Algospeak: How Language Became a Hide-and-Seek Game Between Humans and AI
Imagine you and a friend standing inside a transparent glass box, surrounded by extremely smart but extremely rigid "robot police." They hold a thick dictionary of banned words—the moment you say one, alarms blare and you get kicked out.
To keep talking, you must invent a new way to communicate. You can't say "money," so you say "soft sister coins." You can't say "death," so you say "accepting the bento" or "becoming a star." You can't say "violence," so you say "physically redeemed."
As long as your friend understands you and the robots don't, you win. This is "Algospeak."
In May 2026, a Stanford arXiv paper ("Algospeak, Hiding in the Open: The Trade-off Between Legible Meaning and Detection Avoidance") revealed a profound truth about language evolution: our language is becoming a hide-and-seek game between humans and AI.
What Is Algospeak?
Author Jan Fillies argues that Algospeak is a very special form of communication. Unlike traditional encryption—designed so that *nobody* understands—Algospeak aims to hide in the open: instantly comprehensible to all humans, while leaving AI moderators baffled.
The Language Tipping Point: The MUM Threshold
The paper's most striking finding is a tipping point called MUM (Majority Understandable Modulation). Consider the dynamic:
1. Mild variants: You alter a word slightly. Humans see through it instantly, but AI easily catches it via pattern matching. 2. Moderate variants: You use emerging slang that human netizens intuitively understand but newer AI moderators haven't learned yet. 3. Heavy variants: Words become unrecognizable, even replaced by strings of seemingly meaningless memes.
This creates a brutal trade-off:
- Modify too little, and AI catches you every time (evasion fails).
- Modify too much, and your friends can't understand you either (communication fails).
A Game With No End Point
As Feynman noted, solving one problem often creates new ones. Algospeak evolves in an escalating spiral:
1. Humans invent a new word. 2. AI learns it and adds it to its dictionary. 3. The word dies; humans invent a "new-new" word.
The paper points out that as LLMs' comprehension improves, this evolution accelerates exponentially. It's not just vocabulary change—it's an offense-defense battle over semantics and common sense. Humans exploit AI's lack of "subcultural context" and "real-time emotional connection" to build invisible defenses into language itself.
Why Does This Matter?
This is more than netizens playing games on social media.
The paper reminds us that the nature of language is changing. Language used to exist for *expression*; now, to some degree, it exists for *evasion*.
When algorithms shape our habits of expression—when our dictionaries fill with "unalive," "seggs," and pun-based workarounds—we pay a hidden cost: clarity of communication and depth of thought are being eroded.
To summarize:
We whisper in the sunlight, building a digital underground out of code only our own kind understands. AI becomes ever more like the erudite but rigid scholar—it guards every rule yet cannot grasp the winding turns of human emotion. Meanwhile, we humans, through our endless "nonsense," preserve a sly vitality that algorithms cannot replicate.
The next time you see baffling slang, don't roll your eyes. It may not be posturing—it may be humans waging a tacit guerrilla war against AI to defend the right to speak.
The most beautiful language is often the kind where "I get it, you get it, but it doesn't."