LLMs can get "brain rot"

What the paper is claiming (in lay terms)

• Researchers simulate an “infinite scroll” of social media and mix in different tweet streams:
  • Highly popular tweets (many likes/retweets).
  • Clickbait-detected tweets.
  • Random, non-engaging tweets.
• They use these as continued training data for existing LLMs and then test the models.
• Models exposed to popular/engagement-optimized content show:
  • Worse reasoning and chain-of-thought (“thought-skipping”).
  • Worse long-context handling.
  • Some degradation in ethical / normative behavior.
• Popularity turns out to predict this “brain rot effect” better than content-based clickbait classification.

“Garbage in, garbage out” vs anything new here?

• Many commenters say the result is unsurprising: low-quality data → low-quality model.
• Others argue the value is in quantifying:
  • Which kinds of bad data (engagement-optimized) are most harmful.
  • That relatively early/pre-training damage is not fully fixed by post-training.
• Some see it as basic but still legitimate science: obvious hypotheses still need to be tested.

Data curation, modern training practice, and moats

• Several note that major labs no longer just scrape the internet; they:
  • Filter heavily (e.g., quality filters on Common Crawl, preference for educational text).
  • License or buy curated datasets and hire human experts, especially for code and niche domains.
• Others doubt how “highly curated” things really are, pointing to disturbing outputs from base models and lawsuits over pirated books.
• There’s concern that as the internet fills with AI-generated slop, early players with access to pre-slop data gain a long-term advantage.

Objections to the “brain rot / cognitive decline” framing

• Multiple commenters criticize the use of clinical or cognitive metaphors (“brain rot”, “lesion”, “cognitive hygiene”) for non-sentient models.
• They worry this anthropomorphizes LLMs, muddies thinking, and lowers scientific standards; some call the work closer to a blog than a rigorous paper.

Human brains, media diets, and feedback loops

• The paper prompts analogies to humans:
  • Worries about kids (and adults) consuming fast-paced, trivial content and possible long-term effects.
  • Comparisons to earlier TV eras (e.g., heavy preschool TV watching) with mixed interpretations.
• Commenters note a feedback loop risk:
  • People use LLMs, which may atrophy their own writing/thinking.
  • Their weaker content becomes part of future training data, further degrading models.
• There’s debate over using LLMs for writing: some see it as harmless assistance; others see it as outsourcing thought and producing empty, marketing-style “slop” that is now visibly creeping into research prose.