The question arrives every week: "Can I use ChatGPT to analyse my thesis data?" The honest answer is: it depends on what for, and the difference between good uses and bad ones can cost you your defence. This guide separates what AI does well from what it fabricates, and tells you at what point a real statistician stops being optional.
First, the context: what kind of AI we are talking about
ChatGPT, Claude, Gemini and similar tools are language models. They are trained to predict plausible text, not to execute correct statistics. That means they can give you answers that look like statistics but are wrong, and the plausibility of the text makes it hard to tell correct from invented unless you already have the knowledge to check. They do not have access to your data (unless you paste it in directly), they do not run statistical software, and they do not reason about the internal logic of your research design.
There are exceptions: some systems integrate code tools (Python, R) that do run real analyses. But even then, interpreting the results in the context of your specific question and design requires judgement that the model lacks.
What AI genuinely does well
There are real and legitimate uses:
- Understanding output you already have: paste your SPSS or R results and ask for an explanation. AI is useful here, but always cross-check against the test documentation. The values it interprets are the ones you provide; if the results themselves are wrong (because you chose the wrong test), the interpretation will be wrong too.
- Drafting the results section from correct values: if you already know the values and the interpretation criteria, AI can help you shape the text. It is a writing assistant, not an analyst.
- Finding and understanding literature: exploring concepts, asking for an explanation of measurement invariance or the difference between ICC(2,1) and ICC(3,1). Useful for understanding, not for deciding whether your analysis requires it.
- R or Python code for analyses you already know how to do: if you know which test to run and what options it takes, AI can write the code. The risk is trusting code that looks correct but uses outdated packages or misspecifies parameters.
- Reviewing the phrasing of hypotheses and objectives: improving clarity, academic register, spelling. The risk is low here.
What AI cannot do (and where the risk is high)
- Choose the correct statistical test for your design: without seeing your data, without knowing the type of each variable, without understanding the repeated-measures design or nesting, any recommendation is an informed guess, not a technical judgement. The article on how to choose the statistical test covers the real criteria.
- Verify assumptions: normality, homoscedasticity, independence of residuals. That requires running the tests on your actual data and making decisions about what to do when an assumption is violated.
- Calculate and interpret effect sizes: AI can tell you that a Cohen's d of 0.5 is "medium", but it cannot calculate it correctly from your results table and cannot tell you whether that magnitude makes sense in the clinical context of your variable. The guide to effect size interpretation and the effect size calculator give you the foundation.
- Justify sample size: your committee or reviewer will ask for the four ingredients of the power calculation (expected effect, alpha, power, specific test) and the source of the effect. AI will invent numbers if it does not have them; the article on sample size calculation explains how to do it properly.
- Evaluate design validity: whether your cross-sectional design cannot support causal inference, whether the sample has a selection bias, whether the control variables you chose are appropriate. That is expert judgement.
The most dangerous mistake: AI invents plausible results
Language models hallucinate more frequently on numerical tasks than on textual ones. If you paste your data into the chat and ask for ANOVA results, the model can return plausible-looking values that do not correspond to any real calculation. Thesis committees and reviewers do catch this: results with no internal logic, discrepancies between N and degrees of freedom, effect sizes inconsistent with p-values.
The acid test is always: can you replicate the result in SPSS, R, or any statistical software with your actual data? If you cannot, the number does not exist.
The same goes for the bibliography, and there the check is even faster. If you asked a language model for references, run the list through the reference checker: it looks up every entry in Crossref, OpenAlex, PubMed, Semantic Scholar and DOAJ and returns the real DOI when there is one. A fabricated citation in a thesis reference list is one of the things a committee does check, and does not forgive.
AI as accelerator, not substitute
The smartest use of AI in a thesis is as an accelerator for steps you already master: if you have already decided you will run a logistic regression and have already checked the assumptions, AI can help you write up the results. If you already have the CFA output and know that a CFI of .94 and RMSEA of .057 are acceptable, AI can help you structure that paragraph in APA. But the decision that those indices are acceptable in your context cannot be delegated.
The problematic pattern is the reverse: using AI at the start, when you do not yet know which test to run or which assumptions to check, and trusting the model to make those decisions for you. That is where the errors that committees catch are generated.
When it makes sense to hire a statistician
The clearest signal is when you have been consulting AI for a while and the answers it gives you are ones you cannot verify or fully understand. Another indicator: when the analysis you need (SEM, multilevel models, measurement invariance, survival analysis) has technical conditions that go beyond running the default test. A statistician does not just execute the analysis; they explain what conditions it requires, what it means if those conditions are not met, and how to report it in a way that survives peer review. The article on how to choose a statistician for your thesis covers the criteria for distinguishing a good advisor from an analysis factory.
Is AI giving you answers you cannot defend before your committee?
I am a PhD in Psychology and I carry out the complete statistical analysis of your thesis (test selection, assumption checking, execution in SPSS or R, interpretation and APA write-up) with results ready to defend. Free initial diagnosis.
Book free diagnosis →Frequently asked questions
Can ChatGPT choose the correct statistical test for my thesis?
Not reliably. ChatGPT can guide you on what factors to consider, but choosing the right test requires knowing the type of dependent variable, the study design, and verifying assumptions on your actual data. The model can give wrong recommendations if it lacks that complete information.
Is it dangerous to use ChatGPT to interpret SPSS output?
It can be, if you lack the knowledge to verify the interpretation. AI interprets correctly when you provide exact values and test context. The risk is that the model confuses parameters, uses outdated conventions, or fails to detect that results are produced by an assumption violation.
Can ChatGPT generate the statistical results for my thesis?
No. Language models do not run real statistical calculations. If you ask for ANOVA results from your data, the model can return plausible-looking numbers that correspond to no real calculation. Thesis committees detect these inconsistencies.
Which parts of the thesis can I use AI for safely?
The lowest-risk uses are: improving the phrasing of hypotheses you have already defined, structuring the results section from values you have already calculated and verified, searching and understanding literature, and generating R or Python code for analyses you already know are correct.