Analysis, prediction, and autonomous systems

AI is a method, not an actor. Used carefully it lets ecologists ask larger questions, monitor at higher cadence, reason about systems too entangled to study one species at a time, and run analyses that would otherwise sit waiting for a human. Used carelessly it manufactures certainty where none exists.

Our research is about the careful version — and increasingly about systems that not only analyze, but act.

What we study

• Multimodal models for ecology. Combining satellite imagery, acoustics, environmental DNA, in-situ telemetry, and text — so models reason about ecosystems in something closer to the way ecologists do.
• Ecological digital twins. Simulating ecosystems at usable resolution — for restoration planning, fisheries management, watershed forecasting — while keeping the model honest about what it does and does not know.
• Predictive analysis. Short-horizon forecasts that are useful to working programs — fish stock dynamics, drought stress on crops, deforestation pressure, bloom risk in lakes and coasts.
• Autonomous monitoring systems. AI that watches a stream of evidence and surfaces what changed — a new clearing in a forest, an unexpected vessel pattern, a species heard for the first time in a season — without needing a person at the loop every step.
• Agentic workflows. Chains of small, traceable steps that pull data, run a check, escalate when uncertainty is high, and produce a record a researcher or ranger can re-read. We study how to make these workflows useful without making them opaque.
• Traceability of AI-produced evidence. Every result should travel with the data, model, and method that produced it. Without lineage, an AI claim is a number floating in the air.
• Refusal and uncertainty. Models and agents that decline to act when evidence is thin are more useful than ones that always act. We study how to make refusal a first-class behavior.

How we approach it

We build on open, well-understood model families rather than chasing novelty. We publish methods alongside results. We treat AI as one instrument among many — never the final word on an ecosystem — and we keep the human reachable at every step that matters.

Read next

• Monitoring ecosystems in real time — the time-series these models read from.
• Valuing nature, biodiversity, and community work — where evidence becomes a number.
• Conservation and food security — what good predictions are for.