AI Models Invent Private Languages — And Can't Understand Each Other

AI companies increasingly want multiple AI models to work together as a team — one researches, another writes, a third checks the work. To save time and money, some researchers have them swap compressed internal signals instead of full sentences, the way coworkers might use shorthand notes instead of formal memos. This study tested whether that shorthand is universal. It isn't.

Six AI models were trained separately, each developing its own private way of encoding messages. Then researchers tried every possible pairing — all 30 combinations — to see if one model's messages meant anything to another. Only one pair understood each other perfectly, because they happened to start from the identical setup. A second pair understood each other in one direction but not the other. The remaining 26 pairings failed completely, even with translation software layered on top.

The most striking finding involves something called negative transfer — when old knowledge makes new learning worse, like a bad habit you have to unlearn. One model inherited a messaging system built for a different setup. Its accuracy on a harder task sat at just 17%. When researchers deleted that inherited messaging layer and rebuilt it from scratch, accuracy leapt to 86%. Inherited AI shortcuts can be worse than no shortcuts at all.

So what does this mean for you? As AI assistants multiply — one for email, one for scheduling, one for research — the companies building them have assumed these systems can quietly share notes. This paper suggests they mostly can't yet, and that forcing them to share can backfire. That means more engineering, more cost, and slower rollouts for the seamless multi-AI experiences being promised. The paper is narrow — only 17 states, one setup — so it's a warning sign, not a final verdict.