I use AI across the research lifecycle, from instrument design to the final report, without letting it near the parts of the work where a plausible-sounding error would be fatal. This page sets out exactly where it helps, where it does not, and the checks that make the difference.
Research has a lot of work in it that is necessary, repetitive and slow: checking an instrument for leading questions, chasing down what is already known, cleaning a messy export, coding two hundred open responses, writing the same exhibit note thirty times. AI is very good at that work. It is unreliable at the part that matters most, which is deciding what the evidence means. I use it hard on the first and keep it away from the second.
The draft instrument goes through AI review before it goes near a respondent: leading questions, double-barrelled items, response scales that do not balance, and drift between the Nepali and English versions where a translated item quietly asks something different.
It does not choose the research question, the sampling strategy or the unit of analysis. Those decide what the study can support, and a model that has not met the client cannot make them.
LLM assisted search and synthesis to map what is already known, cluster the arguments, and find the studies a keyword search misses because they use different vocabulary for the same idea.
No citation enters a document unless I have opened the source and confirmed it says what the summary claims. Fabricated and misattributed references are the single most common way AI damages a research reputation.
Automated checks run against incoming data while the team is still in the field: enumerator level outliers, straight-lined response patterns, interviews completed implausibly fast, GPS and timestamp anomalies. Problems surface on day three, not at analysis.
It does not replace supervision. A flag is a question for the field supervisor, not a verdict on an enumerator.
Scripted cleaning with every decision recorded, so the path from raw export to analysis file is reproducible. For qualitative data, the model proposes a first pass of codes across the transcripts, which I then adjudicate and revise.
The coding frame is mine and the final assignment is mine. Model coding is checked against a human coded subsample before it is trusted anywhere.
AI works as a second pair of eyes on Stata, R and Python: drafting analysis code, catching the error in a merge or a recode, and suggesting robustness checks worth running that were not in the original plan.
It does not choose which specification to report or what the coefficient means. Analysis is a set of judgements about a real context, and outsourcing those is how studies end up confidently wrong.
Drafting, restructuring and tightening; keeping exhibit titles, bases and notes consistent across a long report; producing the plain-language version for the audience that will never read the technical annex.
No sentence describes a finding that is not in the data. Every number in a final document traces back to the response file.
The saving shows up in preparation, cleaning and drafting, which is where research time actually goes. The judgement steps take exactly as long as they did before.
When robustness checks cost minutes instead of days, more of them get run. That tends to improve a study more than any single clever technique.
I set this up inside the team: the scripts, the prompts, the checks and the rules about what never goes into a model, so it survives after the engagement ends.
Whether you are running one study or a research unit, the same question applies: which parts of your workflow should AI touch, and which parts must stay human. I set that up with the team and leave the scripts, prompts and rules behind.