AI models can't report their own inner states
Eight open-weight AI models tested across 78,000 probes couldn't reliably detect when their own computations were altered.
Eight open-weight AI models tested across 78,000 probes couldn't reliably detect when their own computations were altered.