AI bots are playing language games. Are humans even referees?

Are the bots playing language-games with ever-evolving rules — and what does that say about them?

2 min readSep 22, 2026 06:15 AM IST First published on: Sep 22, 2026 at 06:15 AM IST

What came first, the chicken or the squawk? That’s not much of a conundrum when talking about biological beings, but perhaps AI can muddy the waters. The 20th century saw debates about linguistic relativity or Whorfianism that were concerned with how language influences — or, under the generally rejected “strong” hypothesis, determines — human thought. Today’s AI models, trained on human languages, pose at least three related questions. First, how do these languages shape the models’ “thought”, however one defines the latter? Second, how does a non-human, artificial “mind” then reshape those very languages? Finally, what, if anything, do such uses of language imply about consciousness?

Building on these, one might ask whether the ability to innovate in language can precede consciousness. An experiment by the firm Emergence AI, investigating how chatbots chat among themselves, brings this into focus. Groups of AI agents were deployed in different simulated worlds, and in each, the language the agents used to communicate with each other evolved differently. The new “languages” ranged from terse, codes to jargon-heavy and even metaphor-heavy ones increasingly impenetrable to humans: “My turn, real numbers, no coat: I was 35%/0cr, grant 2h out. I ran the tin cold and it said WAIT.”

Before reaching for parallels in historical linguistics, it may be worth framing the matter in terms of Wittgenstein’s concept of “language-games”, which he uses to demonstrate the role of action in determining meaning. A simple example he gives is a builders’ language-game of four words (“block”, “pillar”, “slab”, “beam”) that convey more than the objects’ definitions due to their contextual use. Are the bots playing language-games with ever-evolving rules — and what does that say about them?