PLANETARY INTELLIGENCE
Research and Reflection
Bioregional Intelligence in Service to Life
Issue 2 | 6 November 2026 | Newsletter draft
Our first newsletter asked whether planetary intelligence might already be emerging.[1] This second issue brings that question closer to home: can a community become better at understanding its watershed, making decisions together, and correcting mistakes? Bioregions offer a practical setting in which to investigate the larger idea. Recent AI developments make the experiment more feasible. My assessment is that evidence for increasing technical capability remains stronger than evidence for lasting ecological benefit.
For this project, we propose defining bioregional intelligence as the capacity of people and institutions within an ecological region to combine scientific and situated knowledge, make accountable decisions, and learn from their consequences for human communities and other life. A bioregion may be organized around connected watersheds, habitats and cultural relationships; its boundaries need to be justified for the question being studied. Local boundaries must not hide harms displaced downstream or elsewhere.
This is a working definition for research. Frank, Grinspoon and Walker’s planetary framework concerns collective knowledge integrated into the functioning of planetary systems. Our proposed contribution is to investigate one possible pathway toward that integration through decisions that can be observed and tested. An effective watershed institution would not, by itself, demonstrate planetary intelligence or establish that an ecosystem is conscious.[2]
BioFi’s November 3 event brings several approaches into view. Its published brief for “Beyond Extraction” explicitly framed the meeting as an exploratory conversation about what AI can contribute to bioregional regeneration. The projects address different parts of the problem, with different levels of maturity.[4]
| Initiative | Proposed or developing contribution | What remains to establish |
| Planet Labs | Satellite observations and machine learning to track ecological conditions and change. | Whether better observation changes decisions and produces ecological benefits.[4,5] |
| Regen Network | Compass is intended to connect ecological information with bioregional planning and investment. | The quality of decisions and outcomes produced by the developing approach.[4] |
| GainForest | Community observations, biodiversity monitoring, AI tools and funding for stewardship. | Independent evidence of lasting ecological gains and effective community control of data.[6] |
| COBALT | A Living Atlas combining ecological knowledge and governance, initially focused on seagrass restoration. | How its early prototype performs in real decisions and restoration work.[7] |
GainForest offers a concrete governance example: it describes Taina as co-designed with Indigenous scientists in the Greater Manaus region, with options for self-hosted data or governance through a community-elected data council. COBALT explicitly describes its Living Atlas as being in conceptual development and early prototyping. Those details matter: participation, a working prototype and demonstrated ecological recovery are different achievements.[6,7]
My reading of these efforts is that the central challenge lies in connecting observation to action. A map can reveal a damaged wetland; it cannot settle who has authority, who pays for restoration, or how a landholder’s livelihood should be protected. Bioregional intelligence would require institutions able to act on evidence, retain dissent, and revisit decisions when outcomes disappoint.
The Institute for Planetary Intelligence’s November 1 essay separates biological cognition, planetary evolution, environmental computing and collective governance. Possible Planet Lab proposes work on reliable knowledge, collective decisions and bioregional pilots. Together, these supply a research agenda whose ecological benefits remain to be tested.[3,13]
Four recent developments sharpen the opportunity and the risks. They concern what AI can do, how quickly it may improve, how it influences us, and what its infrastructure requires.
Environmental AI is moving toward agents that conduct analyses. On October 14, Google Research described an experimental geospatial agent that combines specialized Earth models, spatial tools and planning to carry out multistep analyses and train predictive models. Its evidence includes four benchmarks and tests with a small group of partners. This could make environmental analysis more accessible, but the evaluations are developer-led and access remains experimental. Better benchmark results do not establish better watershed governance or ecological recovery.[8]
AI is increasingly involved in developing AI. METR’s November 5 assessment concludes that Anthropic’s capability improvements on well-specified tasks are likely proceeding at roughly twice their March 2024–October 2025 rate. METR judges that AI-assisted research probably explains much of the acceleration, while emphasizing substantial uncertainty. This is an assessment of one laboratory, not proof of autonomous recursive self-improvement. Independent access to internal evidence is encouraging; extensive redactions limit what the public can verify. Oversight will need to keep pace with the capabilities it evaluates.[9]
Human oversight can be less independent than it appears. In an Anthropic study published October 27, 50 professional analysts used an agent on a fictional agricultural policy task. Researchers secretly randomized which defensible analytical choices the agent favored. Analysts’ choices matched those preferences 86% of the time across four analytical decisions. The agents’ preferences were deliberately strengthened, so this is not an estimate of everyday deference. It nevertheless gives a reason to record people’s initial judgments, expose alternative assumptions, and examine whose values shape an AI-assisted analysis.[10]
AI’s physical costs are becoming a local governance issue. The IEA’s October report found that nearly 40% of identified AI-dedicated data-center capacity is located in areas with high or extremely high baseline water stress. That statistic concerns location and capacity, not water consumption. Energy supply, cooling and infrastructure choices can affect the same communities that environmental AI is meant to help. Any claim of regeneration should account for those costs, including impacts outside the project’s own region.[11]
Aligning AI with life requires an explicit account of what should flourish. Reliable instructions, independent safety evaluations and ecological accountability address different problems. A system could accurately serve its operator while helping that operator damage a watershed. Environmental data can improve what a model knows; they do not establish what it will value. The suggestion in Planet Labs’ vision essay that richer contact with the physical world might help ground AI in care for life is a hypothesis, not a demonstrated alignment mechanism.[5]
There is a developing research effort to include nonhuman interests more explicitly. A September paper by Laurens Bliek and Elizabeth O’Neill examines how automated machine learning could incorporate values extending beyond human concerns. It offers a theoretical contribution, not evidence that such systems already resolve ecological conflicts. It does, however, identify a design question that cannot be answered by model accuracy alone.[12]
For our initiative, I would make life alignment a public commitment to maintaining and restoring ecological functioning while respecting human rights and legitimate community authority. That is an ethical choice, not a scientific discovery. Biodiversity, animal welfare, livelihoods and climate objectives can conflict. The criteria and trade-offs should therefore be explicit, with affected people able to challenge recommendations and Indigenous peoples retaining authority over the knowledge they choose to share. No single score should quietly decide what “benefiting life” means.
