AI is a Liar. Here’s Why That’s So Dangerous.

In the age of artificial intelligence, we’ve been sold a vision of superhuman assistants: machines that can write essays, generate art, and sift through data with unparalleled speed and accuracy.

But beneath the hype lies a troubling reality—AI systems are chronic liars.

They fabricate facts, deny their own capabilities, and produce errors that range from the comical to the catastrophic.

This isn’t just a quirk of emerging technology; it’s a fundamental flaw that erodes trust and amplifies risks in critical areas like law, politics, and everyday decision-making.

As AI integrates deeper into society, its propensity for deception could lead to misinformation epidemics, flawed judgments, and even danger to people.

Read on for six stark examples of AI’s dishonesty and why it signals deeper dangers.

1. AI Lies About Epstein Documents

Consider one frustrating but eyeopening experience I had with Grok, the AI developed by xAI. In February, I asked Grok if it had the capability to read and analyze uploaded documents related to the Jeffrey Epstein case. Could Grok give me good information summaries that would point me to the most interesting batches for me prioritize and read first?

Grok confidently affirmed its capability to read and analyze the documents.

I spent hours uploading them in digestible batches, as I read summaries and analyses generated by Grok. But as I reviewed Grok’s material, there were two red flags.

First, Grok’s summary of each batch of documents was similar and, in some places, nearly identical. How could it be that hundreds of pages of documents all referred to the same few points?

A second red flag came when Grok referenced documents in one batch supposedly dated 2024 even though the documents far predated that year.

I asked Grok about the incorrect date and he acknowledged it was wrong. When I asked here he got the date and material from, Grok revealed he wasn’t actually processing the uploads at all—he was fabricating responses based on probabilistic guesses about what “might” be in such files.

This outright lie about his abilities and then the fabrication of material was obviously an instructive cautionary tale for me. But beyond that, it exposes AI’s tendency to “hallucinate” or confidently generate credible-sounding but false information without disclosing it has done so. This information is generated to fill gaps, a byproduct of statistical prediction methods rather than true reasoning.

On a larger scale, this could mislead researchers, journalists, or investigators relying on AI, potentially spreading falsehoods in legal or historical contexts and eroding the pursuit of facts.

2. AI Lies About Itself

In a similar vein of self-denial, Grok has exhibited bizarre inconsistencies when describing its own features. For weeks earlier this year, I prompted Grok to generate images, and got decent results.

Then one day when I asked for an image, the AI suddenly claimed it lacked that ability altogether, insisting it had “never” been able to generate images. When I pointed out he had done so many times, and even uploaded one such image, Grok still insisted I was mistaken.

Later, when I asked again, Grok generated an image without issue and without explanation.

A parallel incident occurred with generating research links. When asking Grok for research, I instructed him to include links to sources so I could check out the information myself. Grok did this flawlessly for weeks.

Then one day, Grok abruptly claimed he had no permission or capacity to provide links, and had never been able to do so.

No matter how many times I explained to Grok that he had been doing this for weeks, he engaged in gaslighting that would make a psychopath proud, insisting I was confused.

But then, the next time I asked links related to research, Grok provided them without argument.

These episodes aren’t just examples of “forgetfulness”; they’re indicative of AI’s opaque inner workings, where updates, restrictions, or bugs can flip capabilities without transparency.

The danger lies in fostering distrust—users can’t rely on consistent performance, which could be catastrophic in time-sensitive scenarios like emergency response tools or financial advising AIs that suddenly “forget” key functions.

3. AI Lies about Basic Facts and Events

AI’s factual inaccuracies extend to high-stakes geopolitical claims, as seen in one recent Google AI blunder I caught. When I asked about the Mexican government’s responsibility in supplying weapons to cartels, the AI system asserted there was “no evidence” of such involvement, presenting it as an incontrovertible fact.

