Refusal-first agents for ecological evidence
Note · published 10 May 2026 · Naturecode Project
A model that always answers is a model that sometimes lies. In ecology, where the cost of a false claim can be a misallocated protected area, a misissued credit, or a misled community, that is a serious problem.
We argue for refusal as a first-class behavior: AI systems that decline to answer when the evidence is thin, and that are evaluated on how well they refuse, not only on how well they perform when conditions are easy.
What refusal-first looks like
- The system reports what it does not know, in language a working ecologist or ranger can read. Not "low confidence" — which observations are missing, when they were last seen, and what would close the gap.
- The system distinguishes no evidence from no signal. An absence in the data is not an absence in the world.
- The system surfaces the lineage of its answer — which observations, which model, which time window. A claim without lineage is not a claim, it is a guess.
- The system is benchmarked against scenarios where the right answer is "I cannot answer this honestly yet." If the benchmark does not include those scenarios, the benchmark is incomplete.
Why this matters now
A great deal of ecological reporting is being produced with AI assistance — habitat maps, species distributions, restoration metrics, fishery assessments. Some of this work is excellent. Some of it confidently reports things the underlying data cannot support.
The fix is not to ban AI from ecological work. The fix is to insist that AI systems used in ecology be evaluated against a higher bar: not only can it answer? but does it know when not to?
Open questions
- How to express refusal without paralyzing decision-makers who need some answer.
- How to measure "calibrated refusal" across very different ecological domains — a model that knows when not to identify a bird should not be judged the same way as a model that knows when not to estimate carbon stock.
- How to keep refusal honest when commercial pressure leans the other way.
We publish, listen, and revise.