NotebookLM launches feature to customize and guide audio overviews

Original Article ↗ Hacker News Discussion ↗

Overall sentiment and use cases

  • Many find NotebookLM genuinely useful, especially for:
    • Summarizing technical docs (datasheets, regulations, tenders, research papers).
    • Extracting themes from large corpora (e.g., many postmortems, forum threads).
    • Turning blog posts or academic work into accessible overviews.
  • Audio overviews are praised as good introductions or “better than nothing” where high-quality human content is scarce.
  • Others see it as more entertainment than serious learning and dislike the format.

New customization for audio overviews

  • Users welcome the ability to steer style, roles, tone, and audience.
  • Prompts like “expert + novice host” or whimsical character setups significantly improve clarity and engagement.
  • Some had already been “prompt-hacking” by uploading instruction files; they dislike the new 500-character limit and prefer longer, version-controlled prompts.
  • Small prompt tweaks can dramatically change length and depth of episodes.

Quality, style, and voices

  • Some report surprisingly high quality: accurate, concise, more factual than many human podcasts.
  • Others find the output shallow, cliché-filled, and “overproduced,” especially compared to deep, long-form human shows.
  • The two default voices are divisive: recognizable and helpful for provenance, but monotonous and easily mistaken for the user.

Spam, fake podcasts, and discovery

  • Evidence of thousands of AI-generated shows already being removed from at least one index.
  • Concern that ultra-cheap podcast generation will:
    • Flood directories with low-value content.
    • Degrade search and discovery, similar to SEO spam and AI image pollution.
  • Counterarguments:
    • Podcast space is already 99% ignored; more slop may not matter.
    • Good ranking and curation should keep spam out of listener feeds.
    • Some view this as an opportunity to build better discovery and provenance tools.
  • Debate over whether and how Google should provide watermarking or detection; feasibility is disputed.

Privacy, identity, and Google mistrust

  • Suspicion that, despite assurances, uploaded content could someday train models.
  • Some are uneasy about Google’s history and product shutdowns, fearing lock-in and eventual deprecation.
  • Separate concern: tools like this make it trivial to aggregate a person’s public posts across sites into detailed profiles, raising new privacy and reputational risks.

Alternatives and ecosystem

  • Multiple open-source or indie tools aim to replicate or extend NotebookLM (podcast generation, note-taking, interactive “Duolingo for any subject”).
  • Users ask for an official API and clearer positioning within Google’s broader AI stack.