What Data GitHub Copilot Sends: Context, Telemetry and Training Controls
🔍 WiseChecker

What Data GitHub Copilot Sends: Context, Telemetry and Training Controls

Using GitHub Copilot can send code and other context to an AI service even when optional editor telemetry is disabled. That is because the information needed to answer your request is not the same as a usage counter or crash report. Before enabling Copilot on confidential work, identify the feature, account, plan, model provider and organization’s policy.

There is no reliable single promise that “telemetry off” means “only authentication traffic.” This article separates the relevant categories and checks.

ADVERTISEMENT

Distinguish Request Context From Telemetry

Request context is the information a model needs to generate a response. For inline suggestions, this can include surrounding code and relevant open files. For an agent task, it can also include file reads, search matches, terminal output, prior messages and explicitly attached material. See GitHub’s inline-suggestions application card and VS Code’s agent-context explanation.

Usage and diagnostic telemetry serves purposes such as measuring feature use or investigating errors. It is not the only network activity in the editor. OpenTelemetry export is another, separately configured mechanism for recording agent activity to a chosen collector.

Do not assume configuration files, credentials or personal data are automatically removed from every request. Choose what the agent can access, avoid pasting secrets, and review tool output. A context window is not a promise that only the current line or a fixed number of neighboring lines will be sent.

Check Model-Training and Account Controls

GitHub’s current individual-plan policy differs from older blanket “never used for training” statements. For Copilot Free, Pro, Pro+ and Max, interactions may be used for training unless you opt out. Open the GitHub profile menu, select Copilot settings, and set Allow GitHub to use my data for AI model training to Disabled if that is your preference.

GitHub says Business and Enterprise customer data is not used to train AI models without customer authorization under its data-protection agreement. The individual-plan control is therefore not shown for those seats. Verify which account is actually supplying your license. The authoritative detail is in GitHub’s policy guide.

Training permission, request processing and retention are separate questions. A training opt-out does not stop a model receiving your prompt. Do not infer one universal retention period across IDE chat, cloud agents, third-party providers and saved session features; check the terms for the product and plan you use.

ADVERTISEMENT

Review Editor and Monitoring Settings

In VS Code, search Settings for telemetry.telemetryLevel and choose off to disable the editor’s optional telemetry going forward. This does not erase historical records, switch Copilot to local inference, or establish that every extension has stopped collecting data. See VS Code’s telemetry scope.

For a JetBrains IDE, inspect the installed GitHub Copilot plugin’s actual settings and current documentation. Do not assume there is a universal “Enable telemetry” checkbox that also stops sending code context. Completion preferences are not data-retention controls. GitHub maintains separate IDE setup instructions.

If your team uses Copilot Chat OpenTelemetry, check whether export is enabled, the collector destination, content capture and any managed policy. User settings can be overridden by enterprise settings. Local debug records and remote export are also distinct. Consult the monitoring guide before collecting or sharing traces.

Verify Without Exposing More Information

  1. Document the client version, active account and plan, selected feature and applicable organization policies.
  2. Use a synthetic sample, not a production credential, to inspect which references and tool results the assistant uses.
  3. In VS Code, Developer: Show Telemetry helps examine editor telemetry. It is not a complete audit of all extension or service traffic.
  4. Inspect diagnostic files locally and redact sensitive data before providing them to support. Avoid publishing a raw prompt/session log.
  5. If network-level assurance is required, have the security team perform an approved assessment. An ordinary packet capture of encrypted HTTPS does not reveal full request bodies.

If the requirement is “this data must never reach a cloud model,” enforce that through approved access restrictions or a suitable offline workflow. Do not approve confidential use merely because a setting is unchecked or a log is quiet. If a secret has already been exposed, rotate it and follow the incident-response process.

Official documentation checked: October 14, 2026.

ADVERTISEMENT

Did this guide solve your problem?