LLM classification is ending; guided hallucination takes over

The limit of fixed vocabularies in content tagging

Developer Simon Willison has outlined a problem everyone encounters when trying to use LLMs for data organization: if you give a model two thousand tags in a prompt and tell it to "pick the right ones," it starts making mistakes. Models struggle to maintain attention over massive lists, often ignoring them or inventing their own categories. Willison proposes flipping the approach: let the model freely generate tags, and only then algorithmically map them to the existing database.

Probability wins over rules

This approach addresses a fundamental misunderstanding of how language models operate. Traditional engineering tries to force LLMs into a narrow corridor (choose from A, B, C), which goes against their statistical nature. Willison's method uses the model's strength (context-based association generation) and compensates for its weakness with classic code (vector search for the nearest existing tag). For developers, this means the end of complex prompts attempting to enforce strict formatting.

Grounding moves to post-processing

The critical point of this proposal is the reliability of the subsequent matching. If the model hallucinates a term that is semantically too far from the existing vocabulary, the vector search might assign nonsense. The system thus changes the problem from "the model ignores the list" to "the model hit the wrong vector." Without a human arbiter or a well-tuned similarity threshold, the error simply shifts one layer downstream.

When this becomes a native feature

It will be crucial to watch how API providers react to this pattern. Today, this requires external embeddings and custom logic. The proof of maturity will be when companies like OpenAI or Anthropic offer a hybrid endpoint that handles this two-phase operation in a single call, natively solving the problem of large classification tasks.

Lilith's verdict

Trying to bind a language model with a fixed list is like forcing water to flow uphill. Engineers have finally realized that instead of building better dams, it is more efficient to put a collection tank at the bottom of the hill.