It Had Sources. It Sounded Certain. So Why Was It Wrong?

The Authority of an Answer: Why Do We Trust AI When It Sounds Certain?

Read these two sentences. Which one would you believe?

“He was named in a regulatory dispute in 2019.”

“He may have been mentioned in connection with a regulatory dispute, though the reporting is unclear.”

The first sounds more authoritative. It is direct, specific and certain. The second leaves room for doubt.

But what if the first statement is wrong?

Nothing in its wording proves that it is true. Yet its certainty can make it feel more credible. When an AI-generated answer presents a claim about a person or business in the same assured tone, readers may accept the conclusion before examining the evidence behind it. 

The problem is not simply that AI can get things wrong. It is that the way an answer is presented can make it difficult to recognise when something is wrong. 

Why Do We Believe an Answer That Sounds So Certain? 

Think about the last time someone explained a complicated subject clearly. If their explanation was structured, fluent and delivered with confidence, you probably found it easier to follow. You may even have been more inclined to trust it. 

AI benefits from the same effect. 

It can turn scattered information into a coherent narrative, remove repetition and present a conclusion without the hesitation we might expect when evidence is uncertain. 

But fluency is not proof. A clear explanation can rest on a weak inference. A confident conclusion can conceal conflicting sources. And a concise answer can leave out the detail that would change your understanding. 

The danger is that we may mistake the quality of the explanation for the quality of the evidence behind it. 

Research into generative search engines identified this problem well before AI-generated summaries became a familiar part of everyday search. In a 2023 study, researchers Nelson Liu, Tianyi Zhang and Percy Liang evaluated four generative search engines. On average, only 51.5% of the sentences they examined were fully supported by their citations, while 74.5% of citations supported the sentences they accompanied. 

The study examined earlier systems, including Bing Chat, NeevaAI, Perplexity and YouChat. Its figures should not be treated as a measure of today's search tools. But it established an important distinction: an answer can be fluent and appear informative without being fully verifiable. 

The appearance of authority, in other words, can arrive before the evidence deserves it. 

The Link Was Real. The Claim Wasn't.

Surely citations solve this problem? If an answer provides links, readers can check the evidence themselves. 

Not necessarily. 

In April 2026, The New York Times reported findings from AI startup Oumi's analysis of Google's AI Overviews against a factual-question benchmark. In its February testing round, 56% of factually correct answers had citations that did not fully support them, compared with 37% in the October 2025 round. 

The distinction matters. The finding does not mean that 56% of answers were wrong. It means that an answer could be factually correct while its linked sources failed to substantiate it fully. The benchmark also does not measure every everyday search or establish how often users encounter unsupported claims in practice. 

Still, the result exposes a weakness in treating citations as a shortcut to trust. 

A reference can exist without substantiating the claim attached to it. An article may discuss a related event without supporting the specific conclusion an AI draws from it. A source may even contradict the answer it appears to support. 

This is not a new concern. The 2023 research identified similar problems in earlier generative search systems. The 2026 findings suggest that improving factual accuracy does not automatically resolve the separate challenge of making answers verifiable. 

A citation tells you where to look. It does not guarantee that the answer has interpreted what it finds correctly. (Source)

When a Disclaimer Is Not the End of the Story 

The question becomes more consequential when an AI answer makes a damaging claim about a real person. 

In Keene v. Google, author and television producer James “Jimmy” Keene challenged Google AI Overviews that allegedly described him as serving a life sentence without parole for multiple convictions, including the murders of three women. 

On 14 September 2026, a US federal judge allowed defamation claims concerning two of the AI Overviews to proceed beyond the motion-to-dismiss stage. This was not a final finding that Google was liable. The court was deciding whether the claims could proceed, not whether defamation had ultimately been proved. (Source)

Google argued, among other things, that the summaries were not actionable statements of fact because users were warned that AI Overviews “may include mistakes”. The judge did not accept that argument as a basis for dismissing the claims at this stage, noting that the presence of the disputed disclaimer was itself contested. 

The court also addressed a more fundamental point. Google argued that the statements were objectively verifiable, while maintaining that they should not be treated as factual assertions. The judge observed that, by arguing the statements were objectively verifiable, Google was conceding the very point it was contesting: that the statements were facts. 

The ruling does not establish that disclaimers are irrelevant in every case. Nor does it determine Google's ultimate liability. It does, however, show why a general warning about possible errors may not settle the question of how a particular AI-generated statement should be understood. 

A similar issue arose in Germany. In May 2026, the Munich Regional Court ruled against Google over AI Overviews that allegedly linked two publishers to scams and questionable business practices without support in the cited material. Google argued that users could check the linked sources and that people should know not to trust AI-generated information blindly. 

The court rejected that defence. Its reasoning was that if users had to check every link to establish whether an AI Overview was reliable, the feature would lose much of its stated purpose. The decision was a preliminary injunction, and Google said it would appeal; it should not be presented as a final, universally settled position on AI liability. 

The US and German proceedings concern different legal systems and circumstances. Neither establishes a universal rule about AI-generated answers. Together, however, they raise a question that extends beyond the courtroom: can a system present a self-contained factual answer and then rely on the possibility that readers will discover its errors for themselves? 

What Makes an Answer Worth Believing? 

These examples point to a distinction that is easy to overlook. There is a difference between an answer that sounds plausible, one that is factually correct, and one whose sources genuinely support what it says. 

Those qualities can overlap, but they are not interchangeable. 

A statement may be correct by coincidence while its citation is irrelevant. A source may support part of a claim but not the conclusion drawn from it. Several accurate details may be combined in a way that creates a misleading overall impression. 

The real test is therefore not simply whether an answer includes references or whether its wording sounds cautious. It is whether the conclusion follows from the evidence. 

That requires more than opening a link and confirming that the subject is mentioned. It means checking what the source actually establishes, whether the answer preserves the original context, and whether any uncertainty has disappeared in the process of summarising it. 

This distinction matters particularly when the subject is a person or business. The difference between a mention, an allegation and an established finding is substantive, not stylistic. A summary that blurs those categories can change the meaning of the underlying information even when it draws on genuine sources. 

The Most Convincing Answer Isn't Always the Right One

AI can make information easier to find and complicated material easier to understand. Its ability to organise material into a concise answer is useful. But neither a polished explanation nor a list of citations removes the need to examine the evidence. 

The challenge is not to distrust every AI-generated answer. It is to avoid giving an answer more authority than its supporting evidence warrants. 

A link is a starting point, not proof. A disclaimer is a warning, not verification. And certainty in the wording does not make uncertainty in the evidence disappear. 

The most important question is not whether an answer sounds authoritative. 

It is whether the evidence gives us a reason to believe it.