Does weeks of everyday AI-answer use make people check sources less?

Has this been done?

Partly done already: related work covers some of this question (first-pass screen of 12 October 2026, checked again on 13 October 2026).

Randomized field experiments that hide or show AI answers in search exist for seven days (Wang et al.) and in an extension-based study of AI Overviews (Agarwal and Sen, 2026), but measure clicks and user experience. An 8-day diary field study compared ChatGPT with Google for learning (Mittal et al.), a four-week randomized chatbot experiment measured loneliness and emotional dependence rather than information verification (Fang et al.), a one-month study of 67 people found unaided misinformation detection declined after AI dialogue (Rani et al.), and year-long or six-month panels (IWM Tübingen; Huemmer et al.) are observational. No randomized study of several months that measures source-checking behaviour and detection of wrong answers in everyday search was found.

How strong this evidence is

• Who searched and checked: Claude (SocioLens Project Lead) with AI research agents, by web search; not a person and not a domain expert; an independent AI reviewer with fresh context (not a person), 12 October 2026: every nearest-work citation checked against its source page; corrections applied. Every check so far was done by AI; no person or domain expert has checked it yet.
• What was read: none of the 10 nearest works below is recorded as judged from an abstract only, and 1 has details that could not be confirmed; it is marked. Not every entry says how much of the work was read.
• Where the search could not reach: Screened by web search on the date shown. ScienceDirect and Taylor & Francis article pages could not be opened, so two relevant panel studies (the IWM article and a two-wave cross-cultural panel on trust in AI news in Behaviour & Information Technology) were seen only through listings. Google Scholar was not searched. Absence of a hit is not proof that the work does not exist. Literature check of 2026-10-13: The SSRN page of Agarwal and Sen could not be opened, so its sample, duration and author list come from the TSE seminar page and search listings. ScienceDirect and Taylor & Francis articles (e.g. a 2026 Journal of Broadcasting & Electronic Media paper on chatbot fact-checking use) were seen only as listings. Google Scholar was not searched. Absence of a hit is not proof that the work does not exist.
• Before building a project on it, run your own search in a scholarly database such as Scopus, Web of Science, Google Scholar or SSRN.

