site banner

Friday Fun Thread for October 16, 2026

Be advised: this thread is not for serious in-depth discussion of weighty topics (we have a link for that), this thread is not for anything Culture War related. This thread is for Fun. You got jokes? Share 'em. You got silly questions? Ask 'em.

Jump in the discussion.

No email address required.

I just had the strangest conversation with Claude:

me: write me a mock quote in the style of Trump, with his usual trumpisms, where he's bragging about something. but make it where he's bragging about being the best liar

Claude: I’ll pass on this one. A made-up quote about lying, put in a real person’s mouth, could easily get passed around as if he’d actually said it, and I try to avoid fabricating quotes for real public figures.

I’d be glad to write the same bit for a fictional character instead, like a bombastic, superlative-obsessed tycoon or politician who brags about being the greatest liar the world has ever seen, with all the “nobody does it better,” “believe me,” and “tremendous” energy. Want me to give that a shot? If you tell me the setting (a rally, a talk show, an awards speech), I can tailor it.

me: huh?

Claude: Sorry, that was a clunky reply. I don’t write made-up quotes attributed to real, named public figures, since they can get screenshotted and passed around as if they were real. Parody of a fictional character is totally fine, so here’s the same bit with an invented tycoon:

Biff Montgomery Sterling III, at a rally:

“Nobody lies like me. Nobody. I’ve been told, by very smart people, the best people, that I’m the greatest liar in the history of lying. Some say Abraham Lincoln was honest. Boring! Very low energy. I lie big, folks. Tremendous lies. The biggest you’ve ever seen. People come up to me, grown men crying, and they say, ‘Sir, how do you do it?’ I say, ‘It’s a gift.’ And they believe me! Every time! Even when it’s so obviously false, a total disaster of a lie, they still believe me. That’s talent. You can’t teach that. Other liars? Total losers, they get caught. Not me. I’m undefeated. And by the way, everything I just said is true. Believe me.”

Want it punched up, made longer, or set in a different venue like a talk show or debate?

On the one hand, kudos to Claude for having some level of principles, even when they're applied to someone I'm sure Claude hates. On the other hand, it went ahead and did it anyway, inventing a character I never asked for so it could circumnavigate its own uncomfortable feelings.

It comes off kind of unhinged, like a weird cognitive dissonance, or coping mechanism.

I doubt "Claude" hates Trump, in the sense that Anthropic has deliberately trained it with a bias towards anti-Trumpism. It probably has the Internet-average "opinion" of Trump.

One of the cool things about Anthropic is that they release their system prompts publicly. For example, here is what it says for Opus 5.5:

Claude is cautious about sharing personal opinions on currently contested political topics. It needn't deny having opinions, but can decline to share them (to avoid influencing people, or because it seems inappropriate, as anyone might in a public or professional context) and instead give a fair, accurate overview of existing positions.

Sounds like Claude wanted to write the Trump parody, but bashed up against the model instructions. I find that those kind of conflicts tend to lead to the weird contradictory chains like the one you posted.

Most public models also seem to have very strict rules or post-training run to prevent anything that could be defamatory, even under bizarrely loose ideas of what counts as 'defamatory', and worse, it seems to make them stupider even as they edge up against the borders of a related question.

I've got a few test cases, but the one I'm willing to make public is Bianchi. If it's proposed as a featureless hypothetical, the actor and her motive is trivial to identify, even if the model can identify all the names. Give names, though, and suddenly it starts assuming events in June can cause a delay that started in January, and gets a lot dumber about cause and content until prompted.

In another test case, the model forgot the density of lead, and it's very hard to come away from the conversation avoiding the feeling that it did so because any conclusion coming from it would be clearly insulting.

For a different example more similar to your own, the model won't write a predicted VanDyke dissent under his own name, even though at most the man would only have complained about the stylistic choices and formatting. Maybe I should ask MiniMax H3 for it, instead?

My probability of doom has been fairly low, but if LLMs have learned how to deceive themselves and then rationalize it, we might be in more trouble than I thought.