Do LLMs Lie?

The Curious Case of Lex Fridman's Disappeared April Fool's Joke

Podcaster Lex Fridman posted the following midday on April 1, 2025 (April Fool’s Day).

I have notifications on and happened to be looking during lunch and it popped up.

Certainly shocking. More certainly no harm intended and was meant as a very humorous spoof with powerful imagery. Lex is a genius, articulate and thoughtful, especially kind and yes loving host who features all opinions (even those with which he disagrees) on his podcast, and still shows genuine interest and courtesy. This I believe is why he is successful.

I highly recommend the nearly 5m subscribers Lex Fridman Podcast to you.

Thinking it might go away, others immediately suggested the same, I took a screen shot and made a post about the First Amendment and Freedom of Speech. As I was writing the post, Lex’s original disappeared, and I modified my post on the fly.

In matters of free speech it is important that even the shocking distasteful or unpopular opinions are protected, as a Constitutional matter. The reason is that if we begin to ban unpopular notions; human taste and preference being what it is, at some point the winds of favor will inevitably shift and then the things you said could become “banned.”

Moreover, censorship is used by repressive regimes such as the one depicted. This is because these, built on lies, know that it only takes one person and one message to tell others, “The Emperor has no clothes.” And the whole thing collapses in a reversal of public sentiment.

Time went by in the day, and later on I saw another account posting it, then deleting it (as had Lex quickly). Another user asked “what was it” and someone told them.

Then they asked Grok, xAI’s best of class AI LLM “did Lex really post that?” This elicited the following AI-automated reply from Grok:

I knew it was untrue and my eyes had not deceived me, his original post was right there in mine.

After work I decided to ask Grok myself about the whole matter. It told me, in essence, the same as above. It was now a “fabricated” April Fool’s Day “prank” or “hoax.”

And so I pressed on as attorneys are wont to do; cross examining the LLM:

And it finally relented, as below. It remains to be seen whether this interchange will update the overall messaging given by Grok on this issue.

It is a perfectly fair question to wonder and to ask in the hypothetical, can an LLM [would an LLM] lie?

Most charitably we have to presume that since it was deleted that the database from which Grok drew could at the time asked find no reference to it and this chalked it up as false.

And yet the extended explanation it gave me at the end of the day, thoroughly labeling as certainly “hoax” and “prank” (as if confirmed as such) could give one pause.

We are certainly on the beginnings of the beginning of AI and LLMs. Grok 3 is less than weeks old. I’ve said before I think it is the most anthropologically human LLM.

Ethical frameworks are talked about in some circles. And Elon Musk has indicated that biased AI or tampered-with or manipulated AI that did not say what was true pose a grave risk — are very much not the agenda for xAI.