This month’s Attention Inventory is compiled by our summer intern and Policy Communications Fellow, Katie Culbert.

Every time I walk up to my beehives and hear the buzz of them hard at work, I get excited. I love my honey bees, and no matter how much I learn about them, I always want to learn more. These tiny creatures (less than an inch long!) have continuously challenged my thoughts about cooperation, leadership, and society.

Democracy in the Hive

Before I started researching them, I thought honeybees operated like a monarchy – the hive served their queen, after all. But the queen isn’t ordering around the colony: it’s the worker bees who work collectively to make many of the hive’s decisions. For example, if the worker bees believe that the queen is not performing her job well, they will replace her with a new queen. In this way, honey bees actually function more like a democracy. This completely changed how I thought about them. Living in a democratic-republic myself, I often wonder what humans can learn from that kind of collective decision-making, especially during times of intense polarization.

I remember a time in high school, when we were assigned to work in groups for a final project about any famous figure in history. While we all agreed to do the project on FDR, no one could agree on an approach. I wanted to create a video, a friend wanted to write a storybook, and another wanted to build a website. We were all a bit stubborn and defended our ideas, convinced it was the best option. After making no progress and the deadline quickly approaching, we realized that no one was going to “win”. Instead of arguing for our own preferences, we started to evaluate the strengths and weaknesses of each idea. From that we discovered a solution none of us had considered before: a story map. It combined the accessibility of a website, with the narrative structure of a storybook, and the movement and engagement of a video. Now I see that we didn’t arrive at our best solution because one person had the right answer, we found it because we were willing to listen, compromise, and build off each other’s ideas. Just like the honeybees. Together, we created something stronger than what any one of us could have created alone. While we are taught from a young age that teamwork and compromise are important, it can be hard to actually live that out.

Thinking Bee-yond Ourselves

Another fascinating fact about honey bees is that they are classified as a superorganism, meaning that they act like one singular living organism even though they are made up of thousands of smaller beings. From the moment a bee is born, it has a specialized role that contributes to the survival of the hive. For example, guard bees defend the hive from intruders even if that means sacrificing themselves. Unlike wasps, honey bees have barbed stingers, which means that it is easy to lodge in but hard to take out. As a result, the guard bee dies shortly after they use their stinger. Now humans will probably never be able to function as a true superorganism (and I don’t know if we would want to) but I can’t help wondering what our world would look like if we all just thought a little bit more about the health of the whole rather than just ourselves. What I imagine would not look like a sweeping, grand change overnight, but instead a thousand small decisions that make our communities a better place. Like returning the shopping cart instead of leaving it in the parking lot, picking up a piece of litter that isn’t ours, or letting someone merge into traffic. None of these actions by any means can transform society on their own, but collectively they have the power to shape the communities we live in. I think all it takes is everyone to be a bit more intentional about their actions and the fact that we all share this space.

The Season of Scarcity

Every July I start to pay more careful attention to my bees. Dearth is coming. During dearth, nectar and pollen sources become scarce so colonies have to rely on their own stored honey. If a strong colony does not have enough of its own stores to sustain itself, they can raid a weaker hive for their honey (a behavior known as “robbing”). Learning about this reminded me what most living things, including people, are capable of when resources disappear. While I wouldn’t say I admire this behavior, I understand why it exists. Scarcity doesn’t excuse harmful actions but it helps explain them. It makes me wonder how many conflicts are less about character and more about the conditions people are forced to navigate.

I remember in the early days of the COVID-19 pandemic panicking about not having enough toilet paper. Grocery store shelves that were typically filled with rows and rows of toilet paper, canned food, water, and other essentials were suddenly empty. The shortages became so severe that most stores began placing limits on the amount that customers could buy. People who would never fill their shopping carts with months worth of supplies suddenly did and it made sense. No one knew what the following weeks or months would look like. Most people weren’t trying to leave their neighbors without necessities; they were just trying to protect themselves and their families. Yet these actions created shortages that affected everyone else. Looking back, that experience taught me that in moments of fear or uncertainty, our instincts can change in ways we don’t always recognize. Just like a colony frantically trying to secure enough food before Dearth, many of us felt compelled to gather as much supplies as we could before they disappeared. While the two behaviors are not identical, they are rooted in the same principle: a desire to survive in the face of scarcity.

The Complexity of Good Intentions

Most of the honey bees in the US are European honey bees (Apis mellifera), which were introduced to North America in the 1600s. While they provide invaluable pollination services for agriculture around the country, they can compete with native bees for resources like nectar and pollen. Most of North America’s native bee species are solitary, meaning they do not live in colonies like honey bees. And many are in decline due to pressures like habitat loss, diseases, pesticide use, and much more. So, this competition with honey bees can make conservation of native bees much more complicated. It complicates the idea of something being “good” or “bad”– even good intentions can have unintended consequences.

