Abstract

A claim that cannot be traced to its source is hard to check, and both people and language models lose track of sources. This article reviews four bodies of evidence. First, memory can keep a claim while losing where it came from: judgments of a memory’s source are attributions subject to error, a message from a discounted source can grow more persuasive over time, and repetition makes statements feel true even when people know better (Johnson et al., 1993; Kumkale & Albarracín, 2004; Fazio et al., 2015). Second, people judge online credibility mostly through heuristics, whereas professional fact checkers leave a page to find out who stands behind it, a strategy that students can be taught (Metzger et al., 2010; Wineburg & McGrew, 2019; Wineburg et al., 2022). Third, language models are prone to hallucinated text, including fabricated references and cited sources that do not fully support the statements they accompany (Ji et al., 2023; Walters & Wilder, 2023; Wu et al., 2025). Fourth, belief in false headlines is associated more with a lack of reflection than with partisan bias, shifting attention to accuracy improves what people share, and clarifying questions help when a search request is ambiguous (Pennycook & Rand, 2019; Pennycook et al., 2021; Zamani et al., 2020). On this evidence, Mars deep research first shows a research proposal card for you to confirm, shows each research helper’s progress as a live card, and gives every cited web link its page’s real title; Mars reads public pages as a signed-out visitor, tells you when it could not read a page instead of guessing, and aligns with you through up to five questions before acting. We also state the limits: this evidence supports the design direction and is not an experimental test of Mars itself, and making the publisher more visible does not by itself reduce susceptibility to misinformation (Dias et al., 2020).

Keywords: source monitoring · misinformation · lateral reading · credibility heuristics · hallucination · citation accuracy · accuracy prompts · clarifying questions · deep research · AI research assistant

1. Introduction

Misinformation is persistent. A review of the cognitive research describes retractions as often ineffective in updating memory and notes that attempts to retract misinformation can even backfire (Lewandowsky et al., 2012). A meta-analysis of 32 studies (N = 6,527) found that, on average, correction does not entirely eliminate the effect of misinformation, and that corrections work less well when the misinformation was attributed to a credible source, had been repeated, or was corrected after a delay (Walter & Tukachinsky, 2020). A newer source of the problem has appeared alongside: text from deep-learning language generators is fluent and coherent, yet prone to hallucinating unintended content (Ji et al., 2023).

Checking sources is therefore more than a formality. This article asks why people lose track of sources, how skilled readers check them, why citations written by language models need the same scrutiny, and what that implies for the design of an AI research assistant.

Method: this is a narrative literature review covering cognitive psychology, communication research, information literacy and natural language processing. We favoured peer-reviewed work accessible through a DOI, preferred meta-analyses, preregistered and field experiments and large-scale evaluations, and included a replication and studies that limit popular interventions. Every factual statement carries its source, and the full reference list with original links follows the article.

2. Memory keeps the claim and loses the source

The source monitoring framework holds that judging where a memory came from is an attribution based on qualities of the remembered experience; these judgments use flexible criteria and are subject to error and disruption. Phenomena it accounts for include misattributed familiarity, cryptomnesia and the incorporation of fiction into fact (Johnson et al., 1993).

Losing the source changes how persuasive a message is. In the sleeper effect, people who receive a message together with a discounting cue, such as a noncredible source, are less persuaded immediately after exposure than they are later; a meta-analysis found this increase over time was larger when the message arguments and the cue both had a strong initial impact, and when the discounting cue came after the message (Kumkale & Albarracín, 2004). Repetition adds to the risk: repeated statements are easier to process and are judged more truthful, and this illusory truth effect occurred even when participants knew better (Fazio et al., 2015).

A claim is easier to check when its source stays attached to it rather than left to memory. On this basis, every web link cited in a Mars deep research report carries its page’s real title, on the web and on iOS, so each source in the report can be recognised by what it is (Johnson et al., 1993; Kumkale & Albarracín, 2004).

3. How people judge credibility, and how experts check it

Most people do not evaluate online information effortfully. In focus groups with 109 participants, most users relied on others to assess credibility, often through group-based tools, and routinely invoked cognitive heuristics rather than processing information systematically (Metzger et al., 2010).

