Hypothesis Generation & Study Design
Generating research questions is a core part of behavioural science, but it's not easy. Researchers often spend months narrowing down a testable idea, as this process relies on reading and assessing large volumes of prior work. With thousands of new studies published each year, it's becoming harder to keep up — and easier to miss important gaps in the literature.
AI tools can help by scanning and synthesising large bodies of research to suggest new hypotheses, highlight underexplored areas, and turn vague ideas into testable questions. Some tools can also recommend appropriate methods or generate draft materials such as surveys, interview guides, and behavioural interventions.
What AI can help with
- Analysing existing literature or large datasets and suggesting hypotheses.
- Designing research materials, such as surveys, stimuli, and interview guides.
- Refining vague or complex research questions into clear, precise and testable formats.
- Suggesting methods it would expect you to be using, based on your research question.
- Generating tailored intervention content — personalised health messages, therapy scripts, learning modules.
- Analysing past intervention data to refine strategies.
On this page
Choosing a tool
Most of the examples below use general-purpose large language models — ChatGPT, Claude, Copilot — adapted by researchers for a specific study task. Treat what they return as a draft: an evaluation of six AI tools for finding research gaps (Soriano et al., 2024) found that some returned accurate, useful references while others fabricated sources outright.
Examples
Refining research questions and suggesting hypotheses
A recent study (Tong et al., 2024) combined GPT-4 with causal graphs — maps of cause-and-effect relationships in research — and analysed over 43,000 psychology papers to generate 130 new hypotheses about well-being. The combined approach produced ideas as original as those from PhD students, and stronger than the language model alone.
Designing surveys, questionnaires and interview guides
A recent study (Yuan et al., 2026) explored how AI can support the design of psychological measurement tools, not just their analysis. The researchers fine-tuned a large language model on 169 professional psychological questionnaires so it could generate new scales from a research topic or construct. Against off-the-shelf models it produced items that were more logically consistent (a 28.6% improvement) and adapted more reliably across cultural and regional contexts (over 85% accuracy) — moving from research idea to validated-style instrument faster, while keeping scientific rigour.
Suggesting appropriate research methods
Generative AI can support better experimental design by helping identify mediators (why something works), moderators (for whom it works) and alternative treatment arms (Chang et al., 2024). Language models can help brainstorm these elements, simulate synthetic participants to test ideas, and assess whether a study is likely to scale — though researchers still have to judge which AI-generated ideas are worth keeping.
Generating tailored intervention content
GPT-3.5 has been used to draft 1,150 SMS messages for medication adherence in type 2 diabetes, meeting behaviour-change, readability and tone standards (Harrison et al., 2024). "MindShift" generates personalised persuasive messages that cut problematic smartphone use over five weeks (Wu et al., 2024), and Mass General Brigham found a significant share of AI-drafted replies to patient messages were safe to send unedited (Chen et al., 2024).
Analysing past intervention data to refine strategies
Agentic AI is being applied to personalised health coaching, with one agent identifying barriers to healthy behaviour through motivational interviewing and another supplying tailored strategies using models like COM-B (Yang et al., 2024). It reduces reliance on human coaches, but data privacy and unintended harms remain open problems — see LSE's blog on agentic AI in behavioural science.