OpenAI seems to be silently re-enabling model training on Codex chats
Concerns are mounting that OpenAI’s Codex has quietly re-enabled training on user chats and obscured usage and privacy controls, raising fears that sensitive code and company IP may be ingested despite prior opt-outs. Commenters debate whether this stems from malice, incompetence, or business incentives, and broaden the critique to other AI providers like Anthropic and Google, pointing to dark patterns, weak contractual assurances, and opaque retention policies. Many conclude that organizations serious about data protection should either self-host open models or use tightly governed cloud offerings rather than trust frontier AI labs’ checkboxes and settings.
Reported re-enabling of model training settings
- Several commenters say Codex/OpenAI’s “use conversations to improve models” setting, previously disabled, appears to have been silently re-enabled for them or coworkers.
- Others report their setting remains off, creating conflicting anecdotes.
- Some suspect dark-pattern UI changes (moving settings, changing labels, or toggles that are easy to misclick), while others suggest it could be a bug or “vibe-coded” mishap rather than deliberate malice.
- A few point to prior incidents with other providers where org-wide dials changed from opt-in to opt-out and back, resetting sliders to “on.”
Trust, consent, and legal/regulatory context
- Many express deep distrust of “frontier labs,” citing prior mass scraping of books and web content, and arguing there’s little incentive to respect opt-out checkboxes when enforcement is weak.
- Some claim even with training disabled, providers can keep data for abuse detection, analytics, or synthetic-data generation, functionally “training without training.”
- There is discussion of regulators (e.g., FTC) warning about quietly changing terms or consent settings, but skepticism that authorities will act meaningfully.
- Lawsuits and zero data retention (ZDR) contracts are mentioned as theoretical remedies, but commenters doubt practical enforceability or discovery.
Enterprise IP and security concerns
- Many argue that companies serious about IP protection should not send proprietary code to opaque third-party models at all.
- Suggested mitigations:
- Run your own inference on open models (e.g., via HF, Qwen, DeepSeek, GLM), possibly on confidential compute in major clouds.
- Use gated models via Azure/GCP/AWS or Bedrock with ZDR-style guarantees instead of consumer-facing products.
- Enable advanced account security and audit logs; periodically export data and check caches.
- Others note cost and expertise barriers for small orgs; some enterprises already block external LLMs except in tightly controlled contexts.
Comparisons with other providers and products
- Similar concerns are raised about Anthropic (feedback channels, Daybreak Blue policy changes) and Google Gemini’s “activity” settings, described as dark patterns tying chat history retention to training.
- Some teams report building tooling to monitor many agent accounts and alert if training is toggled on.
Meta attitudes and risk tolerance
- Commenters diverge sharply: from “assume everything is used for training” and “don’t trust this company” to more forgiving views that software quality issues, not malice, may explain behavior.
- There’s ongoing debate about what level of caution is “paranoid” versus simply consistent with pre-AI security norms.