I’ve been facing this idea throughout my internship with OEDP. Through learning about data stewardship and participating in OEDP’s Learning Circle sessions, I’ve learned how much intention and foresight are required when working with community data. For example, at first glance it seems like reporting that the air quality in a community is unsafe is obviously good; it can raise awareness of serious public health issues and motivate action. But good data stewardship requires thinking a step further and asking: what happens next? Could the report cause property values to decline? Could residents be stigmatized because of where they live? Could the report cause panic among the residents? None of these questions mean that the information should be hidden, but they do require careful thought. They remind me that sharing data responsibly means considering the benefits of transparency but also the real-world consequences it may have for the communities it is intended to serve.

Everytime that I think I understand honey bees, a new layer is revealed. They have taught me that the world is far more complicated and nuanced than I had ever imagined, and that most questions rarely have simple answers. If you want to learn more or connect about how I’ve used these learnings in policy spaces, you can check out my introductory video, or email me at kculbert@openenvironmentaldata.org!

Sincerely,

NJS Assembly Bill A2869 and NJS Senate Bill S1256

Welcome to our inaugural Data Stewardship Learning Circles Cohort!

We’re excited to bring people together from across the country to explore questions around how environmental data is collected, managed, and shared. Stay tuned for highlights on each member, coming this summer!

&&&&& other things we’ve been up to

• We’ve grateful to have had a lot of in-person time at events like DC Climate Week and OEDP’s fifth annual staff retreat!

• We’ve wrapped up our Fieldnotes on AI series—keep an eye out for a brief in the next few months.

• Sr. Fellow Shannon Dosemagen co-authored a piece with Gwen Ottinger about Constructing a New Knowledge Infrastructure for Issues in Science & Technology.

• Shannon also published a piece in Southern Cultures about what it means to call a changing place home.

• Cathy recently gave a talk about AI and evaluation when it comes to climate resilience, disaster response, and humanitarian action.

• Brittany co-wrote a blog about a new Office of Management and Budget (OMB) rule would make political alignment a prerequisite for federal science funding.

What We’re Reading, Watching & Listening To…

🦦 Brittany is perusing the Federal Data Field Guide

⚡ Katie is rethinking climate regulation from the ground up via the Volts podcast

👂🏻 Emelia is listening to The Data Fix (a lefty podcast about perpetual tech promises)

🌀 Cathy is reading How to Fall in Love with the Future: A Time Traveller’s Guide to Changing the World by Rob Hopkins

💡 Charles is reading The Politics of Open Infrastructures: Power, Governance, and Justice in Digital Knowledge Practices

💣 Michelle is finally learning about How to Blow Up a Pipeline

Get monthly updates on what we’re reading, watching, or listening to on our website.

Opportunities

Here are a few event (🎟️), funding (💰), professional (💼), and other (🗨️) opportunities we want to share with you all:

🗨️ Nominate impactful EJ practitioners through FracTracker’s Community Sentinel Awards
💼 Map of Life is hiring a Head of Data Science
🗨️ Explore Open Knowledge Foundation’s new AI Learning Labs
🗨️ Submit your own public comment in response to OMB’s new political mandate (see here for guidance on how to do this) by July 13
🗨️ Sign onto the Preservation of Government call to action
💰The Nathan Cummings Foundation is currently accepting LOIs
🎟️ Reserve tickets to the ESIP meeting in Austin from July 28-31

If you have something you’d like to submit for our next quarterly letter, please reach out to dayna@openenvironmentaldata.org with the subject “Newsletter Opp Submission.”

This month’s Attention Inventory is compiled by Katie Hoeberling, Deputy Director at OEDP.

I will preface this with a note that I haven’t been in a particularly inspired mood lately. Good things are happening at OEDP for sure, and I’m personally doing okay. But it hasn’t gotten any easier to watch this administration, spurred on by corporations, dismantle the institutions, which, while imperfect, held some semblance of a line against free-for-all extraction and pollution and war. My attention has been… chaotic and sporadic. But I’m piecing things together and trying to ask good questions. Here’s what that’s looked like lately:

What can we learn when we put aside American exceptionalism?

