Comment on “Beyond disruption and invisibility: Interactional continuity in everyday AI use in India”

Selected via the production queue; critique generated by the AGISS engine.

This is an honest, modest qualitative study (28 interviews in one Bangalore office) about how people fold AI into everyday tasks. The full text fixes most concerns a reviewer might raise from the abstract alone, because the authors repeatedly admit the sample is tiny, from one site, and not generalizable. The real remaining problems are small numbers issues. The most clear-cut: the paper's key statistic — that embedded AI is heavily used but almost never called "AI" — is reported as "9% (n = 2) of respondents," but 2 out of 28 respondents is 7.1%, not 9%; the 9% only works if you divide by 23 (the subset of users), so the stated base ("respondents") doesn't match the number. A related percentage for embedded GenAI use is also computed against 28 while the paper's own table says only 23 people used it. There's also a proposition phrased as a statistical "correlates more strongly than" claim that the descriptive design can't actually test, and a "saturation" claim that sounds stronger than a 2-day, take-whoever-is-available sample warrants. None of this breaks the main argument; they are reporting slips around an otherwise carefully limited study.

The central naming statistic is internally inconsistent: 'explicit naming occurred in 9% (n = 2) of respondents,' but with 28 respondents 2/28 = 7.1%, not 9%. The figure 9% is only reproducible against a base of 23 (embedded-AI users in Table 1), so the denominator silently shifts from 'respondents' (28) to 'users' (23) inside the paper's flagship evidence for naming–use decoupling. The percentage and its stated population are mutually contradictory.

Embedded-GenAI usage percentages are computed against all 28 respondents (67.9% = 19/28, 60.7% = 17/28, 39.3% = 11/28), yet Table 1 reports 'Embedded GenAI Users (n = 23).' If only 23 respondents used embedded GenAI, the purpose percentages should be based on 23 (e.g., 19/23 = 82.6%), not 28. The prose denominator and the table's own user count are inconsistent, making the reported uptake rates ambiguous.

Proposition 1 is phrased as a comparative correlational claim ('will correlate more strongly than'), implying an inferential test of association, but the design — 28 single-site self-reports, single-coder coding, no inter-coder reliability, no correlation statistic computed — cannot support a 'more strongly than' inference. The proposition's quantitative form outruns what a small descriptive qualitative corpus can establish.

The paper invokes thematic 'saturation' ('additional sessions yielded no new themes') to justify stopping at 28, but interviews were a single workplace over 2 days with recruitment that 'reflected availability.' Saturation language implies theoretical sufficiency, which a short, convenience/availability-based single-site sample cannot credibly establish; the claim overstates the evidentiary basis for the stopping rule.

The paper is unusually candid about its own limits and does not pretend to statistical generalizability. It explicitly frames results as 'tendencies and not robust subgroup claims,' states findings 'are not generalizable,' discloses single-coder coding with no inter-coder statistics and no precision/recall metrics (with reasons), and presents its propositions as analytic guides rather than statistical hypotheses. Most reported percentages are arithmetically correct against a 28-respondent base, and the central contribution is a conceptual mechanism illustrated through rich episode-level interview accounts, appropriately scoped to what 28 interviews can support.