This post is part of a series titled “OEDP’s Fieldnotes on AI”, where we offer reflections about AI in environmental and participatory contexts. This week’s reflections center on the question:
When does AI meaningfully help, and when does it increase risk?
We look at the specific situations where AI could reduce burden, increase access, or improve decision-making without increasing risk or dependency. We also examine where AI actively degrades judgment, care, and accountability.
During a grant panel years ago, I read, alongside the other reviewers, what initially seemed like an outstanding proposal: an extensive, engaged project to monitor air quality in an apartment complex. Researchers would work alongside residents to collect and interpret data about the air they were breathing. On its face, it was exactly the kind of community-centered work I’m drawn to. It’s also the kind of work that AI systems, particularly those used for prediction, pattern recognition, and decision support, are increasingly asked to shape.
But as I worked through the proposal, I kept getting a familiar nudge. The project was rich with storytelling about the complexities of living at the site and the importance of knowing what was in the air. Yet it missed a critical part of the story: what happens if the air turns out to be bad? Where were the resources – social workers, medical professionals, health care systems – or at minimum a hint of follow-up that would support the next step in the equation, the move from “my air is bad” to “and now I’m going to do something about it”?
This is a story I’ve encountered again and again. In the early days of the BP oil spill, many residents were wary of learning about potential contaminants in their environment because they lacked health insurance or access to medical care (even community health clinics). A colleague, who recently wrote the excellent book Homesick, traced FEMA trailers from the post-Katrina Delta to places such as the oil fields of Arkansas. The trailers contained high levels of chemicals, such as formaldehyde, but post-Katrina they were already temporary housing for people displaced from their homes. Situated in emergency shelters, once residents learned they were breathing toxic chemicals, where else could they go?
And though many assume it is better to just know, knowing can be terrifying when few resources are available. Where do you go from there?
Reading a ProPublica piece on the use of machine learning models to investigate where Ebola outbreaks might occur brought these earlier experiences back to mind. It underscored how important it is to place clear boundaries around the data we collect and the claims we make from it. In that article, the value of AI lies in surfacing patterns of risk, showing where communities might be vulnerable and how risk might shift due to factors like forest fragmentation or increased human-wildlife contact. In that sense, AI can reduce analytical burden and expand visibility, especially in contexts where resources for monitoring and analysis are limited. The work helped point to where attention should be focused.
But as we know from environmental justice and public health, data and modeling do not substitute for political will. Even when we can identify what is wrong, or trace causal pathways that suggest where intervention would help, we have to be careful not to conflate knowing with doing. When AI systems accelerate the production of knowledge without expanding the capacity to act, it can deepen harm by surfacing risks that people are structurally unable to respond to.
So it is unsurprising that while AI might speed up our interpretation of risk landscapes, building on what air quality monitoring researchers once did with residents of that housing complex, it offers no new method for confronting the political, ethical, and economic realities of environmental health. In situations like these, AI cannot be treated as neutral infrastructure. It is infrastructure deeply embedded in existing power differentials surrounding environmental harm.
Seen this way, AI makes clear that data stewardship is less about maximizing insight and more about holding together knowledge and responsibility. In environmental health contexts, ethical failures are rarely the result of too little data. They arise instead when knowledge is produced without protection, prediction outpaces political will, and visibility ends up leaving people with no recourse. AI can help us see patterns, but stewardship begins where modeling ends.