I had the immense fortune and privilege to spend much of winter traveling in Asia—to the Philippines for holiday celebrations with family, to south India for a dear friend’s wedding, and to Singapore and Sri Lanka for fun (and food). My favorite way to pass time on long haul flights is to read about where I’m traveling, so I brought “How to Stand Up to a Dictator.” Maria Ressa’s memoir of her career in journalism throughout Southeast Asia is also a personal recounting of how Facebook has enabled authoritarians to spread disinformation and consolidate power (from Duterte and Bongbong Marcos in the Philippines to Putin in Crimea and Russia). It was an enlightening and frustrating read, not just because Meta continues to enable the manipulation of public opinion and spread of disinformation, but because Ressa points to so many signs and opportunities we’ve had to prevent Big Tech from steering our recent elections.

Traveling and reading made good reminders of how easily we as Americans fall into the trap of exceptionalism. The word “unprecedented” has become ubiquitous in our media and discourse. And I worry that this thinking blinds us to things we can learn from people and places that have already experienced what feels unprecedented for so many of us. What might we be missing or dismissing? Ressa’s experience and frameworks for ‘standing up to dictators’ could be really useful for resisting the Trump administration and thinking about the stakes of the midterms. For example, the coalition-based #FactsFirstPH approach spearheaded by Rappler (the Filipino digital media company Ressa co-founded), mobilizes journalists, communities, researchers, and attorneys in defending against disinformation, especially ahead of elections. Each group has different but interlocking roles: fact-checking, “algorithmic amplification,” data analysis, and legal protection. If federal and state governments in the US can’t (or refuse to) regulate big tech, what roadmaps can we create for ourselves so we can organize around, and in spite of, them to maintain integrity in our elections?

What creates the conditions for memory and attentive, careful work?

Less than 20 hours prior to arriving at Changi Airport and taking a Grab taxi to a 23-hour Chinese restaurant that served some of the best dumplings I’ve ever had, I was at home in Oakland trying to remember what I’d forgotten to pack (a hairbrush, it turns out). Oakland is a vibrant and beautiful, if rusty, place I’ve been proud to call home for almost a decade. It can also be hard to walk around the city and not feel dismay at the underfunded transit infrastructure, overfunded surveillance infrastructure, and the relentless fights people take to defend their home encampments from being swept up. It was jarring to go from that to Singapore’s well-manicured parks and shiny financial district skyscrapers. I kept thinking “Who isn’t here anymore?” and “Who or what was sacrificed to make this world possible?” These questions ring just as true in the Bay as they did there. In the US, we seem to be in a perpetual state of cultural and institutional forgetting or erasure. Some of this is of course deliberate (deportations and “sweeps,” sanctions on free speech, data removals). But I think a lot of forgetting, institutional or otherwise, happens when we’re pressured to move too quickly or when our attention is captured for profit. I think about what conditions can enable more careful and deliberate work—work that honors and preserves memory, work that makes room for and intentionally integrates what we’ve learned. Speed and distraction threaten our capacity to care.

This is coming up a lot for me as OEDP deliberates on how we do and don’t want to interact with generative AI. We don’t subscribe to the narrative that AI ubiquity is inevitable or that we should allow AI to make decisions for us, but neither are we immune to the appeal of technology that would help a small, scrappy organization do more with the limited time and resources we have. I would love to be able to synthesize and write more quickly. But I also don’t necessarily think it should be easy. We’re grappling with complicated questions that deserve attention, care, and dare I say a little struggle. I recently listened to an essay by Stephanie Krzywonos where she talked about writing as thinking. This really resonated with me as someone who often uses writing to process my thoughts, and thinks of writing as a practice and a muscle. In fact, this newsletter was really hard for me to write. I felt out of practice. What will happen to this muscle if I cede the work, the practice of writing to a machine that cannot care?

In “Pedagogy of the Oppressed,” Paolo Freire says “I work, and working I transform the world.” If work = force x distance, then is using AI to generate writing (or anything else, really) the exertion of limited force over no distance at all? Meaning, are we forgoing work, and thus transformation, by wielding a technology that merely recreates the world as it already exists? If we need new (or previously disregarded) ideas, new ways of telling stories, renewed will to act on the things we know, I’m skeptical that LLMs are going to get us there. I will say I’m excited by some of the alternative, small, and local ways of engaging with AI that OEDP is exploring. Our Fieldnotes on AI series by Shannon, Cathy, and Emelia is a great example of how we can think curiously and critically about AI in environmental spaces.

How can environmental governance accommodate multiplicity?

