Can interface design reverse citation overload's harm to critical thinking?

The paper reports a controlled between-subjects experiment (N=372) on a conversational AI writing task, comparing four citation-presentation interfaces: Collapsible, Hover Card, Footer, and Aligned Sidebar. Its central finding is "the discovery of that effect design can reverse the negative impact of information overload on information evaluation and critical thinking." In the Collapsible condition, critical-thinking scores declined "as the AI provided more citations." The Aligned Sidebar condition reversed this: as citation density rose, Sidebar users "showed significantly improved scores in Synthesizing Multiple Sources and Source-Related Critical Thinking." A second, related result is a trade-off: the Hover Card interface supported a "seamless Cite→Write loop (Prob. = 0.393)" that let users "verify specific claims on-demand without breaking their drafting context," raising scores for Evaluating Evidence Strength "at the cost of synthesis," while the Sidebar's more disruptive "Write→Explore loop" produced "lower perceived usefulness ratings... if information density is low" but "ultimately supported better holistic argumentation."

The authors frame the four interfaces as crossing two design dimensions: visibility ("the explicit display of source details") and accessibility ("the proximity of source details to the in-text citations"). Collapsible is low/low, Footer is high visibility/low accessibility, Hover Card is low visibility/high accessibility, and Sidebar is high on both. Their mechanism for the density reversal is that the Sidebar functions as "an external working memory": by moving sources "from the ephemeral, linear flow of the chat" into "a persistent, spatial layout," it lets users "offload the cognitive burden of tracking evidence" as citations accumulate. For the flow/verification trade-off, they draw on prior work on "cognitive forcing": the Sidebar's friction "reduces over-reliance on AI while simultaneously increasing perceived workload," while the Hover Card's low friction preserves flow but removes the disruption that forces synthesis.

This sits beside Can readers tell truth from fabrication without evidence signals?, which also finds that a well-designed evidence display measurably improves critical engagement with AI-generated text. The two differ in what they manipulate: that study showed participants an idealized, pre-computed density score, while this paper manipulates only where and how existing citations are placed in the interface, and finds the effect depends on citation density rather than appearing from the display alone. It also parallels Do AI writing tools improve online discussion or degrade it?, whose "no single AI tool enhances both producer and consumer experiences" is the same shape of trade-off this paper finds between flow and verification — two studies independently finding that AI-interaction design features which help one valued outcome tend to cost another. It also echoes the design-sensitivity point in Which clarifying questions actually improve user satisfaction?: there, the same underlying feature (a clarifying question) helps or wastes effort depending on its design; here, the same underlying feature (a citation) raises or lowers critical thinking depending on how its interface is built.

The excerpt does not describe the "automated critical thinking assessment" instrument in enough detail to know what exactly the Synthesizing Multiple Sources and Source-Related Critical Thinking scores capture, so the size and nature of the reversal can't be independently checked from what's quoted here. The study is a single writing task on one AI backend (Perplexity's sonar-pro, "temperature of 0.7," US-located users), run once rather than longitudinally, so it cannot show whether the Sidebar's benefit holds across tasks or over repeated use, or whether users habituate to its friction. The authors themselves stop short of recommending one interface: they conclude there is "no 'one-size-fits-all' solution" and propose adaptive interfaces that switch by task phase as future work, not something this experiment tested. The defensible reading is narrower than "add more source transparency": this excerpt supports that citation interface design changes the direction of information density's effect on critical thinking, in this task, not that any single interface is best overall.