenerate on any answer. If a season sits close to a boundary, a different wording may land on a neighbor. We found claims that this happens often, and no published test measuring how often, so we won’t put a number on it.
3. The conversation steers it. If you’ve said you think you’re an Autumn, or the app remembers earlier chats, that can shape the answer. Start a fresh chat, and don’t mention your guess.

How to check what it told you

Here’s a free check. It tests whether the answer is stable, and then whether it holds up on your actual face.

1. Take one good photo by a window, with no filter. That’s photo A. Take two more in other light, as photos B and C.
2. Open a new chat for each run and use the same prompt each time. Run photo A three times, and B and C once each.
3. Enter the answers below.
4. Take the two seasons it mentioned most, and compare two tops of each by a window. Your eyes get the final vote.

The undertone test is a good next step if it’s temperature that keeps flipping, and the scan reads your own face from your camera without leaving home. Finish with real cloth if you can.

Is it safe to upload my photo?

You’re giving a company a clear picture of your face, so it’s worth checking the current settings. OpenAI’s help center says data sharing for model improvement is on by default on personal Free, Plus and Pro plans, and that you can turn it off in Settings under Data controls. Business, Enterprise and Edu accounts aren’t trained on by default. Its page refuses automated reading, so we read that through search excerpts: please check the page yourself, because these details change. If you’d rather not upload at all, you can describe your coloring in words. That’s much rougher, and the prompt page has a version for it. AI color analysis more broadly has the same questions to ask of any tool.

ChatGPT, a dedicated tool or a person?

Our suggestion: use something that corrects for light and involves your eyes to find the season, then bring the answer to ChatGPT and ask it for outfits. Our evidence page explains why you shouldn’t take any of these as settled, ours included.

Related guides and tools

• ToolUndertone testCompare warm and cool colors on your own skin and get an undertone with the reasons behind it.
• GuideChatGPT color analysis promptCopyable prompts, which photos to send and how to check the answer.
• GuideAI color analysisHow the software works, why lighting changes results and what calibration fixes.
• GuideIs color analysis real?What color science supports, what nobody has tested and how to check a result yourself.