How AI code authorship is attributed

What does AI code attribution measure?

AI code attribution describes the evidence that a coding tool contributed to a change. The evidence may come directly from a tool, from Git co-authorship, or from an estimate based on observed activity. Token Intelligence labels these methods so you can interpret the contribution signal.

Attribution is distinct from adoption. Opening an assistant or making requests establishes usage; it does not prove that its suggestions reached a merged PR. Likewise, a tool’s committed-code measure is not automatically a measure of code delivered to production.

What do verified, inferred, and estimated mean?

“Verified” identifies a direct source. It does not claim that every line in a PR has a known origin or that a request can always be traced to a merged change. Read the field definition beside the label.

Commit metadata can identify recorded assistance, but missing metadata does not establish human-only authorship. An activity estimate provides context when direct signals are unavailable. It should remain visibly estimated in any report built from it.

How should you handle multiple tools on one PR?

An engineer can use Cursor, Claude Code, and another assistant on the same task. Tool-level contribution signals may overlap. Adding each tool’s attributed PR count can therefore overstate the number of distinct PRs.

For an evaluation, keep a distinct set of eligible PRs and show tool evidence alongside it. Check identity mapping between the coding tool and source control. Inspect a sample where several tools appear, and document how the report handles that overlap. Do not assume that a tool association divides credit precisely between human and AI work.

Shared branches, rebased or squashed commits, missing telemetry, and disconnected repositories can weaken a comparison. Record unknowns as unknowns. A lack of observed AI evidence does not prove the absence of AI assistance.

Does authorship establish cost or ROI?

Code contribution, billed spending, and output need separate sources. A direct contribution signal may have incomplete cost coverage. A billed amount can cover a whole organization without a reliable PR-level allocation.

Use code output measurement to understand Weave’s expert-hour unit. Use the AI ROI methodology for matched periods, attributed costs, and quality context. Neither contribution share nor output per dollar establishes a causal financial return by itself.

Keep source labels when exporting or presenting findings. A board slide that turns an activity estimate into “verified AI-written code” changes the meaning of the original measurement.

What should a team validate first?

Choose one connected tool, one repository set, and one reporting period. Confirm the native fields available under your account, match identities, and inspect representative commits and merged PRs. Then compare contribution evidence with adoption, cost sources, review outcomes, and rework.

The Cursor and GitHub Copilot coverage guides explain the tool-specific questions to bring to that review. Expand the scope after the first sample is understandable and the missing-data boundaries are recorded.