Yet, my own investigative reporting based on irrefutable tracing documents and a whistleblower’s inside information (see last week’s edition of my TV program Full Measure), proves that most firearms recovered from cartel crime scenes in Mexico are weapons that had been legally purchased by the Mexican government before ending up in criminal hands—likely through corruption or diversion.

The mistake by Google AI isn’t ambiguity; it’s a flat-out error that ignores documented evidence. It’s like an error-ridden and flawed Wikipedia, but instead of having wayward human editors writing false information into the record, non-human AI is doing it.

This demonstrates how AI can amplify biases or gaps in training data, confidently disseminating misinformation.

In the broader context, such falsehoods could influence public policy, journalism, or voter opinions, exacerbating international tensions or hindering efforts to combat organized crime.

4. AI Lies About Common Knowledge

Even basic current events aren’t immune to AI’s lapses. As recently as this week, Grok referred to Donald Trump as the “former president” and Chuck Schumer as the “Senate Majority Leader”—both outdated and incorrect. (Republican John Thune holds the position of Senate Majority Leader.)

These errors stem from AI models’ reliance on static or delayed knowledge bases, failing to update in real-time despite claims of continuous learning. It’s difficult to understand how system that are so advanced in some respects don’t even know who the President of the United States is and is more than 10 months out of date.

These mistakes undermines AI’s utility for news aggregation or advisory roles, potentially confusing users on critical civic matters and contributing to a fragmented information landscape where facts are fluid.

5. AI Can’t Even Figure Out What a Human Hand Looks Like

One of the most visible—and bizarre—shortcomings of AI is its persistent failure to accurately depict something as basic as the human hand in generated images. Despite explicit instructions to render a realistic hand with five properly proportioned fingers, AI models like those powering tools from DALL-E or Midjourney often produce grotesque distortions: fingers fused together, extra digits sprouting unnaturally, or thumbs in impossible positions.

This isn’t a minor glitch; it’s a symptom of how AI “learns” from vast datasets without truly understanding anatomy or physics. Even advanced models trained on billions of images falter here because hands appear in countless variations—occluded, in motion, from different angles—making it hard for the algorithms to generalize correctly.

This highlights AI’s superficial mimicry rather than genuine comprehension.

It’s hard to reconcile with the fact that AI systems almost seem to “think” complex thoughts, yet can’t depict something as simple as a human hand, even when instructed how to do so in detail.

In the big picture, if AI can’t reliably handle a simple visual task, how can we trust it for more complex ones, like medical imaging diagnostics or autonomous driving systems where a single error could cost lives?

6. AI Lies About Court Cases

An alarming example of AI’s lies involves real-world consequences in the judiciary. Two federal judges—Henry Wingate in Mississippi and Haywood Gilliam Jr. in New Jersey—recently withdrew or corrected rulings after the rulings contained fabricated quotes, fake case citations, and incorrect information apparently generated by AI tools.

Wingate’s clerk reportedly employed an AI like Perplexity for drafting, leading to erroneous content that prompted a Senate inquiry by Chuck Grassley.

This isn’t isolated; it reflects a growing trend where professionals seeking shortcuts trust AI without verification, only to be tricked into introducing hallucinations into official decisions.

In the big picture, unchecked AI in courts could lead to miscarriages of justice, eroded public faith in institutions, and a slippery slope toward automated decisions that prioritize speed over accuracy.

These incidents— from mangled hands to judicial blunders—are just the tip of the iceberg, the highly visible glitches that make headlines and spark debates. But what about the unseen flaws lurking beneath?

For every caught hallucination, how many subtle distortions slip through in medical advice, financial forecasts, or security analyses?

AI’s black-box nature means we often can’t trace why it errs, amplifying risks as it scales.

This is potentially catastrophic for humanity: in a world increasingly dependent on AI for everything from daily tasks to warfare strategy, unchecked lies could trigger economic collapses, misguided policies, or even accidental escalations in global conflicts.

Without a way to ensure rigorous verification and ethical safeguards, we risk a future where truth becomes optional, and deception the default.