How accurate is NotebookLM? Citations, hallucinations and a checking routine

What the grounding actually buys you

NotebookLM answers from the sources you selected, cites them inline, and will often reply "the sources don't cover this" instead of inventing something. A study of journalistic use measured roughly 13% of claims as unsupported, against around 40% for open chatbots on comparable tasks. That is a real improvement, and 13% is still one claim in eight. The right posture is neither trust nor paranoia; it is cheap verification.

How citations work

Each numbered chip anchors a claim to a passage. Hovering previews the quote; clicking opens the source at that spot. This is the whole safety system, and it works only when used. The failure mode it protects against has a name, citation drift: a chip that points at a passage which is related to the claim but does not actually support it. The number looks reassuring either way. Only reading the passage tells you which case you have.

Where errors sneak in

• Overreach on thin sources. Asked for more than the documents contain, the model sometimes stretches instead of declining, and the citation points somewhere vaguely nearby.
• Opinion flattening. A source's hedged position ("some researchers argue...") comes back as flat fact. The citation is technically correct, the certainty is invented.
• Garbage in, confidently out. NotebookLM repeats errors in your sources as faithfully as truths. Grounding checks provenance, not correctness.
• Audio amplification. Audio Overviews carry no citations and sound wonderfully confident. Documented failures include inverted negations, where the hosts state the opposite of the source. Verify in chat first; listen second.

The 60-second verification routine

1. Click the citations on load-bearing claims. Not all of them: the numbers, the dates, the quote you are about to put in a deck. Read the actual passage, not just its existence.
2. Run the audit prompt when stakes are high: "Act as a skeptical reviewer. Mark each sentence of your last answer as supported, inferred, or unsupported." Follow with "rewrite using only supported claims" if needed.
3. Make honesty the default with a standing instruction in chat settings: "If something is not in the sources, say 'not in the sources' rather than guessing." This measurably reduces overreach.

Structural habits that reduce errors before they happen

• Keep notebooks single-topic. Mixed notebooks produce cross-contaminated answers where a claim from one project leaks into another's summary.
• Scope questions with source checkboxes. Fewer competing documents means fewer opportunities to blend them wrongly.
• Ask for the counter-case explicitly. "Include cases that contradict the main pattern" recovers the outliers that synthesis quietly drops.

FAQ

Does NotebookLM hallucinate?

Less than open chatbots, since answers are grounded in your sources, but yes: roughly one claim in eight was unsupported in a journalism study. Clicking citations on important claims catches most problems.

Can I trust NotebookLM citations?

Treat them as pointers, not proof. Most are accurate; occasionally one points at a passage that only loosely supports the claim. Reading the passage takes seconds.

Are Audio Overviews as reliable as chat answers?

No. They carry no citations and have documented failure cases like inverted negations. Verify important material in chat, then use audio as the review format.