The other thing I’ve been ruminating on lately—have been really since I started working in science and policy spaces—is how to build systems of data and accountability that can accommodate multiple ways of knowing. Community-collected data has long been a mainstay in conversations about environmental justice. Even as we’ve hyperfocused the last couple of years on protecting and stewarding government data, our partners and stakeholders continue to remind us that, as we re-build tools that have been deleted from public websites, we have an opportunity to do what governments, by and large, have not yet been able to: make room for community voice to meaningfully inform environmental decisionmaking.

These conversations tend to come back to standards—getting agencies to better communicate standards for data, resourcing community researchers to meet them, and building sensors and digital infrastructure that make it easy to collect and upload standardized data. Emelia (our Sr. Research & Policy Manager) and I wrote about this in 2024 and, even though EPA eliminated some of our target audiences (like the Office of Environmental Justice), I still believe state and local agencies could benefit from our recommendations.

But while we work towards standardization, we also have to keep fighting what our long-time collaborator Gwen Ottinger calls “scientific chauvinism.” Gwen describes this concept in detail in their new book, “The Science of Repair,” but to paraphrase here: scientific chauvinism is where researchers and governments privilege credentialed scientists as the sole definers of quality or valid evidence, often through data standards and other norms for knowing well. I’m still reading the book, but I’m encouraged by the possibility of a middle ground Gwen asserts: “We need to bring science into moral community.” Maybe science is limited in its ability to help environmental justice communities hold polluters accountable (at least on its own). But we can bring flexibility into our governance systems: for example, by adjusting the standards for evidence by which we assess the risk of harm to already overburdened neighborhoods—i.e., prioritizing public health significance over statistical significance. (I’m anticipating coming back to this as my own community gears up to try and stop the construction of a new coal terminal at the Port of Oakland. Yes, in 2026.)

None of this will be easy, but I’m emboldened by all the places where I see this conversation showing up, especially the new EJ Data & Mapping Society. We just had our quarterly general meeting—if you want to join the next one or one of the working groups we’re standing up, sign up here!

How do we hope when hope isn’t working?

I realize this wasn’t the most hopeful inventory, but it’s where my head and heart are lately. And honestly, hope isn’t really doing it for me right now anyway. If you’re in a similar boat, I’ll leave you with some righteously furious advice from one of my favorite writers Rebecca Solnit: “If the word hope doesn’t work for you, try ‘Never fucking surrender.’”

I, for one, am trying not to. As anxious and pessimistic as I am about the state of our country and the world, I’m still building new things with PEDP, yelling at OpenAI, going to musicals, and planting seeds (an admittedly hopeful act). This season I’m growing three Asian eggplant varieties, three tomato varieties, summer and winter squash, shiso, and marunggay (wish me luck, my uncle who runs a Filipino nursery out of his backyard warned me it won’t do well in a place as chilly as the Bay, but here we go anyway). I also recently became a proud Master Gardener of Alameda County, and am finally getting to tend to gardens in Oakland after so much travel. If you’re in the Bay next month, come say hi at West Oakland Farm Park!

Sincerely,

What does it mean to care for data and communities when participation itself can become a site for extraction?

We turn to refusal, stewardship, and obligation in the era of AI models. We question what “open stewardship” means when data becomes training material and models become intermediaries.

I talk and think about data stewardship sort of constantly (if you want to talk and think about it together, join me for one of my Data Drop-Ins). Here are some things that I know to be true about the concept:

• It is basically a series of decisions

• It is a relationship

• There are limits

• One must play the long game

• It’s a way of future-proofing

• The process rewards

When we think about stewarding data that then goes to train AI models, does this concept of stewardship change? Not necessarily… there are just more decisions to make within a weirder governance structure.

Standard data stewardship considerations include decisions about what goes where and when. Who are we sharing which pieces of data with? What is the time period for sharing? How do we retain access? In my work, to answer these questions you need to sit down with your team and work up partnership protocols and data sharing agreements. You can soft-legal or hard-legal require specific things in a MOU or contract. If you have less sensitive and more open data, you choose how others access it: aggregated or disaggregated, in raw form, or through visualizations or specific analyses.

A thought partner of mine recently mentioned that someone they knew had a post-it on their desk that simply said, “To what end?” and used this reminder to distill, refine, move forward. It’s a useful reframe in this context: we are making our data available, but to what end? Can we even answer that question when it comes to making data available to LLMs? When we share data with another organization, we ensure that we are values or purpose aligned beforehand; we have a relationship. Can we do that in the age of bots scraping the web to train big models?

