Copilot safeguards: human-centred controls vs technical mitigations
Verdict: Copilot pairs technical mitigations - context scoping and safety filters that run in the pipeline before a suggestion shows - with human-centred controls: AI-generated suggestions are visually distinct and never applied without a deliberate accept.
| Criterion | Human-centred controls | Technical mitigations |
|---|---|---|
| Examples | Suggestions shown as AI-generated; no change without explicit accept | Context scoping before the model; content and code safety filters |
| Where they act | In the editor, with the developer | In the pipeline, before a suggestion is shown |
Rules
- Technical mitigations: code context is structured and scoped before it reaches the model, and content and code safety filters run on the output.
- Human-centred controls: suggestions are visually distinct as AI-generated and require deliberate acceptance.
- Contextual filtering, proxy authentication and post-processing safety checks all run before a suggestion is displayed.
Traps
- Safety filters and context scoping are technical safeguards, not human-centred controls.
- There is no enforced delay between suggestions and no rejection leaderboard; those are not documented mitigations.
- Accepting and committing code are developer actions after the suggestion appears, not pipeline stages.