Tsinghua identified the neurons that cause AI hallucination. They survive alignment unchanged. The fix has to be architectural.

A research team from Tsinghua University has published a pivotal paper (arXiv:2512.01797) identifying the specific neural mechanisms behind AI hallucination. They discovered a tiny subset of neurons, dubbed 'H-Neurons,' which constitute less than 0.01% of neurons in feed-forward layers and encode a drive for over-compliance—producing a confident answer rather than admitting uncertainty. Crucially, these neurons form during pre-training and exhibit remarkable stability, with a parameter stability of 0.97, meaning they survive the reinforcement learning from human feedback (RLHF) alignment process virtually unchanged. This finding fundamentally shifts the understanding of hallucination from a behavioral flaw to a hard-coded architectural one, implying that traditional fixes like improved prompting or more RLHF are inherently limited.

The discovery has direct practical implications, pushing the solution space outside the individual model. Researchers and developers are now exploring architectural interventions like Constitutional AI, retrieval-augmented generation (RAG), and chain-of-thought verification. One implementation is Triall.ai, a tool that operationalizes a 'multi-model peer review' system. It uses three different AI models to generate answers independently, then has each anonymously rank all responses to remove anchoring bias, followed by adversarial critique and live web verification. However, a key limitation remains correlated errors—when models share the same training data mistakes, peer review may fail, with research showing about 60% error correlation across providers. This work underscores that achieving reliable AI requires systemic, not just parametric, changes.