Key points
- In a study of 86,220 volunteers, a computer model reading Facebook Likes agreed with people's own test scores at r = 0.56, compared with r = 0.49 for their Facebook friends (Youyou, Kosinski and Stillwell, 2015).
- A model trained on the language of 66,732 Facebook users predicted Big Five scores in a separate group of 4,824 people, and the language-based scores agreed with self-reports (Park and colleagues, 2015).
- GPT-3.5 and GPT-4, with no special training, inferred Big Five traits from Facebook posts with an average correlation of 0.29 with people's own test scores (Peters and Matz, 2024).
- In all three studies the yardstick was a real questionnaire, and each person had far more text or data than one chat gives.
Leo's two answers
Leo is 29 and tests software for a living. One evening he asks his AI chatbot: "From everything we've talked about, what's my personality type?" It answers with confidence: analytical, independent, a planner. He likes it.
A week later, in a new chat, he asks the same question. This time he's warm, curious and spontaneous. Both answers sound a bit like him. Neither sounds exactly like him.
Leo isn't a real person, but his two answers show the difference between a conversation and a measurement. Here is what chatbots do well, where they fall short, what research says, and how Leo got a result he could compare.
What the research says
Computer predictions from Likes matched people's test scores at 0.56, more closely than their friends' judgments at 0.49, across 86,220 volunteers.
Youyou, Kosinski and Stillwell, 2015Facebook users' words trained a model that then predicted Big Five scores for 4,824 others.
Park and colleagues, 2015Average correlation between GPT-inferred and self-reported Big Five traits, from Facebook posts with no special training.
Peters and Matz, 2024Computers can read some personality from digital traces. But every one of these studies checked the computer's guess against a real test. The test is still the yardstick.
What an AI chatbot does well here
To be fair to the AI, it is genuinely useful for this:
- Explaining tests. What the Big Five traits mean, how MBTI letters work, what DISC measures, how they differ.
- Talking through a result you already have. Paste in your report and ask what it means for a decision, a conversation with a colleague, or a career move.
- Reflecting your writing back to you. Asking "what do my messages suggest about how I work?" can surface patterns you hadn't noticed.
- Running a quiz on the spot. It can write questions and give you a type at the end.
Where a chat guess falls short
| A chat guess | A measured test | |
|---|---|---|
| Questions | Can change every time | The same items, or the same kind of choices, for everyone |
| Scoring | Can change every time | The same rules every time |
| What it sees | Only what you typed in that chat | Your answers to items built for the trait |
| Comparison | None | Your answers next to many other people's |
| Same result next week? | Often not | Designed to be stable |
- It's not a fixed test. A chatbot can write different questions each time and score them differently, so two chats can give two answers, as Leo found.
- It sees only what you typed. If you mostly ask it for help with code or spreadsheets, that's a narrow slice of you.
- It tends to be agreeable. Chatbots often describe you in kind, general terms most people would recognize in themselves. That feels accurate without telling you much.
- There are no norms behind it. "High openness" means something only when compared with how many other people answered the same items.
None of this is a flaw in the AI. It is the difference between a conversation and a measurement.
What research says about reading personality from language
Inferring personality from text and online behavior is a real research field.
- Computers can read some personality from digital traces. Using Facebook Likes from 86,220 volunteers, a computer model's predictions agreed with people's own questionnaire scores more (r = 0.56) than their friends' judgments did (r = 0.49) (Youyou, Kosinski and Stillwell, 2015).
- Language works, moderately. A model trained on the words of 66,732 Facebook users predicted the Big Five in a separate sample of 4,824 users. Its scores agreed with self-reports and with friends' reports, and stayed stable over six months (Park and colleagues, 2015).
- Large language models can do it too. GPT-3.5 and GPT-4, with no extra training, inferred Big Five traits from Facebook status updates with an average correlation of 0.29 with self-reports (range 0.22 to 0.33), similar to models built for the task. Accuracy was higher for women and younger people on several traits (Peters and Matz, 2024).
A correlation of 1 would be a perfect match and 0 no link at all. So 0.29 is clearly better than chance and far from a replacement for the test itself.
Two things follow. First, these studies compare the AI's guess against a real test, so the test is still the yardstick. Second, they worked from a lot of data per person: months of posts or Likes. A short chat gives much less to go on.
Mind your privacy
Before you paste personal messages, emails or diaries into a chatbot to "read" your personality, check what your chat settings do with your history. The Peters and Matz study itself warns that cheap personality inference from text raises privacy concerns.
If you want to use AI anyway
That's fine. Four habits make the result clearer and easier to check. Here is how Leo tried them.
1. Ask it to use a published questionnaire
For example the public-domain IPIP Big Five items, with the items shown and scored the same way each time.
Leo: he asked for the 50 IPIP items, answered them one by one, and saved the scores.
2. Ask how sure it is
Ask what in your answers or writing it bases the guess on, and how confident it is.
