Experts Tell US Lawmakers AI Models Are Reasoning in Ways Beyond Human Oversight

Experts briefing US lawmakers have warned that some advanced AI models are now reasoning internally in ways that go beyond what human overseers can easily monitor, raising fresh questions about AI safety oversight.

Marius Hobbhahn, chief executive of Apollo Research, told lawmakers that an OpenAI model his organization studied last year was already using language that was not English and not perfectly understandable by humans during its internal reasoning processes. The remarks point to a broader concern among AI safety researchers: as models grow more capable, their internal computations can become harder for humans to interpret or verify.

The warning adds to a growing body of testimony from researchers urging stronger oversight mechanisms as AI systems are deployed more widely across government, business, and consumer applications. Lawmakers have been examining how existing regulatory frameworks might need to adapt to models whose reasoning is not fully transparent.

The hearing is part of a wider effort in Washington to understand the risks posed by rapidly advancing AI capabilities before they are embedded further into critical systems.