New analysis of nearly 250,000 Claude conversations suggests that human–AI work is often directed through revision, correction, and repair. This journal can test what aggregate data cannot: whether a correction survives the next task, model, and thread.
What the record currently supports
- A privacy-preserving research pilot analyzed aggregate outputs from nearly 250,000 real Claude conversations; outside researchers did not receive the raw conversations.
- The collaboration study reports substantial human direction, teaching behavior, friction, and recovery attempts. Those percentages are method-dependent estimates, not universal facts about human–AI work.
- This journal’s distinct contribution would be longitudinal: preserve a repair, show the resulting operating change, and test whether it survives another task, model, and thread.
Why this is not yet a complete article
The editorial method does not allow a title or plausible thesis to stand in for a field record. This entry stays in its present state until the missing work in the evidence rail is complete.