Dr David Fraile Navarro, Postdoctoral Research Fellow, Centre for Health Informatics, Macquarie University
A Personal Start: My Reference “Aha” Moment
Last month, while writing a grant, I found myself stuck. Three half-finished paragraphs sat waiting for references. Normally, this would mean an hour trawling PubMed, copy-pasting titles, and staring at a blank page trying to integrate them.
Instead, I tried something different. I pasted one citation I had already collected into ChatGPT and asked it to draft a paragraph placing the study in context. The result wasn’t perfect, I still had to edit and shape it, but it gave me a solid first draft and broke my writer’s block.
That was my “aha” moment: GPT wasn’t replacing my work, but it was shifting it. Less clerical time, more critical editing.
Why References?
References are the scaffolding of scientific writing. Yet anyone who has written a manuscript, prepared a grant, or reviewed a paper knows how time-consuming they are to manage. Searching, formatting, integrating, and checking citations all take hours.
When ChatGPT first appeared in late 2022, it seemed like a potential shortcut. The reality was less helpful: the model confidently produced lists of references, but many were fabricated. Titles and journals looked convincing, but the articles didn’t exist. In healthcare research, that’s not just an error, it undermines trust. Researchers and lawyers were caught in cases with fabricated references being used without proper checking in the real world.
Fast forward to 2025. The situation is changing. GPT can now connect to real databases, search the web, retrieve verifiable citations, and even support more advanced workflows. In this piece, I describe three ways I’ve begun to use my GPT in my own academic work: retrieving references, integrating them into writing, and verifying them during peer review. Along the way, I’ll also touch on newer capabilities like Deep Research and a recent protocol exploring potential future GPT’s role in risk-of-bias assessments.
Use Case 1: Retrieving References Without Hallucination
In the past, GPT “invented” citations. Today, when I ask:
“Find recent publications on automating risk-of-bias assessments in systematic reviews, and give me formatted references in Vancouver style,”
I get a list of real articles, complete with DOIs that I can cross-check on PubMed. Formatting is not always perfect, and coverage depends on access, but GPT no longer hallucinates entire references.
Why it matters:
- Saves time on clerical work.
- Provides a credible starting point for a literature search.
- Depending on the subject some references may retrieve grey literature and less reputable sources.
- Still requires human checking for accuracy and completeness.
Use Case 2: Integrating References Into Writing
Finding a reference is only half the job. The harder part is weaving it into the text. This is where GPT has been most useful in my day-to-day research.
Here’s a real example from my desk:
- Prompt: “Here is a reference: Østengaard et al. 2025. Draft a paragraph explaining why automation of risk-of-bias assessment matters in evidence synthesis, situating this study in context.”
- GPT Output (excerpt): “This study demonstrates that large language models can perform structured risk-of-bias assessments with moderate reliability, highlighting the potential to streamline evidence synthesis workflows.”
- Edited Final: “In evidence synthesis, automating risk-of-bias assessments has long been seen as a bottleneck. Østengaard et al. (2025) show that large language models can achieve moderate agreement with human reviewers, a promising step toward scalable automation, though far from replacing expert judgment.”
GPT gives me a scaffold; I apply the judgment and nuance.
Why it matters:
- Reduces writer’s block.
- Provides multiple stylistic options.
- Keeps control in the researcher’s hands: I choose the source, GPT helps with the framing.
📌 Prompt to Try
“Here is a reference [paste full citation]. Draft a short paragraph situating this study in context for a paper on [your topic].”
Use Case 3: Verifying Citations in Peer Review
As a peer reviewer, I often need to check whether references truly support an author’s claim. A sentence may read: “Automated RoB tools now outperform human reviewers (Smith et al. 2024).”
Instead of pulling the paper manually, I can paste the claim and the citation into GPT:
“Summarise what this paper reports. Does it support the claim?”
GPT then highlights whether the study actually makes that claim. In one case, it flagged that the cited paper only showed moderate agreement with human reviewers, not superiority.
Why it matters:
- Acts as a rapid screening tool.
- Flags potential mis-citations.
- Does not replace full reading but reduces workload in reference-heavy reviews.
Looking Ahead: Deep Research and Risk-of-Bias
Two developments point to where this may be going.
Deep Research, a new ChatGPT feature, can browse and synthesise across dozens of sources. Asking it for a mini-review on risk-of-bias automation yields a structured report, background, methods, challenges, with references and a trail of consulted sources. It’s not flawless, but it edges closer to a junior assistant that can prepare the first steps of a scoping review.
Meanwhile, a new study in BMC Medical Research Methodology tested GPT on risk-of-bias assessments. Researchers sampled trials from Cochrane reviews, asked GPT to rate risk, and compared results with human reviewers measuring agreement between raters, beyond chance.
Risk-of-bias assessment, judging whether a study’s design or reporting could distort results, is complex and nuanced, so this is a strong test of GPT’s “reasoning” limits. Even partial success could help reduce one of the most labour-intensive steps in systematic reviews and help researchers and busy clinicians prepare and deliver high quality reviews.
Responsible Use Today
GPT is not a replacement for researchers. Its role is more like an assistant that needs constant supervision and steering, shifting the balance of our work. To use it responsibly:
- Always verify outputs. Check every reference against the source.
- Don’t outsource judgment. Use GPT to draft or flag issues, not to make final calls.
- Know its blind spots. Paywalled articles, nuanced interpretation, and domain expertise remain weak points. When reading PDFs, formatting issues and tabular information may not be processed adequately.
“If GPT once symbolised unreliability through fake citations, it now symbolises a new challenge: how do we responsibly integrate a tool that is finally beginning to get the references right?”
Lessons for Health Researchers
From these examples, a few principles emerge:
- Time savings are real. GPT reduces clerical burdens.
- Oversight is essential. Verification is non-negotiable.
- Workflows are shifting. Less formatting, more editing and interpretation.
- Transparency is improving. Features like Deep Research’s source trail help accountability.
- Limits remain. Use GPT for scaffolding, not final judgment.
Looking Further Ahead
Beyond references, the next frontier may be workflow integration. Imagine:
- Reference managers embedding GPT to suggest relevant papers.
- Peer review platforms automatically checking whether citations align with claims.
- Hybrid research roles where humans supervise AI assistants for preliminary evidence synthesis.
These shifts could reshape research itself. But the challenge will be ensuring transparency, rigour, and equity as AI integrates more deeply.
Your Turn
How are you using GPT in your own research? Have you tried it for references, or for other parts of the workflow? Share your experiences with us, future Me and my GPT instalments will feature stories from across our community.
Conclusion: From Fabricator to Colleague
In just a few years, GPT has gone from confidently fabricating references to competently retrieving, integrating, and checking them. With Deep Research and tools under academic study, it is edging into automated synthesis.
The transformation is not about replacing researchers. It is about changing where we spend our energy: away from mechanical tasks, toward oversight and interpretation.
And that shift, “me and my GPT” at the desk together, is the story this series will continue to explore.