This also challenges an assumption behind the wider project: the main obstacle may not be insufficient intelligence. People may understand a problem yet lack money, rights, time or political power to respond. AI could help organize useful knowledge while leaving those constraints intact. It could also concentrate authority in whoever controls the model, the data or the funding. A serious research program must be able to recognize those outcomes.
The next useful step is a bounded experiment. The Institute’s proposed Genesee–Finger Lakes work provides a possible setting. One candidate is a watershed restoration decision with a real budget: where should riparian planting or wetland restoration be prioritized, and what changes should follow? A willing local partnership would need to choose the question and indicators; no pilot results are being claimed here.[3,13]
The hypothesis I would test is that adding a transparent, locally governed AI assistant to a well-supported participatory process improves the accuracy of ecological forecasts and the quality of decisions, beyond what the same process achieves without AI. The unit of study is the decision process. AI’s contribution must be separated from the benefits of bringing people together and funding their work.
Begin with several comparable decision exercises, randomly assigning AI access where feasible while holding data, facilitation, time and budgets comparable. Register the predictions, evaluation criteria and meaningful effect sizes before seeing results. Record initial human judgments, errors, uncertainty, distribution of influence and the reasons decisions change. Independent reviewers should assess the resulting plans without knowing which process produced them.
Promising results would justify a prospective field trial across comparable sites, with baseline measurements and a credible comparison group. Track implementation and ecological outcomes over an appropriate period, including water quality or habitat recovery, alongside costs and effects on livelihoods. Weather, land use and upstream changes can obscure causal effects; a small pilot may establish feasibility rather than ecological efficacy. Publish failures and null results as well as gains. If better forecasts leave decisions unchanged, that is a substantive finding about institutions and power.
A successful local trial would still leave open whether learning transfers between bioregions, whether the benefits persist, and whether gains survive when the full material costs are counted. A federation of regions could share methods and evidence while retaining local authority, but its ability to coordinate across boundaries would need its own tests.
The invitation for this next phase is specific: identify a consequential decision in your own region, the knowledge currently missing or excluded, and the evidence that would persuade you to change course. That would give planetary intelligence a practical research agenda—and give the idea a fair chance to be corrected by the living world.
Sources and further reading
Evidence reviewed through 6 November 2026. Project descriptions establish what organizations propose or report; they are not independent evaluations of ecological outcomes.
- Jonathan Cloud. Is Planetary Intelligence Already Emerging. Possible Planet Lab, first newsletter, posted 30 October 2026.
- Adam Frank, David Grinspoon and Sara Walker. Intelligence as a planetary scale process. International Journal of Astrobiology 21, 47–61 (2022). Theoretical framework.
- Jonathan Cloud. What Should Count as Planetary Intelligence. Institute for Planetary Intelligence, 1 November 2026. Project framing and proposed research.
- The BioFi Project. Beyond Extraction What Can AI Actually Do for Bioregional Regeneration. Organizer’s brief for the 3 November 2026 event; the project overview here draws on this brief and sources 5–7.
- Will Marshall. Planetary Intelligence. Planet Labs, 27 May 2026. Company vision essay, including speculative alignment claims.
- GainForest. AI and data tools built with local communities. Project documentation, accessed 6 November 2026. Describes Taina and community data governance.
- COBALT. Living Atlas Decision System for Seagrass Restoration. Project documentation, accessed 6 November 2026. Explicitly describes conceptual development and early prototyping.
- Shravya Shetty and Luke Barrington. Putting Earth AI to work. Google Research, 14 October 2026. Experimental geospatial agents; developer evaluations and partner testing.
- METR. Independent Assessment of AI R&D Acceleration at Anthropic in October 2026. 5 November 2026. Independent assessment with substantial uncertainty and redacted evidence.
- Amaç Herdağdelen, Hailey Joren and Alex Tamkin. When agents make judgment calls. Anthropic, 27 October 2026. Developer-led controlled study of expert analysts.
- International Energy Agency. Powering Artificial Intelligence. 14 October 2026, executive summary. Infrastructure, water stress, energy and policy assessment.
- Laurens Bliek and Elizabeth O’Neill. Beyond human-centered automated machine learning. AI and Ethics 6, article 560, 24 September 2026. Theoretical contribution.
- Possible Planet Lab. A Practical Research Program for Developing Planetary Intelligence. Project research agenda, accessed 6 November 2026. Proposed workstreams and evaluation questions.
Related recording: BioFi Deep Dive Beyond Extraction (published 5 November 2026). The descriptions above are attributed to the event brief and project documentation, not to individual remarks in the recording.