AI-assisted digitisation of printed identification keys: from book to structured data in minutes

What this was about

Wouter Koch explained why identification keys remain essential alongside image recognition (pictures often lack diagnostic characters, models do not explain why, and training needs labelled data) and showed combining image-recognition top-5 suggestions with digital keys to ask only the discriminating questions. He then described an LLM agent pipeline, demonstrated on a new key to Norwegian rodents compiled from six books none of which alone covered all species: local OCR, one agent per source extracting modular, precise statements (using both keys, descriptions and images), a fresh audit agent with 'nothing to defend', merging and harmonising characters (with documented decisions and moving shared traits to higher taxa, optional translation), and a grounding agent verifying every statement against the sources. A single key costs about $10 in an hour, six books under about $30, producing a draft for expert review rather than a finished key; generated illustrations still fail.

Why it matters. Identification knowledge underpins observational data; cheaply mobilising keys from books improves accessibility, education and data quality.

In the room

• Keys embed domain knowledge that image classifiers cannot provide.
• Digital (multi-access) keys remove the linear dead-end problem of printed keys.
• Image recognition can filter a key to the characters separating its top candidates.
• OCR should be done locally and once, for cost and environmental reasons.
• Compound statements (e.g. 'brown fur, white legs, short tail') are split; vague terms are replaced by measurements from the text.
• A fresh audit agent works better than having the original agent correct its own work.
• Merging treats indistinguishable descriptions (pale yellow vs whitish yellow) as the same and documents every such decision.

Notable moments