At 6:41 this morning, as I sat with coffee in my kitchen in Oregon City, a Golden-crowned Sparrow began singing outside my window. My BirdWeather station logged the detection automatically — timestamp, species identification, confidence score. Two minutes earlier it had caught a Dark-eyed Junco. Ten minutes before that, an Anna’s Hummingbird. By the time I opened these two papers, the Macroscope dashboard on my wall showed five species detected, part of a running weekly total of forty-eight species and a life list, accumulated since October 2024, of one hundred and twenty-nine. The system monitors continuously, logging dawn chorus intensity, computing environmental correlations, flagging anomalies against learned baselines. I was drinking coffee. The observatory was already at work.

I mention these numbers not as a boast but as a baseline, because the question that emerged from this morning’s reading is whether baselines themselves are shifting beneath us — and whether the tools we have built to observe the natural world are adequate to the speed of its unraveling.

Thirty Years of the Same Question

In 1997, I compiled Breeding Bird Survey data for eleven species of cavity nesters and old-growth specialists documented as breeders in the Black Mountain Scenic Area of the San Jacinto Mountains. Mountain Chickadee, Brown Creeper, Downy Woodpecker, Northern Flicker, Dark-eyed Junco, Purple Finch, Northern Pygmy Owl, Red-breasted Sapsucker, Pine Siskin, Violet-green Swallow, Western Wood-Peewee — all of them dependent on the structural complexity of mature mixed-conifer forest, and all of them showing thirty-year downward trends in the BBS data. I was making the case that logging of snags and old-growth trees was eliminating critical nesting and feeding habitat, and that the population declines were the predictable consequence.

I was, at that time, director of the James San Jacinto Mountains Reserve, a University of California Natural Reserve System field station perched at 5,400 feet in the mountains above Palm Springs. I knew these birds by their voices, their behavior, their nesting phenology. I had watched the Brown Creeper spiral up bark-furrowed trunks of white fir trees that were older than the California state constitution. The BBS data told the continental story; I was living the local chapter.

Nearly three decades later, a paper published last week in Science by François Leroy and colleagues has confirmed what field ecologists like me have felt in our bones for a long time, and added a mathematical dimension that makes the situation considerably more alarming. They asked a question that, remarkably, had not been systematically addressed at continental scale: is the decline in bird populations itself accelerating?

The Second Derivative

The insight is elegant and sobering. Most large-scale biodiversity studies measure the first derivative — the rate of change in abundance over time. Populations are declining. We have known this since at least the landmark Rosenberg paper in 2019, which estimated a loss of roughly three billion birds from North America since 1970. But Leroy and colleagues looked at the second derivative: the rate of change of the rate of change. Using 1,033 BBS routes, 261 species, and thirty-four years of data from 1987 to 2021, they employed dynamic N-mixture Bayesian models to estimate not just whether populations were declining, but whether the decline was speeding up.

It is. Of the 261 species they analyzed, 122 showed significant declines. Of those, 63 — more than half — are declining at an accelerating rate. A quarter of all species studied are in what amounts to accelerating freefall. The gap between the number of individuals lost and recruited widens each year.

The geography of this acceleration is as important as the fact itself. The hotspots of decline coincide with warm and warming regions — the southern tier. But the hotspots of acceleration map onto a different variable: agricultural intensity. The Mid-Atlantic states, the Midwest corn belt, and California — regions with high fertilizer use, heavy pesticide application, and extensive croplands — are where the rate of loss is itself increasing. My old study area in the San Jacinto Mountains sits squarely in one of those acceleration zones.

Meanwhile, here in Oregon City, the picture is different. Leroy’s smoothed maps show the Pacific Northwest as one of the few regions with a positive change in growth rate — meaning a deceleration of decline. The juncos and flickers and creepers I hear every morning on my walks through Waterboard Park are consistent with that pattern. Seven of my original eleven Black Mountain species appear on my BirdWeather life list here: Dark-eyed Junco, Brown Creeper, Downy Woodpecker, Northern Flicker, Pine Siskin, Purple Finch, even Northern Pygmy-Owl, first detected acoustically in August 2025. Same species, a thousand miles of latitude apart, with radically different population trajectories depending on geography and land use.

