In my trainings the most common question is not whether researchers should use AI. It is how to get output they can actually trust. The answer is mostly about how you ask.
A large language model is like a brilliant, tireless research assistant with no context about your project and no fear of being wrong. Prompting is how you supply the context and install the fear.
Give the model a role and a standard
Start prompts by stating who the model should be and what quality bar applies. For example: you are a survey methodologist reviewing a questionnaire for a household study in rural Nepal, flag leading questions, double-barrelled items and translation risks.
This single habit removes most generic output. The model stops writing for an average reader and starts writing for your discipline.
Work in stages, not one giant ask
Do not ask for a finished literature review. Ask first for themes, then for competing findings within each theme, then for the strongest citations to verify. Each stage gives you a checkpoint where your judgement enters the process.
The same applies to analysis. I ask the model to propose an analysis plan, critique it, and only then draft code. Errors surface early, where they are cheap.
Verify like a supervisor
Treat every factual claim, citation and number as unverified until checked. Ask the model to rate its own confidence and to list what it could not verify. Spot check against sources you trust.
Used this way, AI does not replace rigour. It frees time for the parts of research that need a human: framing the question, understanding the context and standing behind the findings.