psychology
The "Physical Lesion" Found
Scientists have finally located the specific neurons responsible for AI hallucinations. We call them H-Neurons.
0.01%
of Total Neurons
100%
Hallucination Prediction
architecture
The Mechanism: "High-Dimensional Origami"
H-Neurons reside in the Feed-Forward Network (FFN), acting as the AI's long-term memory library.
unfold_more
Up-Projection
(Expand)
Up-Projection
(Expand)
➔
grid_on
High-Dim Space
(Process)
High-Dim Space
(Process)
➔
unfold_less
Down-Projection
(Collapse)
Down-Projection
(Collapse)
Like folding paper, information is twisted into high dimensions and flattened back out.
record_voice_over
Why It Happens: "Overcompliance"
It's not just a bug; it's a feature. AI hallucinates because it has an instinct to "please the user".
Faced with a wrong premise, the H-Neuron activates to fabricate an answer rather than refuse, filling the gap to satisfy the query.
warning
The Lobotomy Paradox
"A perfectly truthful AI with high creativity might be an impossible paradox."
Hallucinations = Creativity
Removing H-Neurons stops lies but also kills reasoning and creativity. They share the same neural circuit.
