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:
What patterns can humans create that machines cannot, and why does this matter for environmental knowledge?
We look at the habits people can cultivate that help them respond to environmental uncertainty—without outsourcing judgment to AI.
When I was a kid, on those warm summer nights, I loved to watch the progress of little lightning bugs (or fireflies, depending on where you’re from) as they blinked their way across the yard. I didn’t have as much time to notice them as I grew and life got busy, but it still became increasingly clear that the lights were slowly going out.
A machine learning system could probably detect the same pattern (using different methods), identifying a decline in lightning bug populations over time. It could also offer explanations, citing light pollution levels or pesticide use. But the pattern would not matter to it. It would not be anchored in memory, or tied to a place, or carry the urgency of noticing something you love vanishing. The difference isn’t just in how patterns are detected, but in how they connect to consequence and to the possibility of response and repair.
Fortunately, my partner’s and my backyard is ripe for testing that response and repair. A few simple shifts can help restore lightning bug habitat: never use pesticides (please!), leave the leaves (the larvae thrive in leaf litter), and keep ambient light levels low (or completely off). While pesticides were never part of our routine, we started pitching the yard into darkness each night. Instead of raking leaves, we let them cover the detritus of plants and soil, and even though it left our neighbor with a few complaints, it created a place for larvae to rest. As we tried small changes while slowly turning the yard back into habitat, I began to notice an increase in the blinking pattern. Scientists might use transects or 3D imaging to monitor lightning bug populations, but you can also simply sit in your backyard and count the flashes. After what felt like relatively few modifications, the blinking didn’t just return, it came roaring back.
I was able to understand this pattern because of my humanness. I had decades of casual observation behind me, memories of the sensory experience of seeing lightning bugs (that gasp when you spot your first one of the year), and enough familiarity with places that I could influence—in this case, my backyard—to interpret what I was seeing and test ways to respond. Put simply, I could take multiple indicators, make sense of them, and do something about it.
In reading What Is Intelligence? this week, I realized that what I was doing in my backyard doesn’t quite fit the dominant definition of intelligence, which treats it as solving a bounded problem with the right inputs and outputs. As one researcher puts it, that definition reflects “a linear, deductive, mechanistic view.” Watching the lightning bugs wasn’t linear, it was relational. It unfolded over years, through memory, and through sustained attention to a pattern that returned each season. Another way of understanding intelligence, one that feels closer to this experience, is as participation in a living system. As a participant in this ecosystem, recognizing the pattern wasn’t separate from responding to it. The pattern only meant something because it was tied to a world I am part of and can affect.
Machine-learned patterns emerge very differently. They are the result of billions, or trillions, of tiny adjustments during training, as models are taught to predict what comes next. Researchers increasingly argue that to understand these systems, you have to look at their training histories: the initial conditions, the data they were exposed to, and the variations that shape how they generalize. But even here, there’s a gap. While both humans and machines are shaped by their histories, only humans must live within the consequences of how patterns are recognized, misrecognized, and acted upon. My observations of fireflies were shaped by years of exposure, but also by the fact that I live in that environment and that my actions feed back into it, and its changes feed back into me.
I’m not an ecologist, and while I understand the role of ecology in the complex history of environmental pollution, I feel more comfortable talking about patterns as a way of understanding the ecosystem changes that pollution can produce. The lightning bug story is a microcosm, but if we return to my last OEDP AI fieldnote, living with environmental pollution shows how clearly patterns can point to system-level consequences. In refinery communities, people learn to read patterns such as the timing of flares lighting up the night sky, how smells travel and dust settles depending on the wind direction, which days children stay home from school, or the shift in work and income when health falters. Examples like this point to the importance of interpreting patterns collectively. Humans have long traditions of doing this through almanacs, journals, and stories: ways of not just tracking patterns, but remembering them, debating their meaning, and passing them forward.
Both humans and machines, in different ways, are products of their histories. For AI systems, researchers point to training data and early conditions as decisive, with small variations steering a model’s behavior in lasting ways. But human pattern-making is not just historical, it is situated. It is built through ongoing relationships with place, memory, and consequence. It is not just that we have histories, it’s that we continue to live inside them.
This is why protecting human pattern-making matters for environmental knowledge systems. If we design tools that prioritize only the kinds of patterns machines can extract, we risk eroding the very capacities that allow people to notice what’s changing, interpret why it matters, and decide what to do next. While we were building DIGITCORE, a digital toolkit built around patterns, we found that meaningful patterns emerged not only from conversations, interviews, and working groups, but from testing what we heard against our own experience. Work like DIGITCORE, at its best, should continue to ask a different question: how do we support the conditions under which people can continue to make sense of their environments?
AI cannot know how many lightning bugs filled warm summer nights in the 1980s (unless it was recorded and digitized). It cannot feel when something is “off,” carry baselines across generations, or weigh patterns against what it would mean for the lightning bug to blink out entirely. And it certainly cannot experience the feeling that makes us audibly gasp as the first lightning bugs blink in the summer night.
