• Home
  • Technology
  • Gaming
  • Entertainment
  • World & Business
  • Science
  • Sports
  • AI
HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
  • HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
    • Home
    • Technology
    • Gaming
    • Entertainment
    • World & Business
    • Science
    • Sports
    • AI
    AI
    Announcement

    Adding AI-generated web text reportedly makes models worse than adding no data under most standard training budgets

    The researcher says almost a third of FineWeb-filtered web text was AI-generated as of September 2026, and the share was rising.

    YA
    MI
    BD
    4 Sources, ,

    TLDR

    A researcher says their team pretrained hundreds of models using different mixes of human-written and AI-generated web text. They found that, across most standard training budgets, adding AI text made models worse than adding no new training data. The researcher also says almost a third of FineWeb-filtered web text was AI-generated as of September 2026, and that the share was rising fast.

    Combined views

    13.1K

    4 Sources, first seen 3h ago

    Combined views

    13.1K

    4 Sources, first seen 3h ago

    211 likes
    3h ago
    first seen 3h ago
    211 likes
    12 comments
    67 saves
    55 reposts

    Sentiment

    Positive——Negative

    Summary

    Not enough discussion yet.

    No sentiment analysis available yet.

    Featured Source
    12 comments
    67 saves
    55 reposts

    Sentiment

    Positive——Negative

    Summary

    Not enough discussion yet.

    No sentiment analysis available yet.

    Today's Rank

    #8

    Today's Rank

    #8

    4 Sources

    @jennajrussellHow much is an AI token worth? We fit scaling laws to find out! In August 2026, Pangram labels 31% of FineWeb-filtered web tokens as AI-generated, up from 10% in June 2024. We pretrained 800 LMs (19.9M–973M params) on mixes of human and AI web text to measure what it does to pretraining 🧵
    @MohitIyyerTurns out that it's not just humans whose brains melt when reading AI slop: we pretrain hundreds of models with different mixes of human-written and AI-generated web text, and we find that across most standard training budgets, adding AI text makes models worse than if we added no new training data at all. This is increasingly pressing since almost a third of FineWeb-filtered web text is AI-generated as of last month, and this number is rising fast!
    @yoavartzi@MohitIyyer Mode collapse?
    @bhuwandhingraRT @jennajrussell: How much is an AI token worth? We fit scaling laws to find out! In August 2026, Pangram labels 31% of FineWeb-filtered…

    4 Sources

    @jennajrussellHow much is an AI token worth? We fit scaling laws to find out! In August 2026, Pangram labels 31% of FineWeb-filtered web tokens as AI-generated, up from 10% in June 2024. We pretrained 800 LMs (19.9M–973M params) on mixes of human and AI web text to measure what it does to pretraining 🧵
    @MohitIyyerTurns out that it's not just humans whose brains melt when reading AI slop: we pretrain hundreds of models with different mixes of human-written and AI-generated web text, and we find that across most standard training budgets, adding AI text makes models worse than if we added no new training data at all. This is increasingly pressing since almost a third of FineWeb-filtered web text is AI-generated as of last month, and this number is rising fast!
    @yoavartzi@MohitIyyer Mode collapse?
    @bhuwandhingraRT @jennajrussell: How much is an AI token worth? We fit scaling laws to find out! In August 2026, Pangram labels 31% of FineWeb-filtered…