Tell HN: OpenAI keeps re-enabling the 'allow training' setting

Hacker News Discussion ↗

OpenAI users report that the “allow training” / “Improve the model for everyone” setting in ChatGPT sometimes appears to turn itself back on, raising concerns that the company may be overriding or failing to persist privacy preferences. Commenters debate whether this is malice, incompetence, or a UI/backend bug, and point out the difficulty of proving what a user actually selected without specialized forensic tools. The thread broadens into skepticism about trusting any major AI provider with sensitive data, comparisons with competitors’ practices, and reminders to use formal opt-out mechanisms or self-hosted models when privacy is critical.

Re-enabling of training / “Improve the model for everyone”

  • Several users report the ChatGPT “allow training / improve the model for everyone” toggle being turned back on after they had disabled it, sometimes after subscription changes or app updates.
  • Others explicitly checked and found their setting still off, in some cases for years, suggesting inconsistent behavior rather than a universal issue.
  • Some worry that once the toggle is silently re-enabled, past conversations may already be used for training before the change is noticed.

Multiple opt-outs and “Advanced Account Security”

  • There are at least two mechanisms mentioned:
    • In‑app setting: Settings → Data Controls → “Improve the model for everyone.”
    • Privacy portal: “Do not train on my content” request form.
  • Commenters are confused whether both are needed; one OpenAI employee says either one is sufficient and that the portal also covers other products and logged‑out users.
  • “Advanced Account Security” is cited (with OpenAI docs) as automatically excluding conversations from training, but some note the main UI doesn’t clearly connect this to training, prompting calls for better documentation.

Bug vs. deliberate dark pattern

  • Some see this as intentional dark patterning, likening it to YouTube autoplay, Facebook privacy defaults, and OS updates that reset or ignore user preferences.
  • Others propose mundane explanations: buggy optimistic UI, failed backend updates, configuration corruption, or product changes that reset settings.
  • There’s debate over whether repeated re-toggling is worse (because it’s demonstrable manipulation) than silently ignoring the setting.

Trust, legality, and data use

  • Many express deep distrust of OpenAI, citing:
    • Training on scraped/copyrighted data.
    • Allegations of IP “scooping” and other ethically dubious behavior.
    • General incentives of profit-driven AI labs.
  • Some argue that if users can’t prove they opted out, the burden should fall on the company; others note that in practice, company logs usually dominate in legal disputes.
  • Suggestions for “proof” include recording TLS sessions with keys or using forensic web capture tools, though this is acknowledged as complex.

Alternatives and coping strategies

  • Several recommend:
    • Using open-weight or on‑prem models instead of cloud services.
    • Switching to competitors (Anthropic, Bedrock, Cursor, etc.), though some report similar or other privacy/dark‑pattern concerns with those as well.
  • Others have left OpenAI products entirely after prior opt‑out bugs or billing issues.