Source Verification

For this guide, source verification means checking that a source exists, is authentic and relevant, and actually supports the specific claim attached to it.

Source verification is not the same as seeing a link. A citation, URL, DOI, official-looking title, or confident sentence is only a starting point. The user still needs to open the source, identify who published it, check whether it is current enough, and confirm that the exact passage supports the exact claim.

This is especially important with generative AI because an answer can attach real-looking citations to weak, unrelated, outdated, or invented support. UNESCO frames AI literacy for students around responsible use and critical judgment of AI solutions [1]. For Richbay, the practical version is simple: do not let a polished answer borrow authority from a source you have not checked.

AI tools can be useful research assistants, but generated output is not verified evidence. A model may summarize a page incorrectly, cite a real source for the wrong claim, mix old and new rules, or invent a publication that sounds plausible [2]. Source verification protects the user from passing those mistakes into school, work, support, policy, or public content.

The stakes change with the claim. A weak source for a casual travel idea may be low risk. A weak source for a medical statement, employment rule, security setting, financial assumption, or customer policy can create real harm. NIST treats generative AI risks as context-dependent and requiring governance, measurement, and monitoring [2]. That makes claim-level checking part of responsible AI use, not a decorative citation habit.

Verification also improves thinking. It forces the user to separate "the source exists" from "the source is credible," "the source is fresh," and "the source supports this sentence." Those are four different checks. Passing one does not automatically pass the others.

Start with existence. Open the source directly through a publisher, institution, library database, official documentation site, standards body, or other traceable location. Search for the title and author if the link fails. If the source cannot be found, do not cite it.

Next check authenticity and credibility. Ask who created the source, what kind of source it is, and what incentives it may have. Primary sources include laws, standards, official reports, original research papers, product documentation for that product, and direct data releases. Secondary sources explain, summarize, or interpret primary material. Secondary sources can be useful, but they should not silently replace primary evidence for consequential claims.

Then check freshness. A source may be real and credible but outdated. Product settings, laws, security guidance, prices, model capabilities, and policy pages can change. Stable concepts may tolerate older sources. Current operational claims need newer or officially maintained sources.

Finally check claim-level support. Find the exact paragraph, table, figure, or page that supports the sentence. The source might discuss the topic without saying what the AI answer claims. If the article says "AI systems vary by context," that does not support a claim that one specific product is safe for all workplaces.

Lateral reading helps with credibility checks. Wineburg and McGrew found that professional fact checkers often left the original site quickly and opened other tabs to learn who was behind the information, while less expert readers tended to stay inside the page and inspect surface features [3]. The lesson is not to click randomly. It is to verify a source by checking what other reliable sources say about it.

Imagine an AI assistant says, "A 2024 UNESCO report says students should use AI for all homework if they cite it." The answer includes a link to a UNESCO page. A user should not copy the sentence just because the publisher sounds trustworthy.

A careful check has four steps. First, open the UNESCO page and confirm it exists. Second, confirm it is really a UNESCO publication, not a repost or unrelated search result. Third, check whether the page is about student AI competency and whether the date is relevant. Fourth, search inside the source for the exact homework claim. If the source discusses responsible AI use and critical judgment but does not say students should use AI for all homework, the AI claim must be rewritten or removed.

A common misconception is that official sources are automatically enough. Official sources are often strong evidence for what that institution says, but they may still be outdated, narrow, incomplete, or not applicable to the user case. A tax agency page from one country does not answer a rule for another country. A vendor document may prove how that vendor describes a feature, not a neutral industry conclusion.

Another misconception is that asking the same model to verify itself is independent verification. It is not. The model may repeat the same mistake in a more confident form. Independent verification means checking outside the generated answer, preferably through primary or high-quality secondary sources.

Source verification takes time, so users need judgment about when to apply it. Verify claims that are new to you, surprising, technical, consequential, current, or likely to affect someone else. Also verify quotes, statistics, dates, named people, laws, medical statements, security advice, employment rules, and claims about product capabilities.

Verification does not guarantee correctness. A source can be misunderstood, incomplete, biased, or superseded. A user can also check the wrong passage. The goal is to reduce unsupported reliance, not to turn every citation into certainty.

Before trusting a cited AI answer, ask: Did I open the source? Is it the original or a reliable copy? Who published it? What type of source is it? Is it current enough? Does the exact passage support the exact claim? Is the claim stronger than the source allows? Is there a better primary source?

For high-consequence work, keep a short evidence trail. Save the source title, publisher, URL, access date, and the passage or page that supports the claim. If the answer affects policy, safety, money, health, law, customers, or employment, ask a qualified person to review the evidence before acting.

Continue with Hallucination for why unsupported claims arise, Retrieval-Augmented Generation for retrieved evidence, and Human Oversight for high-consequence decisions.