Google’s AI thinks I left a Gatorade bottle on the moon
Google’s NotebookLM feature, which generates podcast-style summaries from user-provided documents, is being probed for how easily it can be manipulated by feeding it spoofed or cloaked webpage content, such as a fake claim about leaving a Gatorade bottle on the moon. Commenters liken this to long‑standing SEO “cloaking” tricks but warn that AI‑mediated answers raise higher stakes for misinformation, defamation, and subtle product or political manipulation, especially when users can’t see the underlying sources. Alongside these concerns, people react to the uncanny, shallow quality of the AI podcasts, debate their creative and educational value, and question whether current LLMs justify the enormous investment and energy they consume.
AI-Generated Podcasts: Quality, Style, and Uncanny Valley
- Many find the NotebookLM podcasts unsettling: realistic voices and emotions but no consistent “person,” memory, or evolving viewpoint.
- Listeners note odd reactions (fake surprise followed by detailed knowledge) and shallow, repetitive content.
- Others argue it accurately mimics a large class of existing podcasts: overproduced, padded with small talk, and full of generic praise.
- Some enjoy them as background noise or as satire; others say they’d stop listening within 30 seconds because it “fails the BS filter.”
- The built-in time-wasting banter bothers some, who see it as Google intentionally baking in filler.
NotebookLM Mechanics, Cloaking, and Scope of the “Attack”
- NotebookLM generates summaries and podcasts only from supplied documents/URLs.
- The “Gatorade on the moon” setup relies on serving different content to NotebookLM’s crawler than to normal browsers.
- Some see this as trivial, akin to old-school SEO “cloaking” and not yet shown to poison broader Google AI systems.
- Others emphasize the user-risk: the human sees one page, while the AI consumes a hidden version the user can’t inspect.
- Several expect Google to apply its existing anti-cloaking and spam defenses to NotebookLM, though note that prompt and spam attacks tend to persist.
Misinformation, Manipulation, and Real-World Examples
- Potential abuses mentioned: defamation with plausible deniability, manipulating product recommendations, and influencing elections.
- Another example is seeding a fabricated but plausible detail online and having Gemini repeat it as fact, sometimes with attribution, sometimes not.
- Some argue that LLM training already contends with absurd web claims; others fear LLMs will amplify and legitimize them.
User Experiences and Ethical Concerns
- People report comic results when feeding in resumes, blogs, or fiction; the AI podcasts treat mundane material with exaggerated enthusiasm.
- Some use these tools with children for fun and creativity; others argue this replaces kids’ own creative work with passive AI-generated content.
- Safety filters that block disturbing topics are viewed by some as infantilizing and overly puritanical.
Broader Reflections on LLMs
- Opinions split between seeing LLMs as overhyped, resource-wasting “junk AI” and as an early stage of a rapidly improving technology.
- There is debate over whether fundamental reasoning limits have budged since earlier models.
- Discussion touches on anthropomorphizing language like “thinks,” the risk of a “post-knowledge” culture, and how AI-native generations may use these tools differently.