AI for researchers: delegate the reading, never the rigor

Ask a chatbot for the five most important papers on your research question and it answers in seconds, every reference formatted to the comma. When two information scientists actually checked, prompting ChatGPT for 84 short literature reviews and then hunting down all 636 works it cited, more than half of the older model’s citations turned out not to exist. The better model still invented 18% of its references.

That experiment is the whole bargain in miniature. AI is now genuinely good at the slow parts of research: finding papers, digesting them, polishing prose, drafting analysis code. It is also a fluent fabricator in a profession where a fabricated reference can follow your name around for years. This post is a working toolkit for the four pillars where AI earns a place in a research workflow, held together by two rules that never move: verify every citation it touches, and never paste unpublished work into a tool that trains on what you type.

Adoption is finished. Judgment is the new gap

The debate about whether researchers should use AI has quietly ended. In Wiley’s survey of 2,430 researchers across disciplines, 84% reported using AI, up from 57% a year earlier, and 62% now use it for research and publication tasks specifically. 85% say it has made them more efficient.

The same survey shows the honeymoon ending in a useful way. Concern about inaccuracies and hallucinations rose 13 points year over year, to 64%. Privacy worries rose to 58%. And the share of use cases where researchers believe AI outperforms humans fell from over half to under a third. That is not disillusionment; that is a community learning the tool’s actual shape. One more number worth pausing on: 80% of researchers reach for general chatbots like ChatGPT, while only a quarter use anything built for research. Most of the field is doing scholarly work with a consumer tool, under consumer data terms. That gap, between casual use and deliberate use, is what the rest of this post is about.

Literature search: the first hours AI gives back

Keyword search finds papers that share your vocabulary. The literature that matters often doesn’t: the same phenomenon gets named “intervention decay” in one field and “washout effect” in another, and classic keyword search will never introduce them. AI-era search tools work by meaning instead, embedding your question and every abstract in the same mathematical space and returning neighbors. Semantic Scholar, the free index run by the nonprofit Allen Institute for AI, covers over 200 million papers this way, and a growing shelf of assistants (Elicit, Consensus, scite and friends) sit on top of indexes like it to answer questions with linked citations rather than a results page.

The productive workflow treats these tools as scouts, not oracles. Ask a chatbot to map a field you’re entering: the main camps, the landmark results, the standing controversies. It is excellent at this, because survey-level knowledge saturates its training data. Then walk the actual citation graph: take the two or three real papers it surfaced, pull what they cite and what cites them, and read abstracts until the field’s shape is yours. What you must not do is let the map become a bibliography. Field summaries from a model are leads. A lead becomes a source the moment you have the PDF open, and not one moment earlier.

Reading and synthesis: only trust summaries of what it can see

There are two ways to ask AI about a paper, and they have opposite reliability. Ask from memory (“summarize Smith et al. 2021”) and you get whatever fragments of the paper leaked into training data, smoothed over with confident guesswork. Hand it the actual PDF and ask questions against the text, and you get one of the most useful research tools ever built: what was the sample size, which conditions were excluded, extract every effect size into a table, where do the authors admit the limitation that matters for my study. The difference between the two is grounding, the same mechanism behind retrieval-augmented generation: a model reading a document you supplied can point to its evidence, and a model reciting from memory cannot.

Scaled up, this becomes the real prize: a conversation with your own paper library. Fifty PDFs from your comps reading list or your lab’s reference folder, indexed and questioned together. “Which of these used longitudinal designs? Which contradict each other on the dose effect, and how do their methods differ?” Cross-paper synthesis that used to cost a weekend of spreadsheet archaeology now takes an afternoon of checking the model’s answers, and checking is still part of the job: models misread dense tables and mangle subscripts, so any number you plan to reuse gets confirmed against the page it came from. For reading outside your working language, the same grounded setup is quietly transformative, a near-fluent translation of any paper, on demand.

Every AI citation is fake until you’ve held the paper

Now the rule that protects your name. A language model is a text predictor, not a database, and when it needs a reference it does not look one up; it generates something reference-shaped: plausible authors, a credible year, a journal that fits the topic. Real and invented citations come out with identical confidence and identical formatting, which is exactly how hallucination works everywhere, just aimed at the one artifact academia treats as sacred.

