AI usage
Scroll to the bottom to view the AI usage log.
What AI does here
Two things, both about tidying up other people's reports: it helps
decide when two sightings (one from Acartia, one from iNaturalist) are
the same animal so they show as one marker, and it reads killer-whale
descriptions to estimate the ecotype (Bigg's or Southern Resident).
When AI is used
Most reports never go near AI. For ecotype labels, keyword rules run
first: they read the observer's description for a pod name (J-pod, a
T-number) or words like "Bigg's" or "resident." The model is only
asked when they find nothing, and only for killer-whale reports with a
written description.
For matching, a distance-and-time check flags same-species reports
close in place and time (for an orca, within 2 km and an hour) as
possibly the same animal. Only those pairs go to the model, which
confirms or rejects each. It can drop a flagged pair or join a
borderline one, but never merges different ecotypes like Southern
Resident and Bigg's. Without the model, close pairs merge, borderline
ones don't. Every answer is remembered, so nothing is judged twice.
Which model, and what's sent
The site uses Anthropic's Claude (the Haiku model) through the Claude
API. For matching, it sends both reports' description, species, count,
coordinates, timestamp, and contributor name. For ecotype reading, it
sends only the species and the observer's description. Nothing about
you, the viewer, is ever sent. Observers' words are always shown on
the map exactly as written.
Where it happens
All of this runs server-side, in the pipeline that prepares the map's
data. Your browser never contacts the AI.
Read-only, and no training
The model only reads a report to reach a conclusion. Nothing here
trains a model, and Anthropic does not train on the data the site
sends.
How to tell when AI was involved
A label reads "(inferred)" when a plain keyword match produced it, or
"(AI-inferred)" when the model did. A merged marker says "match
confirmed by AI" when the model made that call. Anything the model
concludes is shown as this site's inference, never the observer's, so
treat an ecotype label as a hint, not a fact.
If the AI is unavailable
The map still works. The plain rules run on their own, some duplicates
may show as two markers, and some ecotype labels are simply left off.
Nothing breaks.
Recent AI decisions
The AI-assisted calls behind the sightings shown right now. It updates
as the map refreshes and covers only what's currently on the map.
- No AI-assisted decisions among the sightings shown right now.