‘Overworked, underpaid’ humans train Google’s AI

Original Article ↗ Hacker News Discussion ↗

Human labor is pervasive across AI companies

  • Commenters say Google is not unique: OpenAI, Scale AI, Surge, Meta, and many others rely on large pools of low-visibility human raters and labelers, often in developing countries.
  • Multiple data-labeling vendors and platforms are listed (Surge/DataAnnotation, Scale/Remotasks/Outlier, Mercor, etc.), with suggestions that “millions” of annotators may be involved industry-wide.
  • Some note this continues a long history from Mechanical Turk and traditional moderation/labeling work.

Exploitation, wages, and “digital colonialism”

  • Reported pay spans a wide range: ~$16–21/hr for US raters in the article, up to ~$45/hr for specialized contractors; however, others cite African and South American workers earning under $2/hr with rates repeatedly cut.
  • Critics argue this exploits legal and safety gaps in the Global South (poor recourse, weak mental health support, PTSD from extreme content) and amounts to “digital colonialism.”
  • Defenders frame it as voluntary market work at local rates; opponents counter that power imbalances and lack of protections make “choice” dubious.

Alignment, RLHF, and whose values

  • Some emphasize that RLHF/RLAIF is essential not just for “values” but basic chat behavior and usability.
  • Others challenge claims about “human values,” arguing models are actually aligned to Google’s and its customers’ commercial and political interests (ads, lock-in, moderation norms), not any universal morality.
  • There’s debate over whether this fine-tuning meaningfully affects “truth” versus just surface behavior.

Evaluation vs training and Google’s statement

  • Google’s line that raters’ work “does not directly impact” models is dissected:
    • Some argue it’s technically true if used only for evaluation/validation, not as gradient-updating training data.
    • Others say that because eval metrics steer future model changes, the effect is indirect but still real, making the statement misleading.

Job quality, harms, and media framing

  • Some participants say the Guardian piece is overblown “ragebait” about a fairly typical freelance desk job, better than many call centers or manual labor roles.
  • Others highlight worrisome anecdotes: pressure to prioritize speed over quality, lack of psychological support, and non-experts asked to process sensitive medical or disturbing content.
  • Several note the broader pattern: safety and ethics are prioritized only until they slow product timelines, after which “speed eclipses ethics.”