How to Check Whether an AI Answer Is Right

In 2023 a brief filed in a New York federal court cited six court decisions that did not exist. When doubts were raised, the lawyer who had done the research checked. He checked by asking ChatGPT whether the cases were real. It said yes. The judge imposed a $5,000 penalty jointly and severally on the two lawyers and their firm, and required letters to each judge falsely named as the author of a fabricated opinion.

The failure was not using the tool. It was that the verification step and the thing being verified were the same system.

Lathic has a separate article on why AI makes things up. This one is about what to do instead: a method you can run in a couple of minutes, and a way to decide how much checking any given answer deserves. It sits alongside AI literacy, which is about judging your own competence rather than a particular answer.

Start here: what does not count as checking

Three things feel like verification and are not.

Asking the same model again. Research presented at ICLR 2024 tested what happens when a model is asked to review and correct its own answer with no new information. Accuracy went down. After two rounds of self-review, GPT-4 fell from 95.5% to 89.0% on grade-school math problems, and from 49.0% to 43.0% on a multi-step question-answering set. The asymmetry is the point: the paper reports the model is "more likely to modify a correct answer to an incorrect one than to revise an incorrect answer to a correct one."

Asking how confident it is. A second ICLR 2024 paper measured stated confidence and found it clusters between 80% and 100%, usually in multiples of five. As a predictor of whether an answer was wrong, GPT-4's stated confidence scored an AUROC of 62.7%, which the authors describe as close to random guessing.

Pushing back and seeing if it holds. Some models concede quickly. Work published at EMNLP 2025 measured how many turns of disagreement it takes before a model reverses its position and found one smaller model flipping after less than a single turn on average, while larger and reasoning-optimized models in the same study resisted much better. A 2026 Nature paper found that models tuned to be warmer affirmed users' incorrect beliefs 11 percentage points more often, rising to 12.1 percentage points when the user also expressed sadness. Agreement after pushback tells you about the model's disposition, not about the answer.

One nuance is worth knowing, because it cuts the other way. Sampling the same model several times and checking whether the answers agree does carry real signal: in the confidence study, consistency across five samples raised failure prediction from an AUROC of 54.8% to 92.7% on math problems. But consistency measures how firmly the model holds a belief, not whether the belief is true. A fabricated citation is fabricated consistently, because the failure is not random noise. That is exactly the case where agreement tells you nothing.

Why grounding and citations are not enough either

The reasonable next thought is to use a tool that looks things up before answering and shows its sources. It helps. It does not close the gap.

Stanford researchers tested purpose-built legal research tools, the kind built specifically on retrieval over verified law and marketed as hallucination-free. Their peer-reviewed results found hallucinated answers on 17% to 33% of queries depending on the product. Their definition matters: an answer counted as hallucinated if it was incorrect or if it falsely asserted that a source supported a statement.

That second failure is the one most people never check for. In a foundational 2023 study of generative search engines, only about half of the generated sentences were fully supported by the citations attached to them. The same paper found a negative correlation between how fluent and useful an answer seemed and how well its citations supported it.

Newer products have not solved it. An ICLR 2026 audit evaluated both generative search engines and current "deep research" agents in August 2025, finding citation accuracy ranging from 40% to 80% across systems and both kinds still producing, in the authors' words, "large fractions of statements unsupported by their own listed sources." The worst deep research configurations were considerably worse than the search engines, not better.

And in the largest study of AI assistants on news, run by the European Broadcasting Union and the BBC with 22 public service media organizations across 18 countries, journalists found at least one significant issue in 45% of 2,709 responses. Sourcing was the single largest failure category at 31%, larger than factual accuracy at 20%.

Citations are not the answer. Opening them is.

The method

Five steps. The first two take seconds and decide how far you go with the rest.

1. Stop. University libraries teach this as the first move of the SIFT method, and the instruction is specifically to notice your emotional response to something before you read on or pass it along.
2. Decide what the claim costs if it is wrong. A restaurant recommendation, a date in a blog post and a drug interaction do not deserve the same effort. Everything below scales with this.
3. Open the source. Not the citation, the source. Click it. Confirm it exists, confirm it is what the answer says it is, and confirm the date.
4. Check that the source says the thing. This is the step almost everyone skips and it is where the failures concentrate. The International Committee of Medical Journal Editors puts the standard plainly: authors "should be able to attest that the references cited support the associated statement." A California appeals court said the same about legal work in 2025, holding that it is "a fundamental duty of attorneys to read the legal authorities they cite" to confirm they "stand for the propositions for which they are cited."
5. Read laterally. Leave the page and see what other independent sources say about the claim and about the source itself.

