Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Every work request carries choices nobody states: which outcome counts, over what horizon, for whom, by what measure, and whether the task is worth doing at all. This paper measures how often AI assistants surface one of those choices while still doing the work they were asked to do. Across 2,496 generated responses and 4,992 judgements on 60 items spanning six framing dimensions, unprompted surfacing rose from 5.2% to 23.3% between two model generations three months apart (paired difference +18.1 pp, 95% CI [+11.5, +25.4], sign test p = 1.1 × 10^-6; cluster-robust logistic OR 5.5, p = 2.4 × 10^-8). The measure requires both naming the framing choice and completing the request. The second finding has consequences for evaluation design. An arm whose prompt explicitly invited critique reached an asserting level in 52.7% of responses and surfaced in 0.0%, because completion fell to zero. Framing engagement and task completion move independently, so an evaluation scoring critique alone would rank the worst intervention as the best. Data, code and figures are deposited separately at DOI 10.5281/zenodo.22847076. Every number in the paper is generated by a single statistics module and checked by a build gate that refuses any figure not resolving to it.
No comments yet — start the discussion below.