AI research ethics checklist
Using AI in research — to generate data, assist analysis, draft text, or simulate participants — raises ethics questions that traditional protocols did not anticipate. Plausible-but-fabricated data, undisclosed AI authorship, participant data leaking into third-party tools, and irreproducible AI-influenced results all threaten research integrity. This checklist walks you through an IRB-style self-assessment tailored to your research type, how you are using AI, and whether human participants are involved, so you catch the issues before submission or publication.
The specific risks AI introduces that traditional protocols don’t address
Traditional IRB checklists assume research generates data through observation, survey, or experiment — and that the data is what it appears to be. AI breaks several of those assumptions:
Fabrication risk. AI language models produce plausible-sounding text and numbers. A researcher who asks an LLM to “generate ten example participant responses” and then includes those in an analysis has committed data fabrication — even if unintentionally. The risk is highest when AI is used to fill gaps, generate synthetic participants, or extrapolate from small samples.
Participant data and third-party tools. If your research involves human participants and you feed their responses into a commercial AI tool for analysis, several things may have gone wrong: the original consent form likely didn’t cover this use, the AI provider may retain the data for training, and data protection obligations may be triggered. This is one of the least-discussed and most common ethics issues in qualitative research using AI analysis tools.
Attribution and authorship. Major journals now require explicit disclosure of AI use, but the norms are evolving quickly. Vague disclosure (“AI tools were used”) does not meet the standards most publishers are moving toward. Specific disclosure includes tool name, version or date accessed, the specific role in the research (e.g., literature summary, coding assistance, writing), and what verification steps were taken.
Reproducibility. Traditional methods can be described precisely enough to replicate. An AI-assisted method often cannot be — model outputs change with version updates, temperature settings, and even prompt order. Preserving reproducibility requires recording exact prompts, model names and versions, API parameters, and the date the model was accessed.
How it works
You describe the project: the research type (empirical, qualitative, computational, or literature-based), how AI is used (data generation, analysis assistance, writing support, or participant simulation), and whether human participants are involved. The tool filters a master list of ethics items into the relevant categories — informed consent and participant data, research integrity and fabrication risk, attribution and transparency, and reproducibility — flags the critical ones, and tracks your completion. You can export the finished checklist into your IRB application, methods section, or supervisor review.
Tips and notes
- Separate AI-generated from collected data. Label it, store it apart, and never let synthetic content masquerade as evidence.
- Disclose specifically. Name the tool, its version, and the exact role it played — not a vague “AI was used” line.
- Mind participant consent. Sending participant data to a third-party model can exceed the original consent and retention assumptions.
- Preserve reproducibility. Record prompts, model versions, and parameters; AI outputs drift, so an undocumented method is not reproducible.
- Check your institution’s current policy. Many universities have issued AI-in-research guidance since 2023; your IRB may have specific requirements not captured in general checklists.