Using AI for data analysis

By Dr. Rafiq Muhammad, MD, PhD · Updated June 2026

AI can genuinely accelerate analysis — writing your code, drafting a first set of qualitative codes, explaining what an output means. What it can’t do is be responsible for the result. The safe pattern is the same everywhere: AI drafts, you run and verify, you own the finding.

Quantitative: code, don’t compute

The single most important habit: have AI write the code, not do the maths. A model will happily compute a t-test in chat and quietly get it wrong; the same model writes correct R or Python that you run on your real data in real software. Ask it which test fits, to write the script, and to explain the output — then sanity-check the numbers yourself.

Qualitative: a first pass, not the analysis

AI can suggest codes, group similar segments, and draft a codebook — a real time-saver on a large dataset. But the analyst reviews and owns every code. Trustworthiness rests on a transparent, reflexive process, so document where AI assisted and never let its labels substitute for your interpretation. A theme the model named that you can’t defend from the data isn’t your finding.

The three rules that keep it rigorous

1. Confidentiality — never paste identifiable or unpublished participant data into a public model. Anonymise first, or use an approved private environment.
2. Verify — every number against the real output of your software; every AI claim against the data.
3. Disclose — record where AI assisted, as your journal or institution requires.

Where it shines

Debugging a cryptic error, translating SPSS clicks into reproducible code, explaining what an interaction term means, drafting a data-cleaning script — these are real, safe wins that save hours without touching your judgement. The line holds: AI helps you do the analysis; it doesn’t make the call.