Leo: it said it was "moderately confident" and pointed to three of his answers.
3. Treat the result as a starting point
Check it against a test built for the job.
Leo: he took Tipolu the same evening and put the two results side by side.
4. Bring the real result back
Paste your test report into the chatbot and ask what it means for you. That is where the AI shines.
Leo: he asked how his result might affect the way he gives feedback to developers.
Where Tipolu fits
Tipolu is a test, not a chat guess. Everyone walks the same kind of 3D trail and makes two-path choices at 12 to 20 crossroads, scored the same way, plus two short open questions at a campfire. You get one of five types (Peak, Horizon, Campfire, River, Root), an accent, and a 0 to 100 score on five traits from the Big Five family. It is free, needs no signup or email, and takes about 5 minutes in the browser. The two campfire answers are used only to read your result and are deleted after 90 days.
It is a self-knowledge game, not a clinical or diagnostic tool. A validation study is planned, and Tipolu won't publish accuracy numbers until it's done. For more on how tests differ, see the most accurate personality test and personality quiz vs test.
How each type tends to use a chatbot for this
Peak: Takes the first answer and moves on. Ask a second time in a new chat before you believe it.
Horizon: Enjoys the conversation more than the result. Ask for a published questionnaire so you end with something you can compare.
Campfire: Likes the warm description. Remember that a kind, general description fits almost everyone.
River: Reflects on what the AI says. Check it against a real week, not just against how it feels.
Root: Wants the method. Ask which items, which scoring and how confident it is.
If Leo took the test
Rootwith a Horizon accentA Root with a Horizon accent explains both of Leo's chat answers. The first chat saw the planner (Root). The second saw the curious side (Horizon). A measured result put both in one picture, with the order clear.
Leo's scores are made up to show what a result looks like.
Lu's tip: Take the test first, then ask your chatbot about the result. You get a measured type and a patient explainer.
Questions people ask
Can ChatGPT give me a personality test?
Yes. It can write and score a quiz, or run a published questionnaire if you ask. The result can vary between chats, so treat it as an informal read, not a measured type.
Is there a good prompt for an AI personality test?
Ask it to use a published Big Five questionnaire, to show you the items, to score them the same way each time, and to tell you how confident it is. That makes the result clearer and easier to check.
Can AI tell my personality from my writing?
Somewhat. Research shows text predicts personality test scores at modest to moderate levels when there is a lot of text. A few messages give a much rougher guess.
Why did ChatGPT give me a different type than last time?
Because a chat is not a fixed test. The questions, the scoring and what the model focuses on can change between conversations. A standard test asks the same things and scores them the same way.
Do ChatGPT and Claude differ at guessing personality?
We don't know of a study comparing them on this, so we won't guess. The research we cite tested GPT models, and the limits apply to any chatbot: little text, no norms, no fixed scoring.
Is it safe to paste my messages into a chatbot to analyze my personality?
Check your chat settings first and leave out other people's private messages. Researchers who study this point to privacy as the main risk of cheap personality inference from text.
Can a chatbot or a test tell me which job to choose?
Neither can choose for you. A measured result describes how you tend to work, and a chatbot can help you think about what that means. See what job fits my personality?.
Sources
- Youyou, W., Kosinski, M., and Stillwell, D. (2015). Computer-based personality judgments are more accurate than those made by humans. Proceedings of the National Academy of Sciences, 112(4).
- Park, G., Schwartz, H. A., Eichstaedt, J. C., Kern, M. L., Kosinski, M., Stillwell, D. J., Ungar, L. H., and Seligman, M. E. P. (2015). Automatic personality assessment through social media language. Journal of Personality and Social Psychology, 108(6).
- Peters, H., and Matz, S. C. (2024). Large language models can infer psychological dispositions of social media users. PNAS Nexus, 3(6).
- International Personality Item Pool, ipip.ori.org, including the 50-item Big Five markers.
You walk a trail instead of filling a form
Tipolu is a short 3D walk. At each crossroads you pick the path you'd really take, and the companion walking with you at the end shows who you are.
12 to 20 crossroadsEveryday choices on a trail, like a market, a lake, a broken well.
Two campfire questionsShort open answers in your own words.
Your mix, not a boxA main type, an accent, and 0 to 100 scores on five traits, with a report that includes work.
Try three crossroads from the real walk
Pick the path you would really take. No right answers.
The trail splits: one path is signed and leads straight to town, the other appears on no map.
Your group is hosting a small dinner in town for friends you've made along the way.
Mist lifts off the lake at sunrise. You have an hour before the others wake up.
Three forks are a guess. The full walk uses 12 to 20 of them plus two campfire questions, and shows your main type, your accent and what that means for work.
Play the free test · 5 min
Which one is walking with you?
Find out how you tend to work before you pick your next path. A short walk, 12 to 20 crossroads, and a report that includes work.
Play the free testAbout 5 minutes · no signup · free