What the Survey Cannot See

There is, however, a caveat that any field station director would notice immediately. The BBS is a beautifully standardized protocol, but its coverage is profoundly uneven. Routes are dense east of the Mississippi and sparse in the West, with enormous gaps in Canada’s boreal forest. Leroy and colleagues correct for detection probability — the chance that a bird present on a route actually gets counted — but the spatial distribution of the routes themselves is a harder problem. Their three acceleration hotspots are also areas with the densest sampling. The deceleration signals in the mountain West and northern Canada may partly reflect the lower statistical power of sparser coverage. The authors apply spatial smoothing to interpolate across gaps, but interpolation is not observation.

This is not a flaw in their study so much as a structural limitation of the BBS itself. It was designed in the 1960s as a roadside survey conducted by expert volunteers, and it remains the best standardized avian monitoring dataset in the world. But the world has changed. We now have tools that can fill the gaps the BBS cannot reach — tools that monitor continuously, that do not require expert observers, and that are deployable at scales from a single backyard to a continent.

Death by a Thousand Cuts

There is another dimension to the decline that continental-scale analysis cannot capture. This morning I took a photograph out my kitchen window: a mossy stone wall, a bird feeder on a pole, raised garden beds coming back to life in early March — and a tuxedo cat sitting calmly on the flagstone path, positioned squarely between the yard and the corridor the birds use to approach the feeder. The predator and the attractant in the same frame.

Leroy and colleagues identify agricultural intensity as the dominant driver of acceleration at continental scale. But at the backyard scale, songbirds face a different arithmetic of attrition. Free-roaming domestic cats kill an estimated 1.3 to 4 billion birds annually in the United States alone, according to a widely cited 2013 study by Loss, Will, and Marra in Nature Communications. Add window collisions — another billion. Add residential pesticide use that eliminates the insect prey base. Add light pollution that disorients nocturnal migrants. Add habitat fragmentation that turns contiguous forest into isolated patches too small to sustain breeding populations. No single one of these factors explains the continental signal, but cumulatively they erode populations from below while agriculture and climate push from above. It is death by a thousand cuts, and many of those cuts happen in places that look exactly like my backyard — places where people genuinely love birds and inadvertently participate in their decline.

This is precisely the scale at which data-driven community stewardship could make a difference. A BBS route cannot tell you that your neighbor’s cat is suppressing nest success in your block’s hedgerow. But a network of BirdWeather stations could detect the absence — the anomalous silence during fledging season, the species that should be present but are not. And a human observer, prompted by that data, can identify the cause and advocate for change.

Aeroplanes for the Mind

Which brings me to the second paper on my desk this morning. In a commentary published in Nature this week, Dashun Wang proposes five principles for scientists using AI agents, built around a central metaphor: if Steve Jobs was right that computers are bicycles for the mind, then AI agents are aeroplanes — faster, more powerful, but harder to control and catastrophic when they crash.

Wang’s framework emerges from his team’s development of SciSciGPT, a multi-agent system in which specialized AI agents divide and coordinate research workflows. His principles are sensible: collaboration beats automation, speed itself is transformative, agents should specialize to domains, trust must be engineered through provenance and transparency, and AI will create a new kind of scientist who learns from machine reasoning rather than merely deploying it. He draws a lovely analogy to Newton’s prism experiments, arguing that what Newton did — recognizing that an apparent anomaly was actually the phenomenon — represents exactly the interpretive leap that automated systems are designed to smooth away.

I find his argument largely correct but oddly confined. Wang imagines AI agents operating within the existing institutional structure of academic science — labs, departments, journals, funding agencies. He calls for coordination between scientific societies, funders, and AI labs. What he does not imagine is the distributed, place-based, citizen-powered version of the same idea — which is, I would argue, what ecological monitoring actually needs.

