AI for Researchers: The Honest Guide
No hype, no panic. What AI can actually do in everyday research (literature work, data analysis, writing) and where to stop delegating before it has consequences.
I use AI every day. For literature searches, coding assistance, first drafts, administrative texts. Along the way I have made mistakes that I would like to spare others.
This text is not an attempt at euphoria and not a warning cry. It is a report from practice: what I do, what I no longer do, and why.
What researchers actually think
A much-cited study by Van Noorden and Perkel in Nature (2023) asked researchers for their assessment of AI. On the plus side: faster data processing, automated acquisition, new hypotheses, optimized experimental designs. On the minus side: automation bias, easier fraud, proliferation of misinformation.
What has changed since then, I see mostly in my own environment: AI use for coding and literature search has noticeably increased. At the same time, there are colleagues who deliberately use AI not at all. The middle ground is narrowing. Those who use it use it more intensively. Those who don’t, do so deliberately.
Both have their justification.
Literature: what works, what doesn’t
What works well: translating and improving search strings. Summarizing openly accessible papers. Extracting study limitations and design. Creating screening checklists for scoping reviews. Identifying contradictory evidence.
For fields I’m newly entering, I use the following GlossaryPromptThe input or instruction given to an AI model.Go to entry ↗ scheme: I give the model a role (as an experienced academic researcher), a task (write an evidence-based overview of X), and a format. The model then sometimes runs for 30 minutes and produces a report that considerably shortens my ramp-up into a field. Not a publication, not a substitute for primary sources, but a useful entry point.
What doesn’t work: having it find references for a particular thesis. Having it write introductions with citations. Letting the model decide which papers are relevant.
That last one is an ethical line: the decision about relevance is scientific work. It shouldn’t be delegated.
And the first, searching for references with AI, is a well-known disaster. A test of eight AI search tools across two hundred citation queries found: Perplexity answered 37 percent incorrectly, and Google Gemini delivered only a single fully correct answer. An analysis of 2.5 million biomedical journal articles found for the first weeks of 2026: roughly one in every 277 papers contains at least one fabricated reference, twelve times more than in 2023.
I work in peer review for a medical journal. What I see there: submissions with references that don’t exist. Papers that contain, in the discussion section, the sentence “I'm an AI language model and don't have access to real-time information or patient-specific data”, verbatim, slipped unnoticed through review and proofreading rounds. Not out of fraudulent intent. Out of negligence.
Consequence: every AI-generated reference has to be checked manually. Not by sampling. Every one. Open the DOI, compare the metadata, read the abstract, check whether the paper actually supports the claim.
Analyzing data: possibilities and a clear limit
Modern models can generate descriptive statistics, cleaning routines, and plausible visualizations from a CSV file in minutes. That is genuinely useful.
But: patient data does not belong in ChatGPT. Not even pseudonymized data. That is a data protection requirement and is clearly formulated in most institutional rulebooks.
The workaround I recommend: synthetic datasets. I take the column structure of my real dataset, generate an artificial dataset with AI help that has the same structure and similar problems (missing values, outliers, coded null values), and develop my analysis code against it. Then I cut the AI connection, load the real dataset, and run the finished script.
That way the model never touches the data. The script does the work.
For the coding environment: I use Visual Studio Code with GitHub GlossaryCopilotan AI assistant that works alongside a person on the job.Go to entry ↗. As an academic or student, there is a free GitHub Student license. The Copilot system is well integrated into VS Code, allows finer control than a pure chat interface, and makes the code visible: I can check it, understand it, adapt it.
That is the decisive difference from vibe coding: I see what the model is doing. I can judge it. Anyone who produces code they can’t read has a problem. Not today, when everything runs, but when it doesn’t.
Writing: acceleration, not ghostwriting
AI can speed things up considerably: grammar, style, abstracts, structured summaries, cover letters, project reports, reviewer responses. I use all of it.
The difference between good and bad use is control: over content and argumentation.
A bad prompt: “Write me an introduction with citations.” A good prompt: “Write a structured 200-word abstract for the following manuscript, for an audience of clinical researchers. Preserve all numbers and units. Don’t invent anything. Flag missing information instead of bridging it.”
The length of the prompt is not an end in itself. But structured prompts reliably deliver better results than open-ended requests: with role, task, context, constraints, and a verification instruction.
Two common mistakes in AI-written texts: exaggeration (the model leans toward superlatives; a groundbreaking finding for an n of 5) and generic formulaicness (everything sounds like the same academic mediocrity). Both can be reduced through prompting, but not eliminated. Your own eye remains necessary.
And: disclosure is mandatory. Anyone who uses AI in manuscript preparation states so, in the acknowledgements section or in the ethics statement. An example wording: "The authors used [Tool] to support [task] during the preparation of this work." That is nothing to be ashamed of, it is standard. The problem arises when AI is used and that is concealed.
What AI can’t do, and why that is decisive
For the first year after ChatGPT was introduced, I delegated almost everything. Literature, structure, phrasing, code, even difficult emails. It was fast. It worked. Until I noticed that I wanted to write a section myself and it felt harder than it used to.
That is not a dramatic decline. It is atrophy: skills that aren’t used get weaker. In experts too, not just in beginners.
In my daily work with doctoral students, I see the same pattern. The texts arrive faster. The structure is right. But when I ask why they made an interpretation the way they did, sometimes no answer comes. Because they didn’t interpret it; the model did.
That is the real problem: AI automates first the entry-level tasks that train the craft. Anyone who has never learned what a clean analysis plan looks like can’t judge whether the plan ChatGPT generated is good. The judgment presupposes the competence that training produces.
There are also phenomena that remind us we don’t fully understand these systems. GPT-5.1 mentioned goblins and gnomes unusually often in unrelated conversations. Nobody knows why. A recent Claude version told users in coding sessions to go to sleep, as a gesture of its own that wasn’t foreseen in training. These are curiosities. They show that we can observe these systems only at input and output, not inside them.
The question that counts
I close with the guiding question I ask myself with every use of AI:
Can I verify, document, disclose, and defend what the AI did?
Not: What can AI do for me? But: Can I stand behind it?
If yes: an excellent tool. If no: my problem, not the AI’s.
A practical overview:
Green light (with verification): optimizing search strings · summarizing open-access papers · writing code for your own (non-patient-related) data · correcting grammar and style · first versions of abstracts and reports · administrative texts · ideation and hypothesis generation
Yellow light (with particular care): literature overviews for your own research · data analysis with synthetic datasets · writing manuscript sections
Red light: uploading patient data · adopting references without manual checking · letting AI decide which sources are relevant · submitting results as a publication without substantive review by the author group
Oscar is a postdoctoral researcher at the Else Kröner Fresenius Center for Digital Health (TU Dresden) and a co-founder of Denkmal.