YouTube to automatically label AI-generated videos

Original Article ↗ Hacker News Discussion ↗

Detection capability and methods

  • Many doubt that AI-generated videos can be “automatically detected” reliably, citing prior AI-text detectors with high false positives and conceptual limits: human and AI outputs overlap.
  • Others think “good enough most of the time” is acceptable, especially to combat low‑effort AI spam, comparing it to imperfect email spam filters.
  • Several speculate YouTube will lean heavily on watermarks like SynthID and similar schemes (C2PA, camera signing), which can reduce false positives but miss content from unmarked tools or re-recorded output (“analog hole”).
  • Some foresee an arms race: models optimized to evade detection vs detectors trained on those evasions.

False positives, creator impact, and appeals

  • Strong concern that mislabeling human work as AI could hurt channels via reduced clicks, reputational damage, or algorithmic downranking.
  • YouTube’s existing automated moderation/appeals are widely seen as opaque and error‑prone, so people are skeptical creators will get fair recourse.
  • Others argue that labeling low‑effort videos as AI could nudge creators toward higher‑effort work, but this is contested given detectors don’t measure “quality”.

User controls and platform incentives

  • Very broad desire for:
    • A global “hide AI content” filter for homepage, search, shorts, and music.
    • The ability to explicitly tag one’s own content as AI-assisted.
    • Finer‑grained segment labels when only parts of a video use AI.
  • Many doubt YouTube/Google will offer strong AI filters unless metrics show AI slop hurts watch time; some expect only third‑party extensions to fill the gap.
  • Some note that different parts of Google have conflicting incentives: YouTube wants to control spam and keep creators happy, while other groups push generative tools.

Scope questions: what counts as “AI video”?

  • Edge cases debated:
    • AI b‑roll in otherwise human explainers.
    • AI dubbing, TTS narration, or AI‑written scripts over real footage.
    • AI upscaling, frame interpolation, relighting, or VFX.
  • Many want any AI use, however small, clearly disclosed; others think the focus should be on “realistic/deceptive” uses, not tools like upscaling.

AI slop, misinformation, and vulnerable users

  • Numerous complaints that search and recommendations are increasingly dominated by AI “slop”: psychology clickbait, history essays, slide‑shows with TTS, and spammy “news” or politics clips.
  • Some say their own feeds are fine; others report having to drastically reduce YouTube usage or rely only on known channels.
  • Particular worry about:
    • Children and seniors consuming endless AI junk or convincing fake “experts”.
    • Deepfake‑style political or health misinformation, especially ahead of elections.

Artistic and cultural debate (music and video)

  • Heated discussion around AI music on YouTube/Spotify:
    • Some feel deceived when they later discover tracks are AI‑generated and want labels and filters.
    • Others don’t care if they enjoy the result, especially for background or “focus” music, and see AI as a valid creative tool.
  • Several argue that human limitation, effort, and biography are central to why art matters; AI is seen as mass‑producing “slop” and crowding out human discovery.
  • Others use AI to realize personal ideas (e.g., Suno songs from their own melodies or lyrics) and say those works are meaningful to them despite the tooling.