Skilled checking looks different. Wineburg and McGrew asked 10 PhD historians, 10 professional fact checkers and 25 Stanford University undergraduates to think aloud while evaluating live websites. Historians and students often fell for easily manipulated features such as official-looking logos and domain names, and read vertically, staying within a site; fact checkers read laterally, leaving the site after a quick scan and opening new browser tabs to judge the original site, and reached more warranted conclusions in a fraction of the time (Wineburg & McGrew, 2019). The strategy can be taught: in a district-wide field study, high school students whose teachers gave six one-hour lessons derived from fact checkers’ strategies (n = 271) grew significantly more in their ability to judge the credibility of digital content than peers in regular classrooms (n = 228) (Wineburg et al., 2022).

Lateral reading depends on knowing what a source is and being able to open the original. On this basis, Mars deep research shows each cited source by its real page title, and for member pages or messages behind a login Mars points you to the original link: a new tab on the web, or the in-app browser on iOS with one tap to Safari (Wineburg & McGrew, 2019; Wineburg et al., 2022).

4. Language models also lose and invent sources

A survey of hallucination in natural language generation describes deep-learning text generators as prone to hallucinating unintended text, and reviews metrics and mitigation methods across abstractive summarization, dialogue generation, generative question answering, data-to-text generation and machine translation (Ji et al., 2023).

Citations are a visible case. When ChatGPT-3.5 and ChatGPT-4 wrote short literature reviews on 42 topics, 55% of the GPT-3.5 citations and 18% of the GPT-4 citations were fabricated, and 43% and 24% of the real citations, respectively, contained substantive errors (Walters & Wilder, 2023). Given the inclusion criteria of 11 systematic reviews, GPT-3.5, GPT-4 and Bard produced hallucinated references at rates of 39.6%, 28.6% and 91.4%, and the authors concluded that any references generated by such models warrant thorough validation (Chelli et al., 2024). A real reference is not enough either: across seven models, 800 medical questions and 58,000 statement–source pairs, between 50% and 90% of responses were not fully supported, and sometimes contradicted, by the sources they cited; even for GPT-4o with web search, about 30% of individual statements were unsupported (Wu et al., 2025).

A model’s citation therefore has to be checkable, and a page the model could not read should not be filled in by guesswork. On this basis, Mars reads public pages as a signed-out visitor; when a page asks for a login or a verification check, it tells you it could not read it instead of guessing, and the final research report skips pages that show only a security check (Walters & Wilder, 2023; Chelli et al., 2024; Wu et al., 2025).

5. A pause to reflect, and a question before acting

Why do people accept false claims? In two studies with 3,446 Mechanical Turk workers, performance on the Cognitive Reflection Test was negatively correlated with the perceived accuracy of fake news and positively correlated with the ability to tell fake from real news, even for headlines aligned with participants’ political ideology; the authors conclude that susceptibility to fake news is driven more by lazy thinking than by partisan bias (Pennycook & Rand, 2019).

Attention, more than values, is often what is missing. The veracity of headlines had little effect on sharing intentions despite a large effect on accuracy judgments; in four survey experiments and a field experiment on Twitter, subtly shifting attention to accuracy increased the quality of news people subsequently shared (Pennycook et al., 2021). Ambiguity at the start of a search raises a related problem: queries are often short and their intent unclear, and user studies and large-scale online experiments found asking clarifying questions useful in web search; the authors also generate candidate answers for each question so that users can select one (Zamani et al., 2020).

On this basis, Mars builds a pause into research. When you ask for an in-depth look at a company, a person or an industry, Mars first shows a research proposal card for you to confirm; when it needs your call, it opens a question panel above the composer with up to five numbered questions, where you pick an option or type your own answer (Pennycook et al., 2021; Zamani et al., 2020). After you confirm, each research helper’s progress is shown as a live card with its current steps, not a spinning black box.

