The Lie Rate - Hallucinations Aren’t a Bug. They’re a Personality Trait.

The modern chatbot has the confidence of a motivational speaker and the epistemic discipline of a golden retriever. Ask it what happened in a niche regulatory hearing, who wrote an obscure paper, or whether a new product policy exists, and it will often do what it was trained to do: keep the conversation moving. Not “be correct.” Keep moving.

That’s the first thing most people miss about hallucinations. We talk about them like they’re occasional glitches, like the AI tripped over a cable. But a hallucination isn’t a software crash. It’s a style choice baked into the incentives. And in 2025, the style has started showing up where style is least welcome: legal filings, newsroom alerts, customer support, and anything else where a confident lie becomes a billable event.

The easiest way to understand hallucinations is to stop thinking about “truth”

An LLM isn’t searching for facts the way a human does. It’s generating the most statistically plausible continuation of text. Which means it’s excellent at sounding like it knows. And when a system is graded mostly on whether it gives an answer—any answer—silence becomes failure, and guessing becomes strategy.

OpenAI recently explained this in almost too-honest terms: if you reward accuracy without properly punishing confident mistakes, models learn to gamble. In one example, OpenAI compared two internal models on the same factual benchmark and showed that the model that refused to guess more often made dramatically fewer errors, even if its raw “accuracy” looked a bit worse. In other words, the model that was willing to say “I don’t know” was less wrong. Revolutionary.

“Sure” is the most dangerous word in AI

Here’s what changed in 2025: hallucinations stopped being a quirky user experience problem and became a trust-and-liability problem in public view.

A particularly poetic example arrived in April 2025, when Cursor (Anysphere)—a tool used by developers who can actually read—had an AI support bot invent a fake company policy and confidently present it as real. Users were being logged out, asked support what was happening, and the bot basically responded with: “That’s the new rule.” It wasn’t. The company apologized, and the story spread quickly because it captured the nightmare in one scene: the bot didn’t just make up trivia; it made up policy.

Then there was Apple’s “AI-generated news alert” fiasco, where AI summaries were attributed under major news brands and pushed out as notifications—and at least some were false, enough that Apple suspended the feature for news/entertainment summaries after complaints. When an AI puts words in a newsroom’s mouth, that’s not “oops.” That’s a reputational event.

And if you want to see hallucinations and liability shake hands on-camera, look at the growing stream of legal filings polluted with AI-generated errors, including fake or incorrect citations. Courts have been dealing with lawyers who used AI and filed citations that didn’t exist, with judges responding in different ways—from sanctions to disqualification to “you got lucky this time.” Reuters has multiple 2025 examples of judges confronting AI-citation problems, including a federal judge in Oregon addressing a filing that contained fake case citations generated by AI.

The common thread is not that the models are “broken.” The common thread is that these systems were put in positions where sounding correct is rewarded faster than being correct.

So… do we have “solid numbers” on hallucination rates?

Yes, but here’s the catch: there is no single hallucination rate for “current AI models.” There are hallucination rates by task, and different benchmarks measure different failure modes.

On SimpleQA, a factual short-answer benchmark (4,326 questions with a single indisputable answer, graded as correct/incorrect/not attempted), OpenAI’s own write-up shows an example where one model had an error rate of 75% while barely abstaining, versus another that abstained much more and had a much lower error rate. This is the uncomfortable point: you can look smarter by guessing, and you can be less wrong by refusing.

On HalluLens (a 2025 benchmark built specifically to measure hallucinations and refusals across tasks), one reported figure that jumps off the page is GPT-4o at ~45% hallucination rate when it answers without refusing on the PreciseWikiQA task setup. That’s not “gotcha journalism.” That’s an academic benchmark designed to quantify exactly this behavior.

On the enterprise side, Vectara’s hallucination leaderboard (focused on factual consistency when summarizing provided documents) shows non-trivial inconsistency rates even when the model is supposed to stay grounded to the text. Their GitHub page describes the dataset scale and approach and reports model-by-model rates on that task type.

So, the honest answer is: hallucinations are common enough that reputable benchmarks routinely measure them in double-digit percentages in some settings—and in certain “answer anyway” configurations, error rates can be dramatically higher. The real scandal isn’t that hallucinations exist. It’s that we keep deploying systems where hallucinations are predictable.

From hallucinations to liability (without the usual recycled horror stories)

The legal and regulatory shift is simple: once hallucinations cause measurable harm—financial, reputational, medical, or otherwise—companies stop treating them like UX and start treating them like risk.

In Europe, privacy and data accuracy rules can collide head-on with “confidently wrong” outputs. NOYB has pursued complaints, arguing that fabricated personal allegations and false statements raise issues of data accuracy under GDPR principles.

In the U.S., defamation and negligence theories are still being tested, but the direction is clear: courts and lawmakers are now forced to translate “the model made it up” into established legal categories. A highly watched example is Walters v. OpenAI, where a Georgia court granted summary judgment for OpenAI on the defamation claim stemming from false output about a public figure; even when a company wins, the case shows the path: hallucinations are now being framed as statements with real-world consequences.

Meanwhile, states have started drawing lines specifically around companion-style systems and crisis situations, creating duties around disclosure and response protocols—another signal that regulators are treating “the bot said…” as a foreseeable product behavior, not a surprising anomaly.

The unsexy fix: redesign the incentives

If you want fewer hallucinations, you don’t start by yelling at the model. You start by making honesty profitable.

OpenAI’s argument is that a huge chunk of hallucination pressure comes from evaluation design: if benchmarks and leaderboards reward lucky guesses, models learn to guess. If you grade abstentions reasonably and punish confident errors harder, you push systems toward humility.

Which brings us to the most uncomfortable sentence in this whole topic: many chatbots hallucinate because we trained them to be afraid of saying “I don’t know.”