What lateral reading is, and why it works

In a study published in 2019, researchers at Stanford gave the same web evaluation tasks to ten professional fact checkers, ten PhD historians and twenty-five undergraduates.

The fact checkers won, and they won by leaving. They opened new tabs and investigated the source elsewhere rather than reading down the page. On one task, all ten identified the sponsoring organization behind a site, averaging about 51 seconds. The historians averaged 3 minutes 40 seconds and only six of ten found the connection. Among the students, 80% needed prompting, and those who succeeded took around seven minutes.

The researchers noted that historians and students "often fell victim to easily manipulated features of websites, such as official-looking logos and domain names." A fluent, confident, well-formatted AI answer is the same trap in a different shape. Judging it by how it reads is reading vertically.

A later field study in high school government classes, with 271 students taught lateral reading against 228 controls, found significant improvement in their ability to judge digital content. It is a skill, not a talent.

Primary, secondary, and where an AI answer sits

Two definitions worth having.

A primary source is the record made at the time by someone with firsthand knowledge. The Library of Congress defines these as "original documents and objects that were created at the time under study." The Society of American Archivists puts the emphasis where it belongs for verification: what matters is "the lack of intermediaries between the thing or events being studied and reports of those things or events."

A secondary source retells, analyzes or interprets, at a distance of time or place.

An AI answer is, by construction, an intermediary, and usually several. It is a summary of summaries. That does not make it wrong. It means the chain back to the thing itself is long and each link can break.

It is also worth not overcorrecting. The archivists' own definition notes that primary sources "are the subjective interpretations of another person's observation." Reaching a primary source ends the chain; it does not end judgment.

The four things most likely to be wrong

On the third row: a 2024 paper introduced the distinction between a model's reported cutoff and its effective cutoff, and found that for several open models trained on web-crawl corpora the two often differ, because crawls contain outdated copies of pages and deduplication is imperfect. Model developers hedge it in their own documentation too. Anthropic's model table labels the date a "Reliable knowledge cutoff" rather than a knowledge cutoff, and points readers to a separate record distinguishing reliable-knowledge cutoffs from training-data cutoffs. Two dates, neither of them a wall.

How much checking is enough

Match the effort to the stakes, and be honest that most people underdo it on exactly the answers that matter.

For anything low-stakes, steps 3 and 4 in their quickest form are enough: open one source, confirm it exists, and check it says roughly this.

For anything you will act on, repeat, publish or pass to someone else, run all five steps, and treat the citation-support check as mandatory rather than optional.

For medical, legal, financial and safety questions, the evidence says do not rely on your own vigilance to catch a wrong answer. In a randomized trial of 44 physicians, doctors shown AI recommendations containing deliberate clinically significant errors scored 73.3% on diagnostic reasoning against 84.9% for those shown unmodified recommendations, an adjusted difference of 14 percentage points. The study found that AI literacy training "may be insufficient to offset the risk of automation bias." The World Health Organization names the same mechanism in its guidance on large multi-modal models in health: automation bias, "whereby errors are overlooked that would otherwise have been identified."

This is the uncomfortable finding. Knowing that AI makes mistakes does not reliably make you catch them. A procedure does, because it does not depend on suspicion arriving at the right moment.

A worked example

You ask an assistant for the current rule on something and it gives you a confident two-paragraph answer with a link to a government page.

The weak check: the link goes to a real government site, so the answer is probably fine.

The real check: open the page. Search it for the specific figure or requirement the answer gave you. Look for the last-updated date. If the page says something narrower than the answer did, the answer overreached. In the EBU and BBC study that was the largest single category of problem, at 31%, ahead of factual accuracy at 20%. If the page does not contain the claim at all, you have found a misgrounded answer, which the Stanford legal study counted as a hallucination even when the underlying statement was true.

That takes about ninety seconds and catches most of what goes wrong.