6. From evidence to design

These findings point to one conclusion: a research assistant that people can check must keep sources visible, make the original easy to open, say what it could not read, and settle the direction before it starts. Mars meets each requirement as follows:

1. Memory separates claims from their sources → every cited link carries its real page titleEvery web link in a Mars deep research report carries its page’s real title, on the web and on iOS. The source of each statement stays visible in the report instead of depending on a memory that may later misattribute it (Johnson et al., 1993; Kumkale & Albarracín, 2004; Fazio et al., 2015).
2. Experts check a source by leaving it → the original is easy to openBecause each source appears under its real title, you can open it and read laterally to judge who stands behind it. For member pages or messages, open the original link: a new tab on the web, the in-app browser on iOS with one tap to Safari (Metzger et al., 2010; Wineburg & McGrew, 2019; Wineburg et al., 2022).
3. Models fabricate and misattribute citations → no guessing at pages Mars could not readMars reads what any signed-out visitor sees, with no account, cookie, password or saved login. When a page asks for a login or a verification check, it tells you it could not read it instead of guessing, and the report skips pages that show only a security check (Ji et al., 2023; Walters & Wilder, 2023; Chelli et al., 2024; Wu et al., 2025).
4. Accuracy needs attention and intent needs clarifying → confirm the plan, answer up to five questionsMars first shows a research proposal card for you to confirm, opens a question panel with up to five numbered questions when it needs your call, and shows each research helper’s progress as a live card. You set the direction and can see the work (Pennycook & Rand, 2019; Pennycook et al., 2021; Zamani et al., 2020).

7. Limitations

First, the studies cited here test how people and language models handle sources, not Mars itself. They support our design direction but do not show that Mars reduces errors in a report; the product’s effect needs its own evaluation.

Second, the samples are specific: online survey workers (Pennycook & Rand, 2019; Pennycook et al., 2020), 45 historians, fact checkers and Stanford undergraduates (Wineburg & McGrew, 2019), high school students in one urban district (Wineburg et al., 2022), and medical questions and literature (Chelli et al., 2024; Wu et al., 2025). Error rates of language models belong to the model versions tested and differed sharply between them (Walters & Wilder, 2023), so they describe a risk, not a fixed rate.

Third, popular interventions are weaker than they sound. A preregistered direct replication of an accuracy-prompt study on COVID-19 headlines was unsuccessful in its first stage (n = 701) and found a small but significant effect only in the pooled sample (N = 1,583) (Roozenbeek et al., 2021). General warnings and fact-check tags reduced the perceived accuracy of false headlines only modestly, and a general warning also lowered belief in true headlines (Clayton et al., 2020); tagging only some false headlines made untagged ones seem more accurate (Pennycook et al., 2020); and making publishers more visible was ineffective, perhaps even counterproductive (Dias et al., 2020).

Fourth, a source label is not a verdict. A real page title tells you what a source is, not whether its content is right. That is why Mars keeps the original link within reach rather than asking you to trust a label, and the judgment of what to believe stays with the reader.

8. Conclusion

Sources get lost in two places: in human memory, which can keep a claim while forgetting where it came from, and in generated text, which can cite works that do not exist or do not support it. The research reviewed here points to visible sources, opening the original to check it, and a pause to consider accuracy and intent as the matching remedies, while labels alone have limited effect. Mars deep research, with its research proposal card, live research helper cards, a real page title for every cited link, signed-out reading and up to five questions before acting, is designed on this evidence.

References

Listed in APA 7th edition style, each with its original DOI link.

