GitHub Copilot starts training on interaction data

This is the kind of move that makes business sense and trust headaches at the same time: GitHub is turning Copilot usage into a model-quality moat, but it is also asking individual users to accept a much broader data tradeoff.

• –The data bucket is broad, covering prompts, accepted edits, code snippets, cursor context, repo structure, feature interactions, and feedback.
• –GitHub is drawing a hard line between consumer plans and Business/Enterprise, which keeps enterprise sales safe but leaves individuals to decide whether model improvement is worth it.
• –Existing opt-outs carry over, but the settings copy needs to be crystal clear or users will read this as a bait-and-switch.
• –GitHub says the data may flow to Microsoft affiliates, not third-party model providers, which should reassure some compliance teams but not eliminate privacy concerns.
• –The move mirrors the broader AI market: consumers subsidize the training loop, enterprises pay for the firewall.