Oregon Regional Ecological Observatories

Consider what already exists. My Macroscope station in Oregon City runs a BirdWeather PUC that identified 48 species this past week and generates continuous acoustic detection data around the clock, year-round. A Tempest weather station feeds temperature, humidity, barometric pressure, wind, and precipitation into the same data stream. An automated pattern analysis system computes environmental correlations — the near-perfect inverse relationship between temperature and humidity, the weaker but ecologically meaningful correlation between temperature and bird activity — and tracks dawn chorus intensity as a metric of ecosystem health. A system I call SOMA, built on stochastic modeling rather than rule-based logic, monitors for anomalies across the sensor mesh and flags events that deviate from learned baselines.

This is one station, in one backyard, operated by one retired ecologist. Now multiply it.

What I have been calling Oregon Regional Ecological Observatories — OREO — is a vision for a loose confederation of natural area stewards, each with their own sensor infrastructure, federated through AI-mediated analysis that discovers cross-site patterns none of them could detect alone. Not a new NEON or LTER network, which would require tens of millions in federal funding and years of bureaucratic construction. A confederation. A BirdWeather PUC here, a weather station there, an iNaturalist observer documenting flowering phenology at a third site, a school garden monitoring program at a fourth. Each site gets its own virtual monitoring agent — the kind of domain-specific AI that Wang calls for — that learns local baselines, detects anomalies, and generates specific field protocols for human observers to ground-truth.

The architecture we are building in yea.earth is designed for exactly this. Each curated location receives an ecological address — coordinates, elevation, ecoregion, climate classification, expected species. Virtual agents monitor incoming data streams and discover what I think of as SOMA-style interrelationships: correlations between weather patterns and bird activity, between phenological timing and acoustic detections, between land-use changes at regional scale and population dynamics at local scale. The AI does the continuous watching. The human goes outside and looks.

This is Wang’s pilot-in-command principle, but distributed across a community of stewards rather than concentrated in a single laboratory. And it directly addresses the spatial bias that limits the BBS. Citizen science networks — BirdNET, BirdWeather, FeederWatch, NestWatch — generate data from backyards, schoolyards, nature centers, and private lands that no roadside survey will ever reach. They fill the temporal gaps too: nighttime detections of owls, year-round monitoring instead of a single June survey, always-on acoustic coverage that removes the human skill bottleneck entirely. You do not need an expert who can identify two hundred species by ear. You need a microphone and an algorithm.

The Observer Persists

There is a thread that runs from the mist nets we strung in the mixed-conifer forest at the James Reserve to the BirdWeather PUC mounted on my porch in Oregon City. The tools changed utterly. The observational impulse did not. In 1997, I was compiling BBS trend lines to argue that snag removal was destroying breeding habitat for cavity nesters. This morning, I read a paper confirming that the decline I documented has not merely continued but accelerated, driven by agricultural intensification across the continent. And I sat in front of a monitoring system that represents, in miniature, what the response should look like: continuous, automated, place-based, human-interpreted observation that operates at the temporal and spatial resolution the crisis demands.

Leroy and colleagues conclude that incorporating the second derivative — the acceleration of ecological metrics over time — into conservation assessments could uncover signals of decline that remain hidden when we focus solely on abundance trends. They are right. But detecting acceleration requires sustained, high-resolution monitoring infrastructure, and the BBS alone cannot provide it. Wang argues that AI agents should amplify human scientific judgment rather than replace it. He is right too. But the judgment that matters most in conservation is local knowledge — the naturalist who knows what should be singing in this forest at this season, and notices when it is not.

The future of ecological monitoring lies not in choosing between standardized surveys and citizen science, between expert observers and AI classifiers, between institutional infrastructure and distributed networks. It lies in federating all of them. What I am proposing with OREO is not a replacement for the BBS or for NEON but a complement — a living, growing mesh of observation points where data flows up from distributed sensors, AI handles pattern recognition and anomaly detection, and humans provide the interpretive context that transforms data into understanding.

The Dark-eyed Juncos are singing outside my window. They were singing outside my window at the James Reserve in 1990, too, though those were Oregon Juncos in the old taxonomy and these are of the same subspecies, having followed me — or I them — a thousand miles north. The Macroscope tells me the dawn chorus was strong this morning, fourteen species, consistent with healthy ecosystem function at this location. Somewhere in the San Jacinto Mountains, on a BBS route I once knew by heart, the same species may be losing ground faster than ever. The question is whether we are building the systems to know the difference, in time to act on it.