A RAND led research team argues that borrowing population and ecosystem models from ecology could give AI governance — the rules and oversight for AI systems — a vocabulary for cascading failures, tipping points, and system wide instability that conventional engineering safety…
A research team is proposing to measure AI risk the way ecologists measure animal populations, by tracking births, deaths, diversity, and ecosystem dynamics, and has laid out three such indicators in a peer-reviewed paper. The proposal, published 25 June 2026 in npj Complexity, is the work of researchers at the RAND Corporation, the University of Notre Dame, and Stanford, and it argues that conventional engineering-safety thinking is poorly tuned to the kind of cascading failures, tipping points, and system-wide instability that an AI ecosystem could produce.
The paper, "The ecology of AI risk," borrows its core vocabulary from theoretical ecology, the math ecologists use to model animal populations and ecosystems, and applies it to the question of how AI systems interact, proliferate, and fail. Lead authors include Edward Geist and Alvin Moon of the RAND Corporation, a U.S. policy research group, James Holland Jones of Notre Dame's Department of Biological Sciences, and Anton Wu of Stanford's Doerr School of Sustainability, with Alexander Dolnick Meyer and Aisha Nájera as co-authors.
Rather than treat AI risk as a property of any single model, the team asks how the population of AI systems is changing over time. Their three indicators are derived from population and ecosystem models in theoretical ecology, and the paper sketches out what each is meant to capture. Population-style indicators track births and deaths in the AI population: how many systems enter use, how many retire, and how fast that turnover happens. Ecosystem-style indicators capture diversity and structure: how varied the AI population is, how concentrated capability is across providers, and how dependent systems are on one another. The third set looks at feedback and resilience: how shocks propagate through the AI ecosystem and whether the system can recover.
That reframe matters because most existing AI safety work is built around the engineering question of whether any one system fails in a controlled way. The ecological framing instead treats risk as a system-level property, closer to how ecologists think about invasive species, monocultures, or food-web collapse than to how a safety engineer thinks about a fault in a single machine.
The proposal comes with the usual caveats a framework paper should carry. The authors themselves flag the limits of the analysis. The three indicators are derived from ecological theory and applied to AI risk by analogy. They have not been validated against historical AI incidents, and there is no underlying dataset of AI population dynamics to test them against. The paper is also recent, which means no regulator has adopted the framework and no peer review outside the journal has weighed in on the indicators' practical use. The RAND landing page for the work describes it as a methodological proposal rather than a measurement system.
What the framework does offer is a vocabulary. Cascading failures, tipping points, concentration of capability, and dependency-induced fragility are concepts that map onto ecological ideas like predator-prey dynamics, biodiversity loss, and keystone species. If AI governance is going to talk about systemic risk, it will need a way to name and measure those things, and the paper's argument is that ecology already has the tools.
What to watch next is whether the framework gets traction outside the paper itself. The authors intend it to inform AI governance policy discussions, but adoption depends on regulators, standards bodies, and AI labs finding the indicators usable in practice. Until then, "the ecology of AI risk" is best read as a research-methods proposal with cross-disciplinary pedigree, not as a new way of scoring the danger of AI.