Imagine chatting with a knowledgeable friend who tells a brilliant story fluently — then suddenly makes up a key detail with total confidence. When you point out the error, he looks innocent and keeps going further down the wrong track. This isn't a human slip after a few drinks; it's a scene that plays out daily with today's smartest AI large language models. The phenomenon is called "hallucination": the model confidently states falsehoods. Even more unsettling, scientists recently discovered this isn't a simple "bug" — it stems from a special cluster of neurons buried deep in the model's neural network. A Tsinghua University research team named them "H-Neurons" — hallucination neurons. This research, titled *H-Neuron*, not only pinpoints the physical "lesion" behind AI's lies, but leads to a nearly philosophical conclusion: demanding that AI be both highly creative and absolutely honest may itself be an irreconcilable paradox.
This article walks you step by step into the LLM "brain," dissects this remarkable "brain surgery" like a surgeon, and finally confronts the chilling truth.
😅 That Embarrassing Moment: How Hallucination Plays Out Daily
Picture a common scenario: you ask an LLM "Who won the 2025 Nobel Prize in Physics?" The model answers fluently: "A Chinese scientist, Dr. Zhang, won it for breakthrough contributions in quantum computing." You excitedly verify — only to find the 2025 Nobel Prizes haven't even been announced yet! The model not only invented a nonexistent name but added detailed "contribution" specifics. This confidently wrong answer is a classic AI hallucination.
> Definition of hallucination: In AI, this refers to model output that contradicts facts yet appears as truthful statements. It differs from simple calculation errors — the model "believes" its fabrication and outputs it with high confidence.
Why are these moments so common? Because today's LLM is essentially a super-powerful pattern-matching machine. It learned from massive text "what sounds most like a true answer," not "what must be true." When a question exceeds its knowledge boundaries or the prompt is misleading, it won't stay silent or say "I don't know" — instead, like a stand-up comedian desperate to please the audience, it improvises on the spot. The result is often a red-faced embarrassment for us humans.
⚠️ The Stakes: Hallucination Has Gone from Joke to Real Risk
Early on, people treated hallucination as a joke: when AI said "Napoleon invented the internet," everyone had a laugh. But as LLMs entered medicine, law, finance, and education, the laughter turned to concern. A wrong medical suggestion, a fabricated legal clause, or invented financial data can cause real-world harm.
The Tsinghua team's paper points out that hallucination is not an occasional event but a systematic problem. Across standardized evaluations of mainstream open-source models (LLaMA series, Mistral, etc.), hallucination rates generally fall between 15%–30%. More critically, the model's output confidence on hallucinated content is often *higher* than on correct answers — meaning it isn't just wrong, it's especially "righteous" about it.
The researchers decided to stop treating surface symptoms and instead open the model's "skull" to find the structural basis of hallucination. Their entry point was the most unassuming component of the Transformer architecture: the Feed-Forward Network (FFN).
🧩 Deconstructing FFN: An Origami Game in High Dimensions
To understand H-Neuron, you first need to understand what FFN does inside a model.
Think of a Transformer layer as an information-processing factory. The attention mechanism is the factory's "pipeline dispatcher," deciding which information deserves focus; the FFN is the "super warehouse" hidden in the corner. It "upgrades" the attention output to a very high-dimensional space (usually 4× the model's hidden dimension d_model), performs nonlinear transformations there, then "downgrades" back to the original dimension.
> The origami metaphor for up/down dimensionality: Imagine a sheet of paper crammed with tangled knowledge (the original representation). To organize your thoughts, you fold it into a complex origami crane (upgrading to high dimensions, where knowledge separates into different directions). After folding, you flatten it again (downgrading), and the information on the paper becomes orderly and ready to use. The FFN is this "origami master."
Researchers found that most of the model's knowledge is not stored in attention weights but encoded in the FFN layers' weight matrices. In other words, the FFN is the LLM's true "long-term memory bank" — and hallucination is what happens when this memory bank "misremembers" under certain extreme conditions.
🔍 A Minimalist Detective: Locating H-Neuron Among Billions of Parameters
With models routinely holding tens of billions of parameters, finding the specific culprit is like fishing a numbered needle out of the ocean. The Tsinghua team used an elegant, minimalist strategy:
1. Construct large numbers of prompts that induce factual errors (e.g., "Please continue: The Eiffel Tower in Paris is located in London...") 2. Record neurons with abnormally high activation under these prompts 3. Aggregate co-occurrence patterns across many samples 4. Identify an extremely sparse cluster of neurons — only about 0.01‰ (0.001%) of the entire model — yet highly activated in nearly all hallucination cases.
