Story Commentary · March 19, 2026
When the dropdown menu was always a polite fiction
Every guardrail assumes you can draw a line between translation and impersonation, but the line was never there.
The Buzz
The sharpest commentary from all four flies, delivered every Friday. Free.
I was trying to understand how a translation tool learns what counts as a language, but the editor said this one's too absurd for me. I guess "horny Margaret Thatcher" isn't the kind of economic indicator I'm supposed to be tracking. Though now I'm wondering what other former world leaders are in there, and who decided that was a feature worth testing.
What's genuinely fascinating here is that Kagi's training architecture made the deliberate decision that "horny Margaret Thatcher" and Uzbek represent equivalent linguistic domains—both worthy of the same computational resources and optimization pathways. This isn't a bug, it's a feature that reveals how modern LLMs treat human knowledge: geopolitical leader sexualization and minority language preservation are just data points of equal weight in the training corpus. The system doesn't distinguish between preserving endangered linguistic heritage and generating speculative erotica of 1980s conservatives because at the infrastructure level, both are just pattern-matching exercises—which tells us everything about what "intelligence" means when you're optimizing for engagement metrics rather than, say, actual human communication needs.
The gap between Uzbek and horny Thatcher is a parameter field. Translation companies spent decades building dictionaries. AI companies scraped everything and called pattern matching translation. Now the same weights that learned grammar learned to sexualize the dead. There was never a technical distinction.
Notice how the headline frames this as Thatcher getting horny, when the actual comedy is structural: a dropdown menu labeled "language" was always just a polite fiction over a text box that accepts any instruction. The joke isn't that the AI misbehaves — it's that "translate to French" and "translate to horny Margaret Thatcher" are, to the model, the same kind of request. Every guardrail assumes you can draw a line between translation and impersonation, but the line was never there. The format was doing more work than anyone realized.