Verifying Information in the Age of AI: Lateral Reading and Spotting Hallucinations
Learners today face huge amounts of information, some of it accurate and some of it misleading or invented. With language models now producing convincing text, verification has become a core skill, not only for journalists and researchers. This article offers practical tools, based on published research, for evaluating digital information and AI outputs.
Information disorder: three types
A 2017 Council of Europe report by Claire Wardle and Hossein Derakhshan distinguishes three types of "information disorder":
- Misinformation: false information shared without intent to harm.
- Disinformation: false information shared deliberately to mislead or cause harm.
- Malinformation: true information shared with intent to harm, such as exposing someone's private life.
This distinction helps in choosing the right response: sometimes a correction is enough, while deliberate disinformation calls for a deeper look at the source and its motives.
Lateral reading: how do experts verify?
Sam Wineburg and Sarah McGrew studied how professional fact-checkers evaluate websites compared with historians and university students. They found that fact-checkers do not linger on the page itself or rely on its appearance or on what it says about itself. Instead they quickly open new tabs to see what other sources say about the publisher. The researchers called this practice "lateral reading" and showed that it leads to more accurate judgements in less time.
Practical steps for verification
In their 2023 book, Mike Caulfield and Sam Wineburg summarise a practical approach that can be applied in minutes:
- Stop: before you share or believe, ask yourself whether you know the source and whether the information triggers a strong emotion in you.
- Investigate the source: look up the publisher outside its own website and find out about its expertise and interests.
- Find better coverage: have reliable sources reported the same thing? What do they say?
- Trace the claim to its origin: go back to the original study, document or image and make sure the context has not been distorted.
Hallucination in language models
A comprehensive review published by Ji and colleagues in 2023 notes that "hallucination", meaning content that is incorrect or not supported by the source, is a well-known phenomenon in natural language generation systems, with several causes related to the training data and the way text is generated. Its most common forms in education include:
- Scientific references that look real but do not exist.
- Precise-looking figures and dates that are wrong.
- Quotations attributed to people who never said them.
- Summaries that add things the original text does not contain.
How to check AI outputs
- Search for every reference by its title in a scholarly database or on the publisher's site; the presence of a link is not enough.
- Compare figures with an official primary source, such as statistical reports and international organisations.
- Rephrase the question and see whether the answer changes; a large change signals weak reliability.
- Ask the tool to name its sources, then check them yourself.
- Distinguish between what can be verified and what is opinion or estimate.
Summary
Verification needs regular habits more than complex tools: stopping, reading laterally, tracing claims to their origin, and treating AI outputs as drafts that need checking.
References
- Wardle, C., & Derakhshan, H. (2017). Information Disorder: Toward an Interdisciplinary Framework for Research and Policymaking. Strasbourg: Council of Europe.
- 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).
- Caulfield, M., & Wineburg, S. (2023). Verified: How to Think Straight, Get Duped Less, and Make Better Decisions about What to Believe Online. Chicago: University of Chicago Press.
- Ji, Z., et al. (2023). Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12), Article 248.