Where does that leave us as we try to instrumentalize data ownership? As it stands, most scraping trains models without giving attributions to original sources or requiring consent. If these models are making a machine god and our data made those models, then is that god ours? What if when your data is used to train a model, you could claim a little bit of that model? (I obviously have more questions than answers this week.) Cooperative and co-ownership of AI models are possible and interesting and happening. There are always going to be limits and barriers. Stewardship is inherently limiting, because if you truly care and attend to something, you can’t care and attend to everything, but we can make it easier to ask those data stewardship questions, do some stewarding, and build larger infrastructural cooperatives that offer alternatives.

Sometimes stewardship is saying no. Refusal in your personal life looks like protecting your peace, and the analogy serves us well here. Choosing not to opt in, not to feed the data into xyz model is a choice that can allow a community to re-orient within their own timeline, play the long game, and build what they want. Te Hiku Media is a great example of an organization and community that have refused offers from Big Tech and created collective models that are fit for their purpose, like the te reo speech recognition model and the Kaitiakitanga License. Creating structures that reinforce collective ownership and sovereignty is a way of future-proofing community data and languages.

Undoubtedly, AI development will continue in new and strange ways. We will have to ask new questions and make new types of decisions. At least in the US context, we are currently in the Wild West of AI governance, and without any meaningful regulation, what we can do in the meantime is bulk up those data stewardship muscles and engage in a bit of AI countergovernance. We can find alternative mechanisms for community-led and worker-led governance to resist or contest power in AI systems. Stewardship might require a bit of friction, resources, and time, but it rewards; bolstering relationships around and with your data is one way to build something different in an increasingly hostile (actual and digital) environment.

What does it mean to care for data and communities when participation itself can become a site for extraction?

We turn to refusal, stewardship, and obligation in the era of AI models. We question what “open stewardship” means when data becomes training material and models become intermediaries.

Open Environmental Data Project first started with acknowledging a gap. Community science and environmental data collection efforts were proliferating, but we still lacked strong translational roles, organizations, and practices that could bridge institutions and communities while maintaining the space of environmental data stewardship. Some of these roles certainly existed, but the work itself was often inconsistently recognized and resourced. Translation in environmental governance is not just about moving information between parties. It requires navigating unequal power relationships, protecting context, building trust, and helping people retain agency over how their knowledge is interpreted and used.

I’ve often referred to these organizations and people as “intermediaries,” whose work is focused on the human role of interpreting and facilitating data exchange between different parties. Before the AI conversation inundated the participatory landscape, developing these translational and stewarding roles was difficult to achieve, but relatively straightforward to imagine. You can see this tension in the work of OEDP’s data stewardship handbook and in many environmental justice and participatory science projects that have attempted to balance openness, accountability, and protection at the same time. Sometimes these roles are meant to help people make decisions for their own sake. Sometimes they are meant to support multiple parties in creating agreements. And sometimes they are meant to provide protections and autonomy to communities who want to share their data while trying to point to a bigger issue.

But a lot of the decisions being made around community-generated environmental data still largely exist at the levels of research, community engagement, or education. The data might support advocacy or public awareness campaigns, or it might shape a local policy discussion. But communitycommunit data often remains connected to the context in which it was gathered and interpreted. The relationship between the people generating the knowledge and the claims being made from it is still relatively visible.

AI systems are beginning to change that relationship. The challenge is not that more data are being collected or shared, but that the distance between sharing data and governing how that data acts in the world is collapsing. Community environmental data is no longer only being used to support a particular claim or intervention. Increasingly, it becomes substrate for systems that generate new inferences, rankings, recommendations, and decisions, far beyond the original conditions and intentions of collection.

For translators and stewards, this creates a very different problem space. From the moment of deliberate sharing to the moment an AI system absorbs data into a model, there is a letting go that happens, which is not always intentional. Once data becomes part of a model, there is often no clear way to understand how it will later be interpreted, surfaced, or operationalized.

And the implications are not abstract. We can easily imagine a legislative assistant using AI tools to help weigh the risks and benefits of an environmental decision. Or a regulatory agency relying on AI-generated synthesis to prioritize interventions. Data that once functioned as situated evidence within a community may instead become generalized training material feeding systems that shape future policy recommendations.

It’s ironic, when you consider the question of legitimacy. Historically, community-generated knowledge has often been dismissed as anecdotal, emotional, or political—until it becomes validated by institutional actors. AI systems may place a veneer of technical authority over community-generated data that suddenly makes it more legible or persuasive to decision-makers. But what gets lost in that translation? What context disappears when lived experience becomes machine-readable inference? What does it mean when institutions cannot hear people directly, but can hear their experiences once translated into machine-readable form? What obligations remain to the communities who generated the knowledge in the first place?

