OpenAI fails to deliver opt-out system for photographers

Original Article ↗ Hacker News Discussion ↗

OpenAI’s failure to ship a promised opt‑out tool for photographers reignites criticism of how AI companies scrape creative work for training without meaningful consent. Commenters argue that requiring creators to individually submit and describe every image is a dark-pattern “opt-out” that shifts all burden onto rights holders, while companies lobby for training on copyrighted material to be treated as fair use. The exchange broadens into whether current copyright law can or should protect artistic styles and online content from large-scale AI training, and what kind of legal or political pressure would be needed to change industry behavior.

Opt-out system & consent

  • Many see OpenAI’s undelivered “Media Manager” / opt-out as evidence they don’t genuinely want data excluded, especially since photographers must submit each work with detailed descriptions.
  • Commenters argue the burden is absurd at scale: creators would have to track and use opt-out mechanisms for many AI firms.
  • Several say consent should be opt‑in, not opt‑out: OpenAI should ask before using works, as most others must.
  • Tech companies’ approach to consent is criticized as showing disregard or even contempt, with analogies to invasive or predatory behavior.
  • Some note precedents like Google’s _NOMAP Wi‑Fi suffix as similarly lopsided “opt-out” schemes.

Copyright, fair use, and training data

  • One side claims training on scraped content is clearly fair use: models create non-expressive abstractions, are transformative, and don’t “copy” in the copyright sense.
  • Others argue it should be infringement, especially as models begin to substitute for the market of original works and occasionally regurgitate them.
  • Multiple people stress the law is unsettled, with many lawsuits pending; any “it’s clearly X” position is disputed.
  • There is debate over analogies to humans learning from books or art:
    • Pro-AI side: learning isn’t infringement; output is only a problem if it reproduces protected expression.
    • Critical side: scale, automation, and corporate profit make this fundamentally different.

Artists’ livelihoods, styles, and compensation

  • Some argue artists should be able to exclude their work and even force retraining of models that used it without consent.
  • Others note that style is generally not protected, and that artists have always learned by copying others.
  • Counterpoint: machines can replicate a style in days and produce near‑infinite derivatives, creating an uneven playing field and disincentivizing innovation.
  • Suggested remedies include mandatory compensation schemes for training use, akin to music royalties, and updated licenses for code and writing.

Legal / policy expectations

  • Several expect courts or legislatures to eventually clamp down, especially under pressure from large rights‑holders (e.g., media companies).
  • Others think powerful AI firms will win favorable rules (e.g., training classified as fair use), especially if framed as essential for innovation or AGI.

Double standards & platform behavior

  • Commenters highlight a perceived two‑tier system: everything online is fair game for training, but model weights and AI outputs are aggressively protected.
  • Policies forbidding training on AI outputs are seen as hypocritical when those models were trained on uncredited human work.

OpenAI, AGI, and trust

  • Strong distrust toward OpenAI is common: accusations of broken promises, bait‑and‑switch from “open” non‑profit roots, and prioritizing profit over creators.
  • Some frame the work as so important (potential AGI, “benefit of humanity”) that copyright concerns are treated as secondary.
  • Several express skepticism that current LLMs can reach AGI, noting hallucinations, lack of true understanding, and mostly incremental scaling rather than paradigm shifts.