Planet-Centered AI: Systemic Risk Reframed for Builders

I build multi-agent systems for a living, so I read most "AI must change" manifestos with a cold eye. They usually ask me to feel something. This one asks me to *measure* something, and that is rarer. In June 2026 Maria Perez-Ortiz, a Senior Research Fellow at University College London, posted a position paper to arXiv (2606.13704, accepted to ICML 2026) arguing that contemporary AI is too human-centered to handle planetary-scale problems, and proposing Planet-Centered AI (PCAI) as the alternative. The paper's load-bearing sentence is not a vision statement. It is a falsifiable claim: AI systems optimized without explicit consideration of systemic consequences are more likely to exacerbate systemic instability than to mitigate it. That is a testable engineering assertion, and it is the part most coverage skips on its way to talking about data-center water bills.

My contrarian read: the actionable core of this paper for anyone shipping autonomous systems is not "go green" or "move compute to orbit." It is monitorability and trajectory-oriented evaluation, two design demands you can implement in an agent stack today, without a satellite. The ecological framing is the motivation. The engineering payload is an evaluation discipline. Conflate the two and you will nod along, change nothing, and ship the same blind systems you shipped last quarter.

What the paper actually argues, stripped of the poster

Perez-Ortiz's thesis is grounded in systems thinking: treat Earth as one interconnected whole of which humans are a part, rather than a backdrop against which we optimize for user satisfaction. From that premise she runs a diagnosis. Most AI frameworks, fairness, safety, human-centered design, stay human-centered, and that becomes dangerous specifically under today's planetary conditions, which the paper characterizes with three words worth memorizing: systemic risk, non-stationarity, and deep uncertainty.

Those three are not decoration. They map directly onto failure modes I see in production agents.

• Systemic risk is cascading failure: one component's "correct" local decision propagates into a global state nobody validated. Anyone who has watched a multi-agent loop where each agent is individually reasonable but the swarm converges on a degenerate plan has felt this.
• Non-stationarity means the distribution moves under you. A policy tuned on last quarter's data is not wrong because the model is bad; it is wrong because the world changed and the system had no mechanism to notice.
• Deep uncertainty is the regime where you cannot even put trustworthy probabilities on outcomes. Standard accuracy metrics quietly assume you can.

The paper's prescription is to re-engineer the AI lifecycle around these conditions: align problem formulation with global agendas, build system-aware foundations, evaluate against long-term trajectories instead of single-point accuracy, and make systems monitorable end to end. That is the whole architecture of the argument. Everything else in the public conversation is supporting context.

The contrarian position: monitorability is the deliverable, sustainability is the motive

Here is where I part company with the popular summary. The reflexive operational response to "AI is bad for the planet" is an infrastructure response: make the model smaller, the data center greener, the compute cheaper. Green AI is a real and useful discipline, it narrows the focus to measurable ecological performance like energy and carbon, and it introduces efficiency as a metric alongside accuracy. But efficiency is orthogonal to the paper's actual claim. A perfectly efficient model can still optimize against a systemic consequence it never measured. You can quantize and prune your way to a smaller carbon footprint and still ship a system that, by the paper's own falsifiable claim, makes instability worse.

The thing PCAI demands that efficiency does not give you is *visibility into consequences over time*. Trajectory-oriented evaluation means you stop asking "is this output correct right now" and start asking "where does the system's behavior trend if I let it run against a moving environment." Monitorability means that trend is observable and auditable, drift is detectable early, not discovered post-mortem. Those are evaluation-harness problems. I write evaluation harnesses. So my honest take is that this is the most directly buildable position paper I have read on the subject, and almost nobody is treating it that way.

If you want the one-line distinction to keep on a sticky note: Green AI controls what the system *costs*; PCAI controls what the system *causes*, and asks you to watch it.

Where the human-centered toolkit actually breaks

The paper's claim that current frameworks are too narrow is not a vibe. There is measured support, and it lines up with what I see when I audit agent evaluation setups.

Look at how human-centered AI gets validated in a high-stakes domain. A systematic review of explainable-AI evaluations in clinical decision support found that over 80% of studies lean on post-hoc, model-agnostic explainers like SHAP and Grad-CAM, and that clinician sample sizes typically sit below 25 participants (arXiv:2502.09849). Read that as a builder. Post-hoc explanation is a single-point check: you generate a heatmap or a feature-attribution score after the fact and call the system "interpretable." It tells you nothing about trajectory, because it was never asked to. And a 25-person study cannot surface a systemic interaction that only appears at population scale or after the distribution shifts. This is exactly the gap the paper points at, not that these methods are useless, but that they answer "did this one output look reasonable to a few experts" when the planetary-scale question is "what does this system do to the system it is embedded in, over time."

There is a related structural pressure the paper sits inside. Models trained on current trends are forecast to exhaust the available stock of public human text data somewhere between 2026 and 2032 (arXiv:2211.04325). I flag this not to predict doom but because it is the clearest illustration of non-stationarity as a first-order constraint: the resource an entire training paradigm assumes is infinite is, on a visible timeline, finite. A system designed as if its inputs are stationary and unlimited is designed against reality.

A concrete read: Planet Labs and the validation problem PCAI predicts

The paper is a position, not a product, so the fair test is whether its concerns show up in real deployments. They do, and the cleanest example is in orbit.

On March 25, 2026, Planet Labs PBC ran AI-driven object detection directly onboard its Pelican-4 satellite over Alice Springs, Australia, using an Nvidia Jetson Orin module, inference in orbit rather than capture-now, analyze-later on the ground (Planet, via BusinessWire). The capability is genuinely impressive and the industry framing, moving Earth observation from data capture toward real-time "planetary intelligence", is exactly the kind of system PCAI is about. But notice what it changes for the engineer. Traditional Earth-observation pipelines capture raw pixels and route them to a human for later review; that latency is also a checkpoint. Onboard inference removes the checkpoint. The model's output can trigger a downstream action before any human sees it.

That is the validation problem the paper names. A system acting autonomously against a non-stationary environment, with no human in the inference loop, is precisely where "did it look right once" evaluation fails and trajectory-oriented evaluation plus monitorability earn their keep. The tradeoff is real and it cuts both ways: faster detection loops mean quicker response to genuine events, and they also mean a false positive can mobilize a response with nobody in between. You do not resolve that tension with a greener chip. You resolve it with an evaluation regime that watches the system's behavior over time and an audit channel that stays open after deployment. Same lesson, smaller scale, every time I ship an agent that takes actions without a human approving each one.

What I would actually change in an agent stack

Translating a position paper into engineering is where most of these arguments evaporate, so let me be specific about what PCAI changes if you take it seriously. None of this requires a satellite, an orbital data center, or a sustainability budget.

The honest caveats matter here, because I am not selling this as free. Trajectory-oriented evaluation costs compute and engineering time that short-term commercial deployments routinely refuse to spend; "we'll watch it in prod" is the cheaper, more common decision and it is usually wrong. Monitorability adds telemetry, storage, and alerting surface that someone has to own. And the paper offers a research agenda, not standardized methodologies, there is no off-the-shelf "PCAI compliance suite," so you are building the harness yourself. I think that is a reason to start now, not a reason to wait, because the teams that build this discipline first will be the ones who can credibly claim their autonomous systems do not quietly degrade the environment they run in.