Read that figure the pessimistic way. The improved model roughly tripled the honesty of its bibliographies, and it would still salt a 40-item reference list with about seven fakes, plus errors in a quarter of the genuine entries. Newer models keep shrinking the rate without reaching zero. So the discipline is binary, no judgment calls involved: every reference a model produces gets resolved to a DOI, and every DOI gets opened, before it enters your manuscript. Search the title in Semantic Scholar or Google Scholar; if nothing comes back, it never existed. Thirty seconds per reference, versus the alternative: a reviewer, or worse a reader after publication, discovering that citation 23 of your paper points at nothing. Science already retracts thousands of papers a year for integrity failures. The fix for this particular one costs half a minute.

Writing: journals allow an editor, not an author

The anxiety around AI writing has settled into surprisingly clear policy. Major publishers have converged on three lines. Springer Nature’s editorial policies state them plainly: AI cannot be an author, because authorship carries accountability and a model cannot be accountable; substantive AI use gets disclosed in the manuscript; and generative AI images are barred from publications. Peer reviewers get their own instruction, to which we’ll return: don’t upload manuscripts into AI tools.

Within those lines sits the most defensible use in this whole post. Grammar repair, tightening, restructuring a muddy paragraph, translation polish: for the majority of scientists writing in English as a second or third language, AI editing removes a career-long tax that never measured the quality of anyone’s science. Usage reflects that. A Stanford-led analysis of 950,965 papers estimated that by February 2024, up to 17.5% of computer-science abstracts were already LLM-modified, against about 6% in mathematics and the Nature portfolio. Those numbers are two and a half years old and predate most of today’s models; the direction has only been up.

The line to hold is the one integrity offices actually enforce: an editor improves sentences that carry your reasoning, a ghostwriter supplies reasoning you didn’t do. Prompting a model to “write the discussion section” and pasting the result crosses it, and not only ethically. The model will happily overinterpret your effect sizes and garnish the argument with citations you now know to distrust. Draft the thinking yourself, let the model make it readable, disclose what it did, and keep the boundary somewhere you could defend to your coauthors. Students navigating coursework face the same line under a different name.

Unpublished work is the one thing never to paste

Everything above concerned published literature, which is public by definition. Your unpublished work is different in kind: the dataset that took two years to collect, the manuscript under review, the grant aim nobody has thought of yet. In research, priority is the currency, and unpublished material is the only place it lives.

The strictest institutions have already drawn the conclusion. The NIH prohibits generative AI in grant peer review outright (notice NOT-OD-23-149, issued June 2023), reasoning that using these tools “requires the sharing of material from applications, violating NIH’s policy on maintaining confidentiality in peer review.” Springer Nature asks manuscript reviewers for the same restraint. Notice what both bodies are asserting: pasting text into a cloud AI tool is a disclosure event. They are right. Free consumer chatbot tiers commonly use conversations for model improvement unless you find the setting that says otherwise, and once text has entered a training pipeline there is no recall button. For a confidential manuscript you were trusted to review, or your own unsubmitted aims page, that is not a risk profile; it is a category error.

The practical policy is a two-tier rule, and it costs you almost nothing. Published papers, public code, your own already-public writing: any tool you like. Unpublished data, drafts, reviews, and anything a collaborator shared in confidence: only environments with a real confidentiality boundary, meaning an institutional deployment under a data agreement, or a model running locally on hardware you control, where the draft never crosses a network at all. Open-weight models on an ordinary workstation are now well past good enough for the summarizing, editing, and coding tasks in this post, which is why regulated industries are converging on the same architecture. A lab with unpublished data is a regulated industry that hasn’t written its regulation down yet.

Manage it like a first-year research assistant

The frame that makes every decision in this post easy: treat AI as a brilliant, tireless, occasionally dishonest first-year RA. You would happily hand that RA the literature scan, the summary table, the reference formatting, and the first pass at a translation. You would check every citation before it entered anything with your name on it, because you sign the paper and the RA doesn’t. And you would never give a first-year the only copy of your unpublished data to carry across town.

Start with one pillar this week, ideally the one eating your evenings. If it’s the reading backlog, put your top twenty PDFs into a grounded chat and interrogate them. If it’s the writing, draft one section yourself and let a model edit it, then diff the two and decide what to disclose. Keep the prompts that worked; they compound into the most reusable methods section you’ll ever write. The drudgery of research was never the research. The thinking stays yours, and now, more of the week does too.

Disclaimer: This article is general guidance, not policy or legal advice. Journal, funder, and institutional rules on AI use differ and change quickly; sources here reflect July 2026. Your target journal’s author guidelines, your funder’s notices, and your institution’s research-integrity office control. When in doubt, disclose and ask.