Users have reported a potential shift in how OpenAI handles data for model training, specifically noting that the 'Improve the model for everyone' setting in Codex appears to have been silently re-enabled. This setting, found within the 'Data Controls' section of OpenAI's interface, determines whether user conversations are utilized to train and refine the company's AI models. The report suggests that even after users had intentionally disabled this feature to protect proprietary information, the toggle reverted to an enabled state without user intervention.
This development highlights a significant concern for companies and developers who use AI tools to handle sensitive intellectual property (IP). Because frontier AI labs process vast amounts of prompt data, there is a persistent risk that confidential code or internal strategies could be ingested into the model's training set. Some observers in the community argue that relying on software checkboxes for privacy is insufficient for high-stakes data protection, suggesting that organizations with strict security requirements should instead opt for self-hosted inference to maintain complete control over their data.
The situation reflects broader industry-wide challenges regarding data transparency and 'dark patterns' in AI services. Commenters have pointed to similar experiences with other providers, such as Anthropic's automated feedback triggers and Google's 'Gemini Apps Activity' settings, which can obfuscate how much data is retained for model improvement. While some users suggest that Zero Data Retention (ZDR) contracts or advanced account security (such as hardware keys) offer more robust protections, the core tension remains: the convenience of using frontier AI models currently requires a level of trust in the provider's data handling that many organizations find difficult to guarantee.