Most researchers ask the wrong question about AI. “Which AI tool is best?” matters far less than “where in my workflow is time being lost on low-value tasks?”
What AI is good for
The honest case for AI in research is narrower than the hype suggests. LLMs are strong at summarising, drafting, restructuring and improving clarity.
Good use cases might be:
- Sharpening a vague research question.
- Generating MeSH terms and search synonyms ahead of a database search.
- Turning dense methodology into plain-language participant materials.
- Reshaping a manuscript discussion section into a conference abstract.
Used this way, AI frees researchers to spend time interpreting findings instead of rewriting the same ethics summary for the fourteenth time.
Where it starts to go wrong
Where AI is poor, and this is the part worth taking seriously, is truth verification, scientific judgement, ethics and clinical reasoning.
Citation fabrication is the clearest example. One recent evaluation of citation claims generated by a leading LLM found only 43.5% were fully accurate.
Nearly 20% were entirely fabricated, complete with plausible journals and DOIs that don’t exist. AI can support a search strategy. It should never generate references for direct citation.
Bias hides in the wording
Bias creeps in more quietly. A prompt asking about “problems rural clinicians face because of limited resources” already assumes deficit before any data has been collected, and it’s a framing LLMs will happily reproduce unless a researcher steps in to correct it. Equity-focused health research is particularly exposed here.
The data test
Then there’s data.
Interview transcripts, patient records and identifiable clinical notes should never go into an unapproved public AI tool. There is a simple test: if you wouldn’t email the raw file to a person unauthorised by ethics approval, don’t upload it.
For teams that genuinely need AI support with sensitive data, the answer isn’t to avoid AI altogether. It’s to use the right environment: enterprise or institution-hosted platforms with proper data governance.
Two frameworks worth keeping close
Two frameworks are worth keeping close.
SCAR (Specify, Context, Audience, Rules) sharpens prompt quality.
SAFE (Secure, Accurate, Fair, Ethical) works as a pre-use checklist:
- Is the platform approved?
- Is the output verified?
- Has bias been checked?
- Does this meet disclosure and authorship standards?
The bottom line
AI can assist, but it cannot take responsibility. The responsibility stays with the researcher, every time. Used well, it doesn’t just make research faster, it clears space for the parts of research that still need a human mind: judgement, interpretation, and accountability.
Stories worth sharing
The text of this article is licensed under the Creative Commons Attribution (CC BY) 4.0 Internationallicense. We’d love for you to share it, so feel free. Images, videos, graphics and logos are not covered by the CC BY license and may not be used without permission from Deakin or the respective copyright holder. For more information on how to share or reuse this content, please contact researchcomms@deakin.edu.au.
Share