Can brief critiques help users think critically about AI suggestions?

An unpublished working draft reports a between-subjects study (n=24) testing "provocations: short text prompts whose aim is to induce critical thinking in the user in AI-assisted tasks," attached to an AI-assisted shortlisting task (choosing a shortlist of candidates — charities or movies — from a larger set). Provocations "highlight risks, biases, limitations, and alternatives of GenAI suggestions." On RQ1 ("Do provocations affect users' critical thinking during shortlisting tasks?"), the paper's "qualitative analysis finds impacts on metacognition and critical thinking at multiple levels of Bloom's taxonomy," coded from think-aloud utterances across the levels of analysis, synthesis, evaluation, comprehension, and application, plus a separate metacognition code. But "with some exceptions, our quantitative comparisons are not statistically significant": no significant effect on shortlist diversity, on "the rate of thinkaloud reflections and user interface actions" on aggregate self-reflection, or on trust — though the Provocations condition had "a higher median shortlist diversity and higher rate of think-aloud reflections," and lower median Rank-Biased Overlap similarity between final shortlists (more varied outcomes), without reaching significance.

The provocations themselves were designed from specific prior-literature principles: short texts (Salomon's technique) capped at two sentences "to avoid information overload," phrased as "plain factual warnings" rather than agentic or anthropomorphic language "to avoid any inadvertent confounding effects of agentic presentation," and attached singly to each AI-generated suggestion "at the same time and location as the suggested factor" so the critique stays specific to that suggestion. On RQ2 ("What important dimensions of the system design, task, or user interact with the effectiveness of provocations?"), the paper names five: task urgency, task importance, user expertise, provocation actionability, and user responsibility — identified as the factors that determine whether a given provocation lands as useful friction or gets ignored, though the excerpt does not detail how each dimension operates.

This sits alongside Can interface design reverse citation overload's harm to critical thinking? as a second design intervention aimed at the same problem — but the two differ in where the effect surfaces: the Sidebar study found a quantitatively measurable reversal of a citation-density effect, while provocations produced a qualitative, coded effect on think-aloud utterances that did not reach significance on any of the quantitative measures tested. It extends Does ChatGPT help writers think less deeply about revision? by testing a design response to exactly that metacognitive gap: rather than documenting metacognitive loss under AI assistance, this study inserts a friction designed to restore it. It also bears on Does frequent AI tool use reduce our critical thinking?: where that survey identifies cognitive offloading as the mediating mechanism of decline, provocations are an attempt to interrupt offloading at the point of use.

The excerpt does not establish that provocations caused the qualitative critical-thinking events reported, since the quantitative comparisons meant to confirm an effect were not significant; the paper's own discussion attributes this partly to "the inadequacy of our quantitative measures, insufficient statistical power, and the relative infrequency of conscious critical thinking 'events'" rather than to an absence of effect, but that is the authors' interpretation, not a demonstrated fact. With n=24 on one task type (shortlisting, across two datasets) and the paper itself labeled "Unpublished working draft," the finding supports only that provocations can produce observable critical-thinking and metacognitive utterances in this setting — not that they reliably improve shortlisting outcomes, generalize to other knowledge-work tasks, or outperform other friction designs such as the Sidebar's persistent display.