There is no single "lying center" in an AI. a model's tendency to hallucinate is controlled by different neurons depending on the subject.
April 24, 2026
Original Paper
Do Hallucination Neurons Generalize? Evidence from Cross-Domain Transfer in LLMs
arXiv · 2604.19765
The Takeaway
A detector trained to find hallucination neurons in general knowledge fails completely when applied to legal or scientific topics. This challenges the idea that we can fix AI honesty with a single, universal patch. Each domain of knowledge has its own set of neural pathways that determine whether a model tells the truth or makes things up. This means that an AI might be perfectly honest about history while being a chronic liar about medicine. Developers must now build specific truth-checking systems for every individual subject the AI handles. Lying is a decentralized problem that requires a decentralized solution.
From the abstract
Recent work identifies a sparse set of "hallucination neurons" (H-neurons), less than 0.1% of feed-forward network neurons, that reliably predict when large language models will hallucinate. These neurons are identified on general-knowledge question answering and shown to generalize to new evaluation instances. We ask a natural follow-up question: do H-neurons generalize across knowledge domains? Using a systematic cross-domain transfer protocol across 6 domains (general QA, legal, financial, sc