A deep research report is not a literature review, and the resemblance is the problem
Deep research agents return something that looks like a literature review: structured, cited, finished. But a review earns trust from its method, not its prose, and the reproducible search that constitutes the method is exactly what these tools omit. Their retrieval is opaque and non-reproducible, and evaluations keep finding citations to be their weakest point. The tools are genuinely useful for exploration. The danger is that their output wears the costume of confirmation, inviting researchers to treat leads as findings. The fix is not to ban them but to refuse to let the form of an output stand in for its method.
Paste a research question into one of the new "deep research" agents and wait. In twenty minutes you get back something that looks like the thing you would have spent a fortnight producing: a structured report, section headings, a confident synthesis, and a list of citations at the bottom. It reads as finished. That is exactly the problem.
The resemblance between a deep research report and a literature review is close enough to be mistaken for the real thing, and the mistake matters, because the two are not the same kind of object at all.
What actually makes a review trustworthy
A literature review earns trust from its method, not its prose. The readable summary at the front is the least important part. What a knowledgeable reader checks is the apparatus: which databases were searched, on what date, with which exact search string, under what inclusion and exclusion criteria, and how the results were screened down to the set that made it in. The field has spent years formalising precisely this. PRISMA-S, the 2021 search-reporting extension to the PRISMA statement, is a sixteen-item checklist whose entire purpose is to make a search complete enough to report and therefore reproducible. The point of all that bookkeeping is not bureaucratic. A search you can reproduce is evidence. A search you cannot reproduce is an assertion.
This is the property that separates a systematic review from an opinion with references. Someone else can run your search, get your numbers, and see whether your conclusion survives. The prose is downstream of that. If the method is sound, the writing is reporting. If the method is hidden, the writing is only persuasion.
Humans were already bad at this
It would be too easy to make this an argument about machines. Before any agent existed, human-authored reviews were failing the reproducibility test routinely. When Koffel and Rethlefsen examined search reporting in systematic reviews from high-impact pediatrics, cardiology and surgery journals, a large share did not contain a reproducible search strategy at all. A 2023 cross-sectional metaresearch study reached the same verdict in blunter terms: search strategies are poorly reported and not reproducible. PRISMA-S exists because the problem was already entrenched.
So why single out the agents? Because they move the incentive in the worst possible direction. A human who cuts the methodological corner at least knows the corner is there. The agent removes the corner from view entirely, and hands back an artifact more polished than most humans would have produced, with the one auditable component quietly missing or unusable.
Where the output is most convincing, it is least checkable
Look at what these systems actually do. OpenAI's Deep Research, launched in February 2025, browses the live web for several minutes to half an hour and returns a cited report. The retrieval path is opaque: you see a summary of steps, not a search you could re-run. Ask the same question tomorrow and the web has moved, the model has sampled differently, and the set of sources may not be the same. There is no fixed corpus, no recorded query, nothing to reproduce.
The citations, the part that makes the report look most like scholarship, are where evaluations keep finding the weakest performance. A 2026 viewpoint in the Journal of Medical Internet Research on deep research agents for medical use reports that citation accuracy varies widely across systems and that fabricated or misattributed references remain common. A growing set of benchmark studies, several still preprints and to be read as such, converge on the same finding: citation quality and factual grounding are the axes on which these agents perform worst, not best. The report is at its most authoritative in appearance exactly where it is least reliable in substance.
The category error
None of this makes the tools useless. Stated at full strength, the case for them is good. For scoping a new area, for discovery, for surfacing a paper in an adjacent field you would never have thought to search, a deep research agent is genuinely fast and often finds things a manual search misses. Exploratory work rewards breadth and speed over reproducibility, and here the agents earn their place. A researcher who refuses to use one for a first pass is being precious, not rigorous.
The hazard is not the tool. It is the confusion of two kinds of work. Exploratory search is confirmatory search's opposite in almost every way that matters: one is allowed to be messy, partial and unrepeatable because its output is a set of leads, not a claim. The other has to be repeatable because its output is a claim. Deep research reports fail as confirmatory instruments not because they explore, but because they dress the exploration in the costume of confirmation. The section headings, the measured tone, the citation list: all borrowed from the confirmatory form, none of the underlying discipline that gave that form its authority.
That is what makes the resemblance dangerous rather than merely imperfect. A tool that looked provisional would be used provisionally. A tool that looks finished invites you to treat leads as findings.
The question to ask any research tool
The corrective is not to ban the agent, nor to romanticise the manual search that was already unreproducible half the time. It is to refuse to let the form of an output stand in for its method. A search you cannot show and cannot re-run is a claim about the literature, not a survey of it, however well the prose around it reads.
So the useful question to put to any research tool, human-driven or agentic, is small and awkward and worth insisting on: show me the search, and tell me whether it will run the same way tomorrow. If it can, you have evidence you can build on. If it cannot, you have a starting point, which is a fine and valuable thing to have, as long as nobody mistakes it for the finish.
This is the reasoning-augmentation case in one example. A system that helps you think better makes its own steps visible so you can check them. A system that thinks instead of you hands you a conclusion and hides the working. The report that comes back looking most complete is, for now, the one to trust least, right up until it shows you how it got there.