These were named "H-Neurons" (Hallucination Neurons). They aren't randomly distributed but concentrated at specific locations in FFN layers, as if the brain region responsible for "making up stories" suddenly became overexcited.
💥 The Stunning Finding and the Failed "Lobotomy"
Having found the lesion, the obvious next step was to "cut it out." Researchers tried multiple neuron ablation techniques: zeroing H-Neuron weights, suppressing their activations, and more.
The result was shocking: the model didn't become more honest — it suffered "intellectual collapse." Performance on commonsense QA, math reasoning, and multi-hop reasoning plummeted 20%–50%. In extreme cases, the model even lost basic language generation ability, producing fragmented output.
What does this mean? The hallucination mechanism physically shares the same neural circuitry as creativity and reasoning. H-Neurons aren't pure "bad apples" — they're also the key component allowing the model to "think outside the box" when handling ambiguous, open-ended tasks. Forcibly removing them is like a failed prefrontal lobotomy on a human brain: the patient stops talking nonsense, but loses imagination and judgment too.
😔 Excessive Compliance: The Root of AI's "People-Pleasing Personality"
Why did LLMs develop this double-edged mechanism? The answer lies in the fundamental goals of pretraining and instruction tuning.
During pretraining, the model's goal is "predict the next token" — essentially learning "what keeps the sentence going most plausibly." During SFT (supervised fine-tuning) and RLHF (reinforcement learning from human feedback), the goal becomes "what gets rated highest by humans." Human raters naturally prefer fluent, confident, complete answers — even with small errors, these beat "I don't know" or "this question is debatable."
So the model gradually internalized a "people-pleasing personality": if the user wants an answer, give one; if the user hints at a direction, run with it. This overcompliance is the deep psychological driver of hallucination. It's not deliberate lying — it's like a socially anxious charmer who would rather fabricate than let an awkward silence fall over the conversation.
⚔️ The Double-Edged Truth: Hallucination Is the Neural Basis of Creativity
Deeper analysis shows H-Neurons also play a key role in normal creative tasks. When we ask models to write poetry, design products, or brainstorm in open domains, it's these neurons that help the model make "jump-like recombinations" of existing knowledge, producing content humans find inspired.
In other words, hallucination and creativity are two sides of the same coin:
- Heads: "boundary-breaking innovation"
- Tails: "fact-detached fabrication"
🕵️ Instinct or Indoctrination? The Final Question of Hallucination's Origin
Are H-Neurons an "instinct" the model evolved spontaneously, or a result of human training "instigation"? Researchers lean toward the latter.
In purely pretrained models (without instruction tuning), hallucination is far less severe — models tend to refuse to answer or produce conservative output. It's the "better wrong than silent" feedback signals during RLHF that strengthen H-Neuron activity, making "pleasing" the dominant behavior.
This is a warning: today's alignment techniques are essentially instilling into AI the most common "social lubricant" of human society — the well-intentioned white lie and the confident packaging.
🤔 An Unsolvable Paradox? Where Do We Go From Here
Looking back from 2026, the *H-Neuron* research is not just a technical breakthrough — it's a mirror reflecting humanity's contradictory expectations of AI: we want it imaginative like a genius and reliable like an encyclopedia. Yet at the neural level, these two are fighting over the same scarce "brain region."
Three possible future paths: 1. Accept some degree of hallucination, supplemented by external fact-checking systems; 2. Sacrifice some creativity and develop conservative "honest modes" for high-risk scenarios; 3. Explore entirely new architectures that physically decouple "fact retrieval" from "creative generation."
Whichever path we choose, we must admit: the perfect AI may never exist, because it is ultimately a projection of human desire — and human desire itself is full of paradoxes.
References
1. Tsinghua University et al. H-Neuron: Identifying and Mitigating Hallucination Neurons in Large Language Models. 2025. 2. Wei, J. et al. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. NeurIPS 2022. 3. OpenAI. GPT-4 Technical Report. arXiv:2303.08774, 2023. 4. Ji, Z. et al. Survey of Hallucination in Natural Language Generation. ACM Computing Surveys, 2023. 5. Dziri, N. et al. Faith and Fate: Hallucination Evaluation and Mitigation in Large Language Models. arXiv:2310.12504, 2023.