The design deserves a moment, because it is what lifts this above anecdote. Two topics, two styles of phrasing for each, one week of collection in January 2026, and a comparison that holds the subject fixed while varying only how the question is asked [3][4][5]. When the summaries diverge under that arrangement, the wording is the plausible cause rather than the topic [6]. The queries came from open forums and Google Trends rather than from the authors' imagination [3], which matters more than it sounds: a leading query invented at a desk can be leading in a way no real person would ever type.
The asymmetry between the two topics is the part I would put weight on. Chemtrails has been circulating since the 1990s, and after researchers criticised Google in 2015 for giving the belief visibility, the company demoted the problematic results [15][16]. The overviews now barely mention it [8]. The 15-minute city, a planning term recast by conspiracists as a scheme to monitor residents through technology [17], carries no such remediation history, and it is there that 12.5% of the analysed summaries set false claims beside true ones as a debate with two legitimate sides [12].
The quoted example shows the mechanism, and it is subtler than a summary asserting something untrue. The overview explains how a 15-minute city and a smart city overlap, then adds that "the control aspect is where critics focus their backlash, blurring the lines between convenience and surveillance" [14]. There is no checkable false statement in that clause. It is a framing decision, inherited from the pages a control-flavoured query pulls in, and delivered with the authority of the box at the top of the page [20].
The thing this write-up does not tell you is the denominator. 12.5% is exactly one in eight, so the analysed set is eight overviews, or sixteen, or twenty-four, and the summary article does not say which [22]. At the low end, one coded summary is worth the entire 12.5 points [23]. Nor does the design reach readers: it collects search output, not participants, so the effect on what anyone came away believing is unmeasured [24].
Where debunking did appear, and it usually did [9], it often stopped at the word misinformation [10]. The authors' argument is that a label is not an argument, and that people arriving at these topics need the factual and logical case [11]. The pre-AI baseline was not clean either: the same simulated conspiratorial searches returned YouTube videos and Facebook posts promoting the beliefs [18]. What changed is position, since the summary is where reading now tends to stop [20]. As an aside worth its own study, the overviews leaned more commercial in the sources they surfaced [19].
My read, with its condition attached: this looks like a retrieval-and-phrasing failure concentrated on narratives young enough to have no moderation history [8][12][16], which means every remedy is topic-shaped and arrives after the narrative it is meant to answer.
