Hallucinations and Mental Health

It’s common to hear that a language model “hallucinates” when it produces false or nonsensical outputs. The metaphor is vivid: the AI is imagined as a fragile mind, wandering in dreams, conjuring phantoms, perhaps even needing therapy.

Charming, but deeply misleading.

---

The Metaphor Problem

• Hallucination implies subjective experience — perception independent of reality.

• Mental health language implies cognition, emotion, and consciousness.

• LLMs have none of these. They generate sequences according to probabilistic patterns, not perception or imagination.

The metaphor frames statistical divergence as an inner psychological event. Users interpret errors as “misperception,” rather than the predictable output of relational constraints applied to tokens.

---

Why This Is Misleading

1. Projects human phenomenology onto algorithms — treating computational patterns as mental states.

2. Obscures relational mechanics — hallucinations are not failures of cognition; they are natural consequences of pattern instantiation.

3. Encourages misdiagnosis — a model does not “see” or “believe” anything; it outputs aligned correlations.

By calling them hallucinations, we import an erroneous ontology of sentient error onto statistical machinery.

---

Relational Ontology Footnote

In relational terms, what is labeled a “hallucination” is an actualisation of potential token alignments outside the constraints of factual accuracy. There is no mind wandering — only relational patterns unfolding under probabilistic rules.

---

Closing Joke (Because Parody)

If LLMs truly hallucinated, your AI assistant would be wandering around the office, describing imaginary colleagues and offering unsolicited existential advice — and yet still forgetting your password.