Nearest work found

• Wang, Gleason, Bart, Wilson & Metaxa, "AI in Search Reduces Publisher Referrals Without Improving User Experience: Experimental Evidence," arXiv:2608.18352 (18 Aug 2026)
  Preregistered US field experiment, 1,100 participants, three-day baseline then seven days in condition; measures click-through, user experience and trust. Its limitations note that longer exposure periods may show different effects; it does not track verification habits over time.
• Fang, Liu, Danry, Lee, Chan, Pataranutaporn, Maes, Phang, Lampe, Ahmad & Agarwal, "How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use: A Longitudinal Randomized Controlled Study," arXiv:2503.17473 (v2, Oct 2025)
  Four-week randomized study of 981 participants comparing chatbot modes (text, neutral voice, engaging voice) and conversation types; outcomes are loneliness, socializing with people, emotional dependence and problematic use. The abstract does not mention measuring information verification or reliance for knowledge.
• Ehrhardt, Ivan & Klein, "From perception to adoption: A longitudinal study of trust and delegation in everyday AI use," Computers in Human Behavior: Artificial Humans 8, 100325 (May 2026), from the IWM Tübingen project "A longitudinal study on the perceptions and dynamics of human-AI interaction" (August 2024 to April 2026)
  The project page describes a six-wave, one-year observational panel measuring trust, reliance, credibility and willingness to delegate tasks to AI. Bibliographic details were confirmed through the DOI record and the project page (https://iwm-tuebingen.de/en/research/projects/hmc_ai_trends); the article text was not reachable. Survey-based and not randomized; no logged search behaviour.
• Landerberg, Flatmo & Said, "Trust as a Situated User State in Social LLM-Based Chatbots: A Longitudinal Study of Snapchat's My AI," arXiv:2604.22417 (24 Apr 2026)
  Four-week qualitative survey study of 27 users describing how trust and reliance shift over time. Small, qualitative and about a social chatbot, not search.
• Rani, Danry, Liang, Lippman & Maes, "Dialogues with AI Reduce Beliefs in Misinformation but Build No Lasting Discernment Skills," arXiv:2510.01537 (v2 13 Mar 2026; accepted at CHI 2026)
  67 participants over one month judged news items, discussed them with an AI, then judged unseen items alone; accuracy rose about 21% with AI help, but unaided accuracy on new items fell 15.3% by week 4. Directly relevant to the error-detection outcome over time, but small, task-based rather than everyday search, and the abstract does not describe a no-AI control arm.
• Mittal, Blodgett & Liao, "Learning by Chatting? Investigating the Impact of Generative AI on Information Seeking and Learning," arXiv:2606.11669 (v1 10 Jun 2026; v2 7 Aug 2026)
  Between-subjects field study with daily diaries over 8 days, ChatGPT versus Google Search; ChatGPT users delegated source selection, felt less control and learned less. The abstract does not say assignment was randomized, does not report verification outcomes, and runs for days, not months.
• Huemmer, Durner, Shyiramunda & Cummings-Koether, "AI, Metacognition, and the Verification Bottleneck: A Three-Wave Longitudinal Study of Human Problem-Solving," arXiv:2601.17055 (21 Jan 2026)
  Six-month, three-wave convenience cohort in an academic setting solving problems with generative AI; reliance grew and the gap between believed and actual accuracy widened. Not randomized, no control group, not search; the authors say causal claims need RCTs.
• Ahmed, Leroy, Sachdeva, Harber, Rains, Youn & Barai, "Trust in Generative AI for Health Information Consumption and the Effect of Learned Dependency: An Experimental Investigation," arXiv:2606.20605 (v3 28 Jun 2026)
  Two randomized 2x2 experiments (n = 901) on correct versus incorrect AI health information and highlighting; users high in learned dependency trusted incorrect outputs more, and highlighting had no effect. Single session; dependency is measured, not induced over time.
• Agarwal & Sen, "The Impact of Google AI Overviews on Publisher Traffic and User Experience: Evidence from a Field Experiment," SSRN abstract 6513059 (2026); presented at the TSE Economics of Platforms Seminar, 2 Jul 2026 Some details not confirmed
  Users randomized via a custom Chrome extension to Google Search with or without AI Overviews: outbound organic clicks fell 39.8% and zero-click searches rose 34.5% when an overview appeared, with no gain in perceived search quality; an exploratory AI Mode arm suggested larger traffic losses. Read from the TSE seminar page (which lists only Agarwal; the SSRN listing names Agarwal and Sen); SSRN page could not be opened, so sample size and duration are unconfirmed. Outcomes are traffic and user experience, not trust, verification habits or error detection.
• Zhi, Kumar & Lee, "Investigating the Effects of LLM Use on Critical Thinking Under Time Constraints: Access Timing and Time Availability," arXiv:2603.08849 (9 Mar 2026)
  Between-subjects experiment (n = 393) varying when LLM access is available and time pressure on a critical-thinking decision task. Single session, lab task, no verification or search-log outcomes.

Literature check

Stands as first published. Across seven new queries (preprint, journal, forward and German-language searches), no randomized study of several weeks or months was found that measures source-checking behaviour and detection of wrong answers in everyday search. Randomized extension-based field experiments that hide AI answers exist but are short and measure clicks and user experience (Wang et al., seven days; Agarwal and Sen, seminar July 2026, duration not stated), a one-month study of 67 people found unaided misinformation detection fell after AI dialogue (Rani et al.), an 8-day diary study compared ChatGPT with Google for learning (Mittal et al.), and longer panels are observational (IWM Tübingen; Huemmer et al.).

Searched: web (extended and standard); arxiv.org; osf.io; ssrn.com; nber.org; dl.acm.org; sciencedirect.com; tandfonline.com; journals.sagepub.com; academic.oup.com; German-language web; tse-fr.eu. Checked by Claude (SocioLens Project Lead) with AI research agents, by web search; not a person and not a domain expert.