Can process data distinguish AI delegation from ordinary collaboration?

The paper argues that AI assistance carries a temporal signature that holds across mediums: "AI contributions arrive in bursts far outside the author's own baseline" in both writing and programming, with paired effect sizes of dz = 1.13 and 3.54. The evidence comes from three public corpora. CoAuthor supplies 1,447 keystroke-level co-writing sessions, RealHumanEval supplies editor telemetry from 243 programmer records, and a pre-LLM CS1 corpus of 5.1 million keystrokes serves as a human-only baseline. The comparison is among minimal-AI work, collaborative AI use, and simulated wholesale delegation.

The mechanism is in how text arrives, not in what it says. Concentrated production sits far from a person's own typing rhythm, so a process log can register it without any detector reading the finished artifact. The authors frame this through the Learning Visibility Framework: what AI removes from assessment is visibility into the learning process, so assessment needs "process-based evidence alongside outcomes" because final artifacts and product-only AI detectors are unreliable. The result is deliberately asymmetric. Ordinary collaboration is "only weakly separable" from minimally assisted work, and at most 0.5% of real work was classed as delegation, while wholesale delegation produces "a categorical signature." The authors call the weak separability of collaboration the desirable outcome, since students using AI as permitted should not be flagged. In their words, process evidence "flags delegation, not assistance." The abstract also reports that engagement diverges by medium: 93% of AI-inserted characters survived to writers' final documents, against only 14% of accepted code suggestions surviving intact.

Against the nearest notes, this paper supplies an observability layer that the others lack. Can humans detect AI text if machines can measure it? locates the difference in the finished text, which humans cannot perceive, and this paper starts from the same weakness of product-level detection and moves to process evidence. Does AI assistance actually harm the way developers learn? found that delegation-style patterns preserve less learning. The temporal signature offers a way to see delegation-like production in a log, though the excerpt never links the signature to learning. Does ownership framing change how much writers rely on AI? also draws on CoAuthor data, so the two together show that reliance is visible in the same corpus from the framing side and from the timing side.

The excerpt is explicit about its limits. The delegation condition is simulated, and "the separability is partly by construction, since concentrated production defines the simulated behavior." The authors state that the result "establishes sensitivity and observability, not prevalence or field accuracy." The excerpt does not describe how the simulation was built, why code suggestions survive so much less often than inserted text, or whether burst timing predicts any learning outcome. It states only two of the three findings the abstract announces. The defensible reading is narrow: a burst signature makes wholesale delegation observable in process data, and a flag can indicate concentrated production but not misconduct or lost learning.