AI systems can now spot when they're lying in real time, without needing a separate judge model

Researchers built a method that monitors an AI's internal reasoning as it generates answers, checking whether those answers are actually supported by the information the system retrieved. The technique runs fast enough to work during generation and can be verified publicly without revealing the model's weights, meaning deployed systems could prove they're being honest.

Retrieval-augmented generation (RAG) — where an AI looks up facts before answering — was supposed to stop hallucination. It doesn't. The system can still invent details that don't match the retrieved evidence, and until now there was no way to catch it at inference time without running a separate AI auditor, which defeats the purpose. This method spots faithfulness failures by reading the model's own internal activations, which means you can monitor honesty without computational overhead or a second model. The real stakes: if this works in deployment, it moves AI honesty from 'trust the system' to 'the system proves it's honest, and you can verify that proof.' That's a structural change in what's auditable.