Can ChatGPT do statistics on your data?
Partly. ChatGPT runs real Python on an uploaded file, so its arithmetic is usually right. The weak points are the choices around it: which test to use, whether its assumptions hold and what the result means. Those depend on how you ask and can change from one chat to the next. For a quick look that's fine; for a decision, a report or a thesis, check every step or use a tool built for statistics.
By Gevorg Petrosyan, founder of StatEVN · Updated 19 November 2026
What ChatGPT does well
When you upload a file, ChatGPT can write and run Python on it. That makes it good at:
- describing a dataset: counts, averages, missing values, quick charts
- explaining statistical ideas, such as what a p-value or a confidence interval means
- writing code for a test you have already chosen, which you can read and rerun
Because the numbers come from code that actually runs, simple calculations are usually right.
Where it goes wrong
The risk is not the arithmetic. It's the decisions around it, which a chat makes quietly:
- The wrong test. A t-test on a small, skewed sample, or a straight-line regression for a yes/no outcome. The code runs, and the answer looks fine.
- Skipped checks. Unless you ask, it often doesn't check the assumptions behind a model, such as uneven errors, overlapping inputs, or errors that carry over in time.
- Code that answers a different question. For example, adding up values across years when the question was about one year. This is hard to spot without reading the code.
- Confident interpretation. A correlation described as a cause, or "significant" read as "important".
- Different answers in different chats. Ask the same question twice and it may choose a different method, so the result can change.
- How you phrase it matters. Studies of ChatGPT on research data found descriptive statistics reliable, but its choice and use of inferential tests depended heavily on how detailed the prompt was.
If you use ChatGPT for statistics
Treat it like a fast assistant whose work you check:
- Ask which test it chose and why, and whether another would fit better.
- Ask it to check the assumptions of that test and show the results.
- Ask for the code, read it, and rerun it yourself.
- Never accept "significant" without the effect size, the p-value and the sample size.
- Check your account's data settings before uploading data that isn't yours to share.
ChatGPT and StatEVN side by side
| ChatGPT with a file | StatEVN | |
|---|---|---|
| Who picks the method | The chat, differently depending on how you ask. | Fixed rules based on your data and your question. Advanced mode lets you choose. |
| Assumption checks | When you ask for them. | On every result, with a plain warning when a result shouldn't be trusted. |
| Test on unseen data | When you ask for it. | Every model is scored on rows it never saw. |
| The explanation | Written by the language model. | Built from the fitted numbers by fixed rules. Nothing is guessed or invented. |
| Same file, same answer | Not guaranteed. | Yes. |
| Your file | Handled under your ChatGPT account's data settings. | Used only for your analysis. Never sold or used to train anything. |
| Best for | Exploring, learning, writing code you will check. | Answers you'll act on, report or put in a thesis. |
Why StatEVN doesn't let AI write the numbers
StatEVN is built the other way round from a chatbot. Established statistical methods do the analysis, with real p-values and diagnostics. Language is used only to explain the result, and each sentence is built from the fitted numbers, so it can't say something the model didn't find. When a result shouldn't be trusted, it says so before anything else.
Questions
Can ChatGPT calculate p-values?
Yes. With a file uploaded, it runs Python and can compute p-values correctly. The risk is whether it used the right test for your data, which you need to check.
Is ChatGPT good enough for thesis statistics?
It can help you learn and write code, but your supervisor will ask why you chose each test and whether its assumptions hold. Check every step, or use a tool that runs and reports those checks for you.
Does StatEVN use AI?
Not to produce results. The models are standard statistical methods, and the plain-language explanation is built from their numbers by fixed rules. Nothing is generated or guessed.
Why can ChatGPT give a different answer to the same question?
Each chat decides afresh which method to use and how to read the result, so the same data and question can lead to a different choice and a different conclusion.
Read next
Try it on a sample file
One click runs a real model on one of StatEVN's sample spreadsheets and explains it. No sign-up.