We have entered an uncomfortable era in which two sources people traditionally rely upon for information—government leaders and search engines—can both produce information that is false, misleading or impossible to verify.

A politician can exaggerate, misrepresent statistics or repeat a claim that has already been disproved. An artificial-intelligence chatbot can produce a convincing answer, complete with invented facts and nonexistent citations. And increasingly, AI can create photographs, videos and audio recordings that appear authentic.

The result is not simply that people encounter more misinformation. It is that the public has a harder time knowing what deserves to be believed. That makes an old journalistic rule more important than ever: Don’t ask who said it. Ask what evidence proves it.

Artificial intelligence is particularly dangerous when it sounds certain. The National Institute of Standards and Technology, or NIST, calls one of the major problems with generative AI “confabulation.” The agency defines it as situations in which AI systems generate and confidently present erroneous or false information. NIST notes that AI can even produce fabricated reasoning and citations that appear to support an incorrect answer. It can even invent the citations.

Stanford researchers have independently found serious weaknesses in AI fact-checking. In testing, standard AI models relying on their internal knowledge performed poorly when asked to determine whether claims were true or false. Researchers warned that a chatbot that confidently mislabels a claim can actually reinforce misinformation rather than correct it. 

That means an AI response should be treated like any search engine or Wikipedia-like database, as a starting point for research, not the final authority. An AI chatbot can help locate a government report. It cannot make the report true.

The same principle applies to elected officials. A statement from a president, governor, mayor or council member establishes that the person said something. It does not establish that the statement is true.

For example, the Associated Press recently examined claims by President Donald Trump that documents released by his administration proved allegations of widespread fraud in the 2020 election. AP’s review found that the documents did not substantiate the claims. That distinction is crucial: “The president said X” is a fact. “X is true” is a separate question. Journalists, researchers and citizens should not confuse the two.

The same standard should apply regardless of political party. Political affiliation is not a verification system. A Democratic mayor, Republican council member, independent candidate or government agency can make an inaccurate claim or false statement. Government documents deserve serious consideration because they can contain valuable primary-source information. But “government document” does not automatically mean “correct.”

A striking recent example involves a U.S. Census Bureau report concerning noncitizen voting in the 2020 election. A recent investigation by WIRED reported that researchers had identified significant problems with the data and methodology used to estimate noncitizen voting. Critics said the reported numbers were likely affected by errors in matching databases rather than representing actual illegal votes. The lesson isn’t that government statistics should be discarded. It is that statistics must be examined.

• Who collected the data?
• What population was measured?
• What definitions were used?
• What assumptions were made?
• Can another researcher reproduce the result?
• Were alternative explanations considered?

Those questions are more valuable than simply asking whether the document carries an official government seal.

What can we trust?

There is no single source that deserves unlimited trust. Instead, the most reliable approach is triangulation—finding independent evidence that converges on the same conclusion. For important claims, look for several different kinds of evidence:

1. Primary documents

Whenever possible, go to the original source. If a politician says a law was passed, read the legislation. If an agency announces a program, read the agency’s actual regulations, budget documents or official records. If a city says it spent $10 million, look at the adopted budget, financial statements, contracts and expenditure records. A news story reporting that something happened is useful. The underlying document is usually better.

2. Independent experts

Look for researchers who have no obvious financial or political interest in the conclusion. Universities, professional organizations and peer-reviewed academic research can be particularly useful—but even academic research should be examined for methodology, funding and limitations. A single study is rarely the last word.

3. Multiple independent news organizations

Do not rely on one article. Compare reporting from organizations with different ownership structures and editorial perspectives. If Reuters, Associated Press, a local newspaper, a government document and an academic researcher independently report the same underlying facts, confidence increases. The AP, for example, has maintained systematic verification of claims made by public officials for decades and now also verifies online material and AI-related content. 

4. Data that can be reproduced

The strongest statistics are those another person can independently examine. A claim such as “crime increased dramatically” should lead to actual crime statistics. “Taxes went up” should lead to tax records or enacted legislation. “Housing increased” should lead to permits, completed units or other defined measurements. The question should always be: “Show me the numbers.”

Don’t trust a citation simply because it exists

AI has introduced another trap: fake authority. A chatbot may produce a citation that looks completely legitimate. It may give the name of a university professor, an academic paper and a URL. But the citation may not exist. Or it may exist, but it says something completely different.

NIST specifically warns that generative AI can produce confabulated citations. Therefore, every important AI-generated citation should be independently opened and checked.

Don’t ask: “Does the AI have sources?”

Ask: “Do the sources actually say what the AI claims they say?”

That is a much higher standard.

Beware of photographs and videos

Visual evidence used to be considered particularly powerful. That assumption is disappearing.

New AI systems can create realistic photographs, audio and video. Research published this month found that humans and even sophisticated detection systems can have difficulty distinguishing modern AI-generated videos from authentic footage. Consequently, a photograph should no longer automatically be treated as proof.

• Find the original photograph.
• Determine when and where it was taken.
• Check whether it has been edited.
• Look for other photographs of the same event.
• Look for news articles or reports that were published around the same time as a particular event or situation. These reports can provide immediate context, reactions, and information relevant to the event as it was happening.
• And, whenever possible, identify the original photographer or organization.

The five-question truth test

Before repeating a controversial claim, ask five questions:

1. Who is making the claim? Is it a politician, government agency, company, activist, journalist, researcher or AI system?

2. What is the original evidence? Find the document, dataset, recording, law, court filing or study.

3. Can the evidence be independently verified? Can another person reach the same conclusion?

4. What do credible sources that disagree say? Don’t merely search for information confirming what you already believe.

5. What would change my mind? This may be the most important question of all. If the answer is “nothing,” the problem is no longer lack of information. It is confirmation bias.

A new hierarchy of trust

In the age of AI, perhaps the best hierarchy is not:

• Government
• media
• social media
• AI

Instead, it should be:

• Evidence
• independently verified primary sources
• transparent research
• accountable journalism
• expert analysis
• statements from interested parties
• anonymous social-media claims
• unsupported AI output.

Even that hierarchy isn’t absolute. A government database may be better evidence than a newspaper article for one question, while an independent scientific study may be better than a government press release for another. The key is transparency, corroboration and accountability.

Resources for checking the truth

NIST’s Generative AI Profile explains AI “confabulation” and other risks associated with generative AI. NIST AI Risk Management Framework

Associated Press Verification explains how journalists verify claims, photographs, videos and online information. AP Verification

Stanford research on AI fact-checking provides evidence that AI systems can perform poorly when asked to determine whether claims are true without carefully curated evidence. 

Academic research can be searched through Google Scholar, university libraries and scholarly databases. For important claims, look for peer-reviewed studies rather than merely articles discussing research.

The responsibility ultimately belongs to us

The most dangerous response to misinformation is not skepticism. It is cynicism. If people conclude that “everyone lies” and therefore nothing can be known, misinformation has already won. The challenge is that finding the truth now requires more work.

There are facts. There are reliable records. There are reproducible scientific methods. There are public databases, court records, government documents, academic studies and professional journalists who can be held accountable for errors.

In the past, the question was often: “Is this source trustworthy?”

Today, the better question is: “Can I independently verify the evidence?”

That is the standard citizens, journalists—and increasingly, artificial intelligence itself—should meet. Trust should not be given because someone has authority, a title, a large audience or an impressive computer behind them. Trust should be earned by evidence.