Typing rhythms can reveal when AI wrote the whole thing, but they're blind to AI that only helped.

This explores whether the record of *how* a piece of work was produced, not the finished product, can show the difference between handing a task to AI and working alongside it. The corpus has one direct study, and its answer is lopsided: process data catches delegation well but can't see collaboration at all. In writing and programming logs, AI-delegated content shows up as concentrated bursts that break from the author's usual rhythm. Large chunks appear faster than that person ever types or edits. That gives a clear yes-or-no signal for wholesale delegation. Lighter collaborative help (a suggestion here, a rephrasing there) gets absorbed into the author's normal pace and looks the same as work with almost no AI help Can process data distinguish AI delegation from ordinary collaboration?.

That flips the question's premise in a useful way. The signature doesn't separate two kinds of AI use. It separates work done at a human pace from work that wasn't. Collaboration stays invisible because it doesn't disturb the person's rhythm. The detectable feature is the handoff itself: the point where the person stops shaping the work as it happens.

Why watch the process instead of judging the output? Research on agent workflows gives a strong reason. When frontier models are handed long document tasks, they silently corrupt about a quarter of the content over repeated passes. The errors pile up and stay hidden in spot-checked results Do frontier LLMs silently corrupt documents in long workflows?. Delegated output can look fine and still be degraded, so evidence of how it was made carries information the result alone doesn't. This ties to a broader framework for delegation, which treats verifiability (can anyone actually check the outcome?) as the foundation for deciding what to delegate at all What makes delegation work beyond just splitting tasks?. A process signature is one way to verify the *how* when you can't fully verify the *what*.

The agent-systems literature attacks the same problem from the opposite direction. Instead of inferring delegation after the fact, it builds systems where delegation is recorded by design. One voice-assistant design keeps delegated requests, their results, and the surrounding conversation in a single ordered timeline, so you can always see what was handed off and when it came back Can frontends handle delegation while staying conversationally engaged?. Another line of work anchors cryptographic fingerprints of agent actions and approvals. The process record then can't be altered later, and the sensitive content itself is never exposed Can commitments protect sensitive agent data while enabling verification?. Both are versions of the burst signature that are explicit and can't be faked: they don't rely on rhythm to guess.

The corpus is thin here. One empirical study speaks to human process signatures directly, and nothing yet tests whether the burst pattern survives people who deliberately pace or retype AI output. The adjacent material points to a real shift, though. As people delegate more, the trustworthy evidence moves from the artifact to the timeline that produced it.