March 27, 2026 Microsoft is updating GitHub Copilot to train on real-world developer interactions, expanding beyond public code datasets to include prompts, generated outputs and usage patterns. The change matters because it shifts Copilot’s improvement strategy toward live developer workflows, while introducing an opt-out model that automatically enrols eligible users unless they disable it.
The update applies to Copilot Free, Pro and Pro+ users, whose interaction data may be used for training starting April 24 unless they opt out. According to GitHub, enterprise tiers, including Copilot Business and Enterprise, as well as enterprise-owned repositories are excluded from this policy.
GitHub said interaction data can include prompts, generated suggestions, accepted or modified code, file context, repository structure and developer feedback. The company emphasised that data “at rest,” such as private repositories, issues and discussions, is not used for training, and that information is not shared with third-party AI model providers.
“We believe the future of AI-assisted development depends on real-world interaction data from developers like you,” said Mario Rodriguez, chief product officer at GitHub, adding that the approach is intended to improve model accuracy and better reflect real development workflows.
The decision is a move across AI tooling toward training on usage data rather than relying solely on static datasets. GitHub previously trained Copilot models on public repositories and curated examples, but has already incorporated internal Microsoft data and employee usage signals, which it said led to measurable quality improvements.
From a product standpoint, the goal is to refine outputs across multiple dimensions: more accurate code suggestions, improved contextual awareness and earlier detection of potential issues. Interaction-level telemetry also gives the model visibility into how developers actually edit, accept or reject suggestions, a signal that static training data cannot provide.
However, the rollout introduces a clear trade-off between model quality and data governance. Eligible users are automatically opted in unless they actively disable data sharing in their settings before the April 24 deadline. Users who have previously opted out will remain excluded without further action.
GitHub said interaction data used for training may be shared with affiliates, including Microsoft, but reiterated that it will not be provided to independent third-party model providers. The company also maintained that participation is optional and does not affect access to Copilot features.