1. Chelli, M., Descamps, J., Lavoué, V., Trojani, C., Azar, M., Deckert, M., Raynier, J.-L., Clowez, G., Boileau, P., & Ruetsch-Chelli, C. (2024). Hallucination rates and reference accuracy of ChatGPT and Bard for systematic reviews: Comparative analysis. Journal of Medical Internet Research, 26, e53164. https://doi.org/10.2196/53164
2. Clayton, K., Blair, S., Busam, J. A., Forstner, S., Glance, J., Green, G., Kawata, A., Kovvuri, A., Martin, J., Morgan, E., Sandhu, M., Sang, R., Scholz-Bright, R., Welch, A. T., Wolff, A. G., Zhou, A., & Nyhan, B. (2020). Real solutions for fake news? Measuring the effectiveness of general warnings and fact-check tags in reducing belief in false stories on social media. Political Behavior, 42(4), 1073–1095. https://doi.org/10.1007/s11109-019-09533-0
3. Dias, N., Pennycook, G., & Rand, D. G. (2020). Emphasizing publishers does not effectively reduce susceptibility to misinformation on social media. Harvard Kennedy School Misinformation Review, 1(1). https://doi.org/10.37016/mr-2020-001
4. Fazio, L. K., Brashier, N. M., Payne, B. K., & Marsh, E. J. (2015). Knowledge does not protect against illusory truth. Journal of Experimental Psychology: General, 144(5), 993–1002. https://doi.org/10.1037/xge0000098
5. Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y. J., Madotto, A., & Fung, P. (2023). Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12), 1–38. https://doi.org/10.1145/3571730
6. Johnson, M. K., Hashtroudi, S., & Lindsay, D. S. (1993). Source monitoring. Psychological Bulletin, 114(1), 3–28. https://doi.org/10.1037/0033-2909.114.1.3
7. Kumkale, G. T., & Albarracín, D. (2004). The sleeper effect in persuasion: A meta-analytic review. Psychological Bulletin, 130(1), 143–172. https://doi.org/10.1037/0033-2909.130.1.143
8. Lewandowsky, S., Ecker, U. K. H., Seifert, C. M., Schwarz, N., & Cook, J. (2012). Misinformation and its correction: Continued influence and successful debiasing. Psychological Science in the Public Interest, 13(3), 106–131. https://doi.org/10.1177/1529100612451018
9. Metzger, M. J., Flanagin, A. J., & Medders, R. B. (2010). Social and heuristic approaches to credibility evaluation online. Journal of Communication, 60(3), 413–439. https://doi.org/10.1111/j.1460-2466.2010.01488.x
10. Pennycook, G., & Rand, D. G. (2019). Lazy, not biased: Susceptibility to partisan fake news is better explained by lack of reasoning than by motivated reasoning. Cognition, 188, 39–50. https://doi.org/10.1016/j.cognition.2018.06.011
11. Pennycook, G., Bear, A., Collins, E. T., & Rand, D. G. (2020). The implied truth effect: Attaching warnings to a subset of fake news headlines increases perceived accuracy of headlines without warnings. Management Science, 66(11), 4944–4957. https://doi.org/10.1287/mnsc.2019.3478
12. Pennycook, G., Epstein, Z., Mosleh, M., Arechar, A. A., Eckles, D., & Rand, D. G. (2021). Shifting attention to accuracy can reduce misinformation online. Nature, 592(7855), 590–595. https://doi.org/10.1038/s41586-021-03344-2
13. Roozenbeek, J., Freeman, A. L. J., & van der Linden, S. (2021). How accurate are accuracy-nudge interventions? A preregistered direct replication of Pennycook et al. (2020). Psychological Science, 32(7), 1169–1178. https://doi.org/10.1177/09567976211024535
14. Walter, N., & Tukachinsky, R. (2020). A meta-analytic examination of the continued influence of misinformation in the face of correction: How powerful is it, why does it happen, and how to stop it?. Communication Research, 47(2), 155–177. https://doi.org/10.1177/0093650219854600
15. Walters, W. H., & Wilder, E. I. (2023). Fabrication and errors in the bibliographic citations generated by ChatGPT. Scientific Reports, 13, 14045. https://doi.org/10.1038/s41598-023-41032-5
16. Wineburg, S., & McGrew, S. (2019). Lateral reading and the nature of expertise: Reading less and learning more when evaluating digital information. Teachers College Record, 121(11), 1–40. https://doi.org/10.1177/016146811912101102
17. Wineburg, S., Breakstone, J., McGrew, S., Smith, M. D., & Ortega, T. (2022). Lateral reading on the open Internet: A district-wide field study in high school government classes. Journal of Educational Psychology, 114(5), 893–909. https://doi.org/10.1037/edu0000740
18. Wu, K., Wu, E., Wei, K., Zhang, A., Casasola, A., Nguyen, T. P., Riantawan, S., Shi, P., Ho, D. E., & Zou, J. (2025). An automated framework for assessing how well LLMs cite relevant medical references. Nature Communications, 16, 3615. https://doi.org/10.1038/s41467-025-58551-6
19. Zamani, H., Dumais, S., Craswell, N., Bennett, P. N., & Lueck, G. (2020). Generating clarifying questions for information retrieval. Proceedings of The Web Conference 2020, 418–428. https://doi.org/10.1145/3366423.3380126

Cite this article

Mars Research. (2026). Why sources must be checked: source memory, credibility and model-made citations. MarsDataAI. https://marsdata.ai/research/checking-sources-in-research