Responsible enrichment · Field guide 05
AI photo tagging needs a review system.
The useful question is not whether a model can generate tags. It is whether the archive can preserve provenance, route uncertainty and correct mistakes without obscuring what a person actually confirmed.
Keep three states distinct
Every AI-assisted field should have an explicit state: suggested, reviewed or rejected. A suggestion is not a fact. A reviewed value should identify the reviewer and date. A rejection is valuable operational data because it reveals where the model or workflow fails.
Retain the original metadata separately. If an imported caption says “elk survey, western unit,” an automated system should not silently replace it with a more confident-looking species label. Enrichment belongs beside the source until a reviewer approves a working value.
Match review effort to consequence
| Use | Risk | Review approach |
|---|---|---|
| Broad visual browsing | Low | Suggestions may aid internal discovery if clearly labeled |
| Species or habitat description | Medium | Qualified staff confirm uncertain or important identifications |
| Public caption | Medium to high | Editorial review checks accuracy, context and credit |
| Sensitive-species location | High | Restrict access; do not expose coordinates through generated tags |
| Scientific or enforcement decision | High | Use an approved domain workflow with documented validation |
Design the review queue
Prioritize records with low confidence, disagreement between sources, high-value collections or high-consequence uses. Group similar images so reviewers can work efficiently, but keep the ability to correct individual records. Show the image, source metadata, proposed value, confidence, model provenance and any related sequence.
“Human in the loop” is not a complete control. NIST’s AI Risk Management Framework emphasizes defined roles, documentation and ongoing risk management. Reviewers need authority, training, time and a clear standard for approval. They also need a way to challenge the system rather than merely rubber-stamp it.
Measure the right outcomes
- Acceptance rate by field and collection: broad averages can hide weak categories.
- Correction and rejection reasons: taxonomy mismatch, visual ambiguity, domain error or missing context.
- Override rate: how often people change accepted suggestions later.
- Coverage: what portion of the archive receives a usable reviewed value.
- Time saved: compare against a human baseline for the same task.
- Downstream harm: track incorrect publication, exposed restrictions and search failures.
Preserve provenance for future reprocessing
Models, vocabularies and agency needs change. Record which model produced a suggestion, its version or service date, the input asset, any threshold, the vocabulary offered and the review outcome. Then the agency can reprocess suggestions without confusing old machine output with confirmed metadata.
AI should make uncertainty easier to manage, not harder to see.
A minimum policy
Define permitted uses; prohibited or restricted data; required human review; approval roles; confidence handling; retention of source values; model and prompt logging; vendor data-use terms; correction workflows; performance monitoring; and an off-switch. Revisit the policy when models, collections or uses change.
This framework is for archive workflow design. Agencies should apply their own privacy, security, records, scientific and procurement requirements.