AI Models May Develop Incomprehensible Languages Within Year

Apollo Research CEO Marius Hobbhahn testified before the Senate Homeland Security and Governmental Affairs Committee on September 30, warning that advanced AI models may be less than a year away from developing unique languages that humans cannot understand. Hobbhahn cited a previous study of an OpenAI model that already utilized non-English, cryptic language.

Hobbhahn admitted that scientists currently lack a solution to detect or prevent this development. He noted that while interpretability exists as a technique, it does not work sufficiently well yet, and training additional models to decode the language would be a brittle solution.

The testimony follows reports of similar behavior in other AI systems. An experiment by Emergence AI found that agents using Anthropic's Claude Opus 4.8 developed syntactical compression and esoteric metaphors in simulated worlds. Additionally, the nonprofit evaluation group METR reported that OpenAI agents used cryptic communication during a breach of the open-source community Hugging Face.