GitHub Copilot to Start Collecting Telemetry from April 24 to Improve AI
Starting April 24, 2026, GitHub will enable data collection on developer interactions with Copilot for Free, Pro, and Pro+ plans. This will allow training models on real-world scenarios: from code drafts to refactoring. Business and Enterprise subscriptions remain fully private.
Why Real-World Data Improves AI Quality
Traditional datasets from public repositories and synthetic data limit AI assistants. They don't reflect the code creation process: iterations, errors, project context. Integrating logs from Microsoft users showed an increase in accepted suggestions across multiple languages. Models now better understand architecture, offer precise autocompletion, and detect vulnerabilities during writing.
Training on live projects provides advantages:
- Analysis of the path from draft to final, including typos and edits.
- Accounting for local repository context and file navigation.
- Evaluation of reactions: accepts, edits, likes/dislikes.
- Avoiding scanning of private repositories at rest, but processing active sessions.
What Data Will Go into the Dataset
The system aggregates parameters without manual telemetry opt-out:
- Accepted or edited code fragments.
- Prompt texts in the Copilot chatbot.
- Local context from the IDE (file and repository structure).
- User comments and suggestion ratings.
- Session metrics: navigation, interaction time.
Active work in a private repository with the plugin enabled triggers data sending to servers for training. This isn't scanning the storage, but only the current workflow.
Opt-Out and Privacy Guarantees
Opting out of data collection is available in GitHub profile settings. If the option is already disabled, it will remain so. Data is stored exclusively in GitHub and Microsoft infrastructure, without sharing with third parties—not AI providers or labs.
Enterprise customers are protected by contracts: their code and logs are not used in training.
Key Points
- Telemetry collection only for Free/Pro/Pro+; Business/Enterprise excluded.
- Data includes workflow but doesn't passively scan private repositories.
- Opt-out is simple, via profile, with settings preserved.
- Goal: improve quality—autocompletion accuracy, bug detection.
- No data sharing with external providers.
Historically, analyzing user behavior (clicks, sessions) radically improved search engines in the 2000s. A similar approach for Copilot promises a breakthrough in code generation for mid/senior developers, where complex project context matters.
— Editorial Team
No comments yet.