This is part of why I found myself nodding along when Sylvie Delacroix described “data stewards” as one of the missing careers of the twenty-first century. To really crack the nut on stewardship in the age of AI, we need to think much more seriously about how people are trained to use data. And I’m not talking about communities creating data—I’m talking about AI developers, computer scientists, data analysts, procurement officers, consultants, and end users of AI systems in state and local government. Increasingly, these actors are becoming governors of environmental knowledge systems whether they recognize it or not.

While most technical systems are still designed as though mediation, negotiation, and contextual interpretation are inefficiencies that automation can remove, the more environmental knowledge becomes intermediated through AI systems, the more important human translation becomes, not less.

This is where intermediary roles and ombuds models become increasingly important. Colorado’s Environmental Justice Ombudsperson role, for example, exists in part because environmental governance is not only a technical problem. It is also a problem of accountability and attempts at conflict resolution. Communities need mechanisms that help them navigate institutions while also preserving autonomy and recourse. We need more roles that recognize stewardship not just as maintaining data infrastructure, but as maintaining relationships, obligations, and consent over time. And translators become all the more important when we start thinking about negotiation, exchange, and consent itself as infrastructure. That infrastructure includes refusal.

Part of the difficulty here is that many contemporary AI systems are not emerging from neutral conditions of knowledge production. Meredith Whittaker described AI as being “born out of surveillance,” pointing to the enormous infrastructures of extraction, tracking, behavioral monitoring, and data accumulation that underpin many large-scale AI systems. The concern is not just about privacy in the narrow individual sense, but about the way contemporary technical systems normalize continuous observation and treat human behavior, communication, and participation as material to be captured and operationalized.

That framing matters for environmental and participatory science communities because so much of this work has historically depended on trust, voluntary participation, and negotiated forms of openness. Communities often participate because they believe they are contributing to accountability around a shared environmental (and/or health) issue. But when participation itself becomes raw material for systems built on extractive logic, the meaning of consent starts to shift. The question is no longer only whether people agreed to share data, but whether they meaningfully understood the downstream systems their participation was feeding.

In participatory science and open source communities, participation is often framed as inherently good. Sharing is progress, and openness is framed as virtue. While I deeply believe in the importance of participation and collective knowledge production, stewardship in the age of AI may increasingly require us to create conditions where communities can also withhold participation, redirect it, or weigh its terms.

Carole McGranahan writes about refusal not as simple rejection, but as a political and generative act, one that creates the possibility for different relationships, obligations, and futures to emerge. Refusal, in this framing, is not withdrawal from engagement altogether. It is a way of asserting that participation cannot be assumed, and that communities retain the right to shape the terms under which they are known, represented, or incorporated into larger systems. Refusal becomes a way of maintaining autonomy and protecting the possibility of alternative political arrangements rather than accepting the structures already on offer.

That framing feels increasingly important in environmental and participatory science contexts. Refusal is not only obstruction. It can be a way of preserving autonomy, slowing processes, protecting context, and keeping alternative futures imaginable. It can also be a way of insisting that consent is ongoing rather than part of a single transaction, where communities retain some ability to shape not only whether their data is collected, but how it travels and mutates once it enters technical systems.

This is perhaps the hardest shift for many of us who came from traditions of open science, open data, and participatory technology. We may be entering an era where stewardship no longer means just helping information move more freely. It may also mean helping communities understand what sits on the other end of systems that increasingly use participation itself as raw material.

If AI systems are becoming intermediaries between people and environmental knowledge, then stewardship cannot end at the moment of collection or release. We won’t be able to think about consent as a box checked or a singular agreement. Governance will have to span full lifecycles to persist as data moves through institutions, models, interfaces, and decisions. The future of open stewardship may depend less on how much information we can circulate, and more on whether communities retain the ability to partake in the terms under which knowledge is transformed into mechanisms of action and levers of decision-making.I have three graduations this summer. Two niblings are leaving high school, one is leaving college. (My money is also leaving my pocket.) But this period has also made me reflect a lot on a time when they were little, and a younger version of myself was entrusted to keep them whole and within eye shot. I joke that it’s a miracle that we’re here—not because I/we didn’t take our job seriously, but because it seems kids do not come with a self preservation mechanism built in. But we made it. The kids were stewarded. And now I’m the one in the passenger seat.. I often remind them that the responsibility to get me home in one piece is on them now. They hold the keys.

Stewardship on its own is relatively straightforward. It’s the careful and responsible management of something entrusted to one’s care. In the case of data it typically means that you collect data, you share it, you make it discoverable and reusable. The open science movement pushed institutions toward transparency, toward FAIR principles. But things get complicated, largely because we’re humans. Careful management also means dealing with corporate entities that can easily discover, reuse, and monetize data they never collected. Responsible management also means ensuring that the data isn’t misused by bad actors. AI models ingest, they then predict and transform, then learn from individual interactions, and the owners of the model benefit. But what about the community whose knowledge fed into it? Arguably, without them, there would be no AI model.

However, when we talk about what developers and institutions owe the communities they rely on, we tend to reach for the word consent. Did people agree to have their data used? But consent assumes that people knew what they were agreeing to, that the terms were legible, that the power between the person clicking “I agree” and the institution holding the data was roughly equal. Yet, when you have to use their platform to submit all your documents for your new job, is there much of a choice?

David Kernohan sees this as a structural gap: “when you’re ill, you rely on doctors whose professional obligations require them to prioritise your wellbeing. When you need legal representation, lawyers are bound by duties to act in your best interests. But when your health data, location history, or browsing patterns are being negotiated for use in AI training, algorithmic decision-making, or commercial exploitation – who represents you?” We could say it’s the responsibility of data policies such as GDPR, but we all know HIPAA wouldn’t be as effective without the people responsible for ensuring that it’s enforced.

As the Citizens and Technology Lab put it, “those who create the tools of science exert a powerful influence on what can and can’t be known.” The questions researchers ask determine what gets measured. What gets measured shapes what is considered a problem. What is a problem determines who gets to contest it and how.

And that’s where refusal comes back into play. While we may not always have the power to do so, Ufuoma Ovienmhada writes that (when we can) there is an obligation to refuse to produce and present research that might endanger marginalized populations or entrench harmful narratives. Specifically, they describe refusal as “the act of withholding knowledge, withholding data, and withholding consent.” This may sound like the opposite of what research is supposed to do. Scientific knowledge is supposed to be disseminated or be FAIR. But that assumption is itself a choice, and it is not a neutral one. It assumes that the production and circulation of knowledge is always good, that more visibility is always better, that the communities whose lives are being studied always benefit from being studied. None of that is guaranteed. Refusal, understood this way, is not a failure of openness. It’s a protective act. It’s a recognition that some knowledge, in the wrong hands or at the wrong moment, does harm.

AI adds another layer to this. There is the data that helped create the model, but then there’s also the model that creates even more data (accurate or not). The questions the model can answer become the questions that count. And the model’s questions were set by whoever designed it, whoever funded it, whoever decided what the training data should look like and what the benchmark should measure. This is why we’ve found that it’s important that we make efforts to reduce harm to communities, in particular, develop data use and sharing agreements to be clear on who has the keys, when, and how that responsibility will manifest.

What does it mean to care for data and communities when participation itself can become a site for extraction?

We turn to refusal, stewardship, and obligation in the era of AI models. We question what “open stewardship” means when data becomes training material and models become intermediaries.

I have three graduations this summer. Two niblings are leaving high school, one is leaving college. (My money is also leaving my pocket.) But this period has also made me reflect a lot on a time when they were little, and a younger version of myself was entrusted to keep them whole and within eye shot. I joke that it’s a miracle that we’re here—not because I/we didn’t take our job seriously, but because it seems kids do not come with a self preservation mechanism built in. But we made it. The kids were stewarded. And now I’m the one in the passenger seat.. I often remind them that the responsibility to get me home in one piece is on them now. They hold the keys.

Stewardship on its own is relatively straightforward. It’s the careful and responsible management of something entrusted to one’s care. In the case of data it typically means that you collect data, you share it, you make it discoverable and reusable. The open science movement pushed institutions toward transparency, toward FAIR principles. But things get complicated, largely because we’re humans. Careful management also means dealing with corporate entities that can easily discover, reuse, and monetize data they never collected. Responsible management also means ensuring that the data isn’t misused by bad actors. AI models ingest, they then predict and transform, then learn from individual interactions, and the owners of the model benefit. But what about the community whose knowledge fed into it? Arguably, without them, there would be no AI model.

However, when we talk about what developers and institutions owe the communities they rely on, we tend to reach for the word consent. Did people agree to have their data used? But consent assumes that people knew what they were agreeing to, that the terms were legible, that the power between the person clicking “I agree” and the institution holding the data was roughly equal. Yet, when you have to use their platform to submit all your documents for your new job, is there much of a choice?

David Kernohan sees this as a structural gap: “when you’re ill, you rely on doctors whose professional obligations require them to prioritise your wellbeing. When you need legal representation, lawyers are bound by duties to act in your best interests. But when your health data, location history, or browsing patterns are being negotiated for use in AI training, algorithmic decision-making, or commercial exploitation – who represents you?” We could say it’s the responsibility of data policies such as GDPR, but we all know HIPAA wouldn’t be as effective without the people responsible for ensuring that it’s enforced.

As the Citizens and Technology Lab put it, “those who create the tools of science exert a powerful influence on what can and can’t be known.” The questions researchers ask determine what gets measured. What gets measured shapes what is considered a problem. What is a problem determines who gets to contest it and how.

And that’s where refusal comes back into play. While we may not always have the power to do so, Ufuoma Ovienmhada writes that (when we can) there is an obligation to refuse to produce and present research that might endanger marginalized populations or entrench harmful narratives. Specifically, they describe refusal as “the act of withholding knowledge, withholding data, and withholding consent.” This may sound like the opposite of what research is supposed to do. Scientific knowledge is supposed to be disseminated or be FAIR. But that assumption is itself a choice, and it is not a neutral one. It assumes that the production and circulation of knowledge is always good, that more visibility is always better, that the communities whose lives are being studied always benefit from being studied. None of that is guaranteed. Refusal, understood this way, is not a failure of openness. It’s a protective act. It’s a recognition that some knowledge, in the wrong hands or at the wrong moment, does harm.

AI adds another layer to this. There is the data that helped create the model, but then there’s also the model that creates even more data (accurate or not). The questions the model can answer become the questions that count. And the model’s questions were set by whoever designed it, whoever funded it, whoever decided what the training data should look like and what the benchmark should measure. This is why we’ve found that it’s important that we make efforts to reduce harm to communities, in particular, develop data use and sharing agreements to be clear on who has the keys, when, and how that responsibility will manifest.

How can prototyping sharpen inquiry—or capture it?

We consider how fast, imperfect prototypes can clarify thinking (rather than replace it) and environmental questions (without locking communities into technical pathways).

A thread that runs through my work has been open physical tools for environmental monitoring, part of open science hardware (OScH). When experimenting and prototyping with physical objects, the results can be seemingly quirky, but they are actually a perfect example of what we should be aiming for in digital spaces: things that are responsive to the environments in which they’ll be used.

I spent a decade immersed in this kind of daily prototyping while working alongside the Public Lab community. Public Lab’s aerial mapping rig is a good example. To capture images once the camera was high up in the air, you basically had two options: use a hacked SD card that allowed for continuous shots, or find something to depress the shutter button. Once, on a trip into the Gulf of Mexico, I failed to bring my usual “depressors”, which were the head of a Q-tip or a small stone. They were easy-to-find objects, but less easy to imagine as components that enable the capture of hundreds of images from thousands of feet up. It was a critical miss on my part. Looking around the boat, though, the solution presented itself: a tiny bone from a fish, caught on a previous voyage. That spirit of prototyping persisted.

In the prototyping I’ve done, ideas are bound to a specific time or place, built with the materials at hand, shared digitally, and then repurposed in other contexts with their own questions and constraints. The point is never perfection: it is about learning along the way, asking better questions, noticing what use (or a thing or an idea) reveals, and allowing the object itself to reshape how we understand the problem. In prototypes, we learn and begin to see our engagement with the world differently.

There is something important here for our approach to AI. In physical prototyping, imagination is not abstract. It is grounded and emerges from constraint, from proximity to problems and questions, and from the need to make something work with what’s available. It is iterative, revisable, and accountable to the conditions in which it operates. The prototype does not close the question; it opens it up further. In contrast, many AI systems arrive already formed by presenting outputs that feel complete, even when they are built on partial data or unexamined assumptions. They compress the space in which imagination might otherwise operate. Instead of asking, what could this be?, we are asked to evaluate what has already been generated. This is where the risk of capture begins.

When a prototype is treated as a solution, meaning a model output is taken as an answer rather than a prompt, it can narrow the range of possible interpretations before communities have had the chance to interpret, contest, or redirect what is being proposed. And yet, prototyping in AI could look very different if we looked more closely at other examples of how prototyping unfolds in the physical world. There are instructive examples from earlier “emerging technology” moments, including open science hardware, the early Internet of Things, and even personal computing kits of the 1970s and 80s, where devices were often situated, iterative, and shaped through use. And in the AI space, embodied AI work such as CurrentAI’s “Suno Sutra” points toward a vision for locally-run, shared AI devices that are less dependent on centralized control.

Rather than building toward scale or optimization, we might build toward clarity. As we discussed in the intro to our OEDP fieldnote series, we’re interested in small, local models, and collaborative workflows at a localized scale. And through experience what can help us get there are events like clinