The way we search for information is undergoing a structural transformation. Where traditional engines returned ranked lists, inviting users to browse, compare, and judge, generative AI produces a single answer and presents it as the final results. Research comparing Google with AI-powered engines finds that these systems drastically diverge in which sources they surface and how they define what counts as authoritative (Kirsten et al., 2025).
The divergence is not neutral. Audits of ChatGPT, Bing Chat, and Perplexity show bias towards US commercial media framing, with academic and non-Western outlets underrepresented (Li & Sinnamon, 2024).
An additional risk of AI summaries is the tendency to strip out the uncertainty language present in their sources, Formulations such as "evidence suggests" are replaced with more assertive ones, presenting answers as more definitive than the underlying research warrants. Different systems make different editorial choices in doing so, yet users encounter only a single version of events, with no means of knowing what was omitted or how another system might have framed the same question. (Huang et al., 2025).
The damages resulting from this new information funnels are already measurable. Tracking 161,382 Wikipedia article-language pairs, Khosravi & Yoganarasimhan (2026) find AI Overviews reduce daily traffic by roughly 15%, reallocating attention away from the sources AI itself depends on for training.
What is much harder to see however is what researchers lose. Generative search breaks the feedback loop that allows systems to improve, end-to-end synthesis doesn’t allow users to understand what is the best path for information retrieval (Dai et al., 2025). As Marchionini argued before the current paradigm, meaningful research is not about lookup, it is exploratory search. It is an iterative process through which understanding is built, not retrieved (Marchionini, 2006). AI answers are optimised for the former. The cost of worsening in the latter will have to be accounted for.