Take the doubling at face value and the review budget stops being a rounding error. More than 124,917 of the 249,834 conversations carried consequential work, on Anthropic's summary of the finding [13], and high-stake exchanges run close to twice as long as ephemeral ones [3]. Weight that half by length and it accounts for roughly two thirds of all the text a reviewer would have to read [17]. Oversight cost does not scale with the count of flagged cases. It scales with the words inside them.
The sizing is not settled, because the two sources do not say the same thing. Anthropic's thread reports that over half of the conversations involved consequential tasks, meaning work that affects other people or is hard to undo [9]. The paper's own text applies its criticality categories to the conversations analysed that contained actionable tasks [10], which is a smaller base. Whether the consequential slice is a majority of all traffic or a majority of task-bearing traffic changes the headcount of anyone planning to inspect it.
Friction will not do the routing. It appears in roughly half of all conversations, is often productive, and users apply active strategies to recover from it [6]. Two sets that each cover about half a population need not overlap at all [18], so nothing in the reported figures licenses the assumption that breakdowns cluster in the consequential half. A breakdown detector, on this evidence, finds a large and mostly benign population.
Agency is reported the same way. Human-led collaboration dominates, but the paper is explicit that users construct their level of agency through how they prompt and iterate [4]. That makes "human in the loop" a description of user habit rather than a property of the deployment. Active teaching turns up in 67 percent of conversations, about 167,000 of them [5][14], and the paper says the benefits are unevenly realised [5]. Majority behaviour, uneven payoff.
The access route deserves the same scepticism as the findings. Anthropic says this is the first time external researchers have been given a way to study AI's impacts using real, privacy-preserved Claude usage data, work previously possible only inside AI labs [7]. The Stanford group behind the paper, Shao et al. [12], along with Oxford's Human Information Processing Lab and METR, worked from aggregated outputs of conversations in a two-month window, April to May 2026 [8]. Aggregated outputs from a vendor's own pipeline, over two months of one product, is a narrower instrument than the conversation count makes it sound.
Placement matters more than the totals. High criticality delegation concentrates in advisory and professional domains [2], the domains that already carry duties of care and records obligations. An escalation rule triggered on stakes, in that setting, escalates the median case rather than the exception.
