OpenAI quietly resumes model training on Codex chats
openai training
| Source: HN | Original article
OpenAI appears to have quietly resumed training its models using data from Codex chat interactions.
OpenAI appears to have resumed using data from its Codex chat interface to train future models, a shift that has so far gone unannounced. Monitoring tools that track changes in OpenAI’s data‑usage policies flagged the re‑appearance of language in the terms of service that permits the collection of conversational snippets for model improvement. The move follows a period during which the company had publicly limited such training on user‑generated code discussions, a step that was welcomed by developers concerned about the confidentiality of proprietary code.
The change matters because Codex chats are a primary channel through which developers seek real‑time assistance with programming tasks. If those interactions are fed back into the training pipeline, the resulting models could inherit snippets of private or commercially sensitive code, raising questions about intellectual‑property rights and data privacy. The practice also touches broader regulatory scrutiny of how large language‑model providers handle user data, especially in jurisdictions that are tightening rules on automated processing of personal and proprietary information.
Stakeholders will be watching for an official statement from OpenAI clarifying the scope of the data collection and any accompanying safeguards. Regulators may probe whether the company’s current disclosures satisfy emerging transparency requirements. Developers using Codex are likely to reassess their reliance on the service, possibly shifting to self‑hosted or open‑source alternatives if confidence erodes. The next few weeks should reveal whether OpenAI will adjust its policy, introduce opt‑out mechanisms, or face formal inquiries from data‑protection authorities.
Sources
Back to AIPULSEN