NotebookLM for Literature Reviews: Where It Helps and Where It Still Needs Human Judgment
A NotebookLM literature review works only if you are clear about which half of the job it does. The tool reads the sources you load and answers from them, with pointers back to the passage it used. It does not search for papers, make inclusion decisions, or check that the references you have are correctly attributed. Treated as a source-grounded notebook, it is useful. Treated as a systematic-review engine, it will quietly do less than you think.
What it does well
When the corpus is already curated, NotebookLM is strong at summarising across documents, answering cross-paper questions, and pulling recurring themes out of a set. Google's documentation for Gemini Notebook, formerly NotebookLM, describes it as returning "grounded information based on your sources with clear in-line citations," and the FAQ is explicit that "if the answer isn't in the source material, it won't provide a response." For a reading-heavy synthesis step, a tool that refuses to improvise past your sources is doing something valuable.
Where human judgment stays in the loop
It is not a search
Every screened set starts with a search across named databases. PRISMA 2020 asks reviewers to "specify all databases, registers, websites... searched" and to "present the full search strategies for all databases" (PRISMA 2020 checklist). NotebookLM searches none of them. Whatever you did not upload does not exist as far as the tool is concerned.
It does not screen
Inclusion and exclusion decisions carry methodological weight and need a documented rule applied by a person, ideally two. Asking a grounded model to "pick the relevant ones" produces a plausible list with no auditable basis and no record of what it dropped.
It does not verify citations
Because it can only cite what you gave it, it will not invent a paper. It can still misread one — summarising a pilot as a definitive trial, or attaching a finding to the wrong subgroup. That is the same interpretation risk covered in the honest verdict on NotebookLM synthesis, and it survives grounding intact.
The retrieval limit inside a big notebook
With many sources loaded, the FAQ notes the tool "retrieves the most relevant information based on your question, then builds a response from it" rather than reading everything each time. A vague prompt against a crowded notebook returns a shallow answer that can look like an evidence gap when it is really a retrieval miss. Narrow the question to the population, intervention, or outcome before you trust a thin response.
The caveat Google prints itself
Google's documentation states the tool "can make mistakes and its answers don't reflect Google's views" and advises consulting "a qualified professional for medical, legal, or financial advice." In a clinical review that is not filler to skim past: a grounded summary of a trial's harms or effect size still needs a reader who knows the field to confirm it before it shapes a conclusion.
Where it fits in the review
After the search is done, the screening is logged, and the surviving references are confirmed real, NotebookLM is a reasonable place to synthesise the final set: draft the narrative summary, run cross-document comparisons, check your own claims against your own sources. That ordering — grounded synthesis last, not first — is the point of a source-grounded research workflow, and it is compatible with the gated approach in AI in systematic reviews without compromising rigor.
AI for Academic's literature search and full-text fetch at aiforacademic.world assemble and export the screened set, and its citation check confirms the references are real before you load them anywhere. Free to start.