Our organization exists to advance education, research and public understanding of artificial intelligence. That makes ReefBeacon, our AI-using flagship, a place to demonstrate something we consider foundational: AI earns trust only when it is used where it helps and kept away from where it must not go.
The rule
The rule is simple to state. Numbers that matter are computed by deterministic code from published data; AI is used for what AI is actually good at — orchestrating the daily pipeline, interpreting images, drafting text, answering questions — and never as the calculator.
In practice
Concretely, on ReefBeacon:
- The heat-stress numbers — sea surface temperature, HotSpot, Degree Heating Weeks — come from NOAA Coral Reef Watch’s published satellite products, and the threshold judgements (DHW above 4 alerts, above 8 marks high risk) are plain arithmetic in deterministic code. No model output can move a reef between risk tiers.
- The daily pipeline — reading the data, checking it, grading 65 stations, generating the short assessments that accompany alerts — is orchestrated with AI function-calling: the model coordinates the tools; the tools produce the values.
- Survey-photo interpretation uses AI vision where it genuinely helps — grading bleaching severity across thousands of images — with a confidence level attached to every judgement and low-confidence cases routed to human review.
- Historical questions are answered from a cited event library, and when the library does not hold enough history to answer honestly, the system says so instead of improvising.
Why this matters beyond reefs
For a nonprofit that also teaches AI literacy, this is the lesson we most want the public to internalize: the question is never “can a model do this?” but “should a model be the one deciding this?” A temperature reading is a measurement. A risk tier is a rule applied to that measurement. Neither becomes better because a language model produced it — and both become worse the moment no one can check them. Traceability is not a feature we added; it is the architecture.
Teaching what we practice
The same principle now runs through our curricula and explainers: show learners the whole chain — data, rule, output, doubt — and let them pull on it. If you teach or study AI and want to use ReefBeacon as a classroom case, write to hello@fledged.org.