Someone in the NotebookLM subreddit recently threw a polite but pointed tantrum: the app now makes gorgeous slides, videos and even podcasts, and then merrily gets the facts wrong. The original poster’s gripe was simple and relatable: multimedia outputs that look delightful but recite a muddle of blended sources, omit facts, or present flat-out contradictions. They insist it’s not a prompt or source problem; the fidelity has dropped since the “pretty things” push. (Thread: https://www.reddit.com/r/notebooklm/comments/1p7908r/notebook_is_now_style_over_substance/)
Some shrugged and suggested alternatives (ChatGPT Projects, SciSpace, Nouswise). Others said this was a familiar product lifecycle: start focused on function, chase features, then, hopefully, settle into a stable combo. A recurring worry: has NotebookLM pivoted toward students and flashy studio artifacts at the expense of reliability for professional workflows?
The drama matters because NotebookLM occupies an odd, useful niche: many people use it as a RAG (retrieval-augmented generation) tool to pull facts out of their documents and trust the tool for low-hallucination, source-grounded answers. If the grounded-search part loses accuracy while the multimedia front-end blooms, trust, which is the real product here, erodes fast.
Why users are upset
- Hallucinations and conflation: The OP describes the system blending multiple sources into incoherent conclusions, or conflating separate items into one. That’s not a UI problem; it’s a model-behavior problem.
- Consistency regression: Several replies echo the sentiment that things “used to work” and now don’t, which is a particular kind of frustration. You’re paying attention to a tool, you learn its quirks, and then those get shuffled around by a new release.
- Product focus: A few commenters suspect a strategic pivot to multimedia and student-focused features, leaving enterprise-grade precision on the bench.
How to handle this now (practical advice)
- If you need accuracy, treat NotebookLM like a fast clerk, not a final authority: extract facts and verify sources rather than accepting polished artifacts verbatim.
- Supply one clear, authoritative source when possible. The OP claims they tried that and still had issues, but the single-source approach remains the lowest-risk route.
- Keep an eye on alternatives: ChatGPT Projects, Nouswise, SciSpace, and classic RAG pipelines are mentioned in the thread. Test them for your use case before migrating fully.
- Log bugs and examples. If a product is pivoting, hard examples help push priorities back toward reliability.
Context and why this is happening
There’s a tension at work in many AI products: aesthetic features, like slides and videos, are shareable, demo-friendly, and important for adoption, while core model reliability is compute- and research-heavy to maintain across modalities. Teams sometimes ship the flashy layer first to drive growth and then iterate on substance, and that interim phase is exactly where users feel betrayed.
This update has sparked interesting reactions from the community: some are brushing it off as early-stage growing pains; others see a worrying product pivot away from the tool’s original strength. If you rely on NotebookLM for work, double-check outputs, explore backups, and keep constructive evidence ready for the product team. Meanwhile, Reddit was busy debating whether beauty should get priority over brains, and as usual, both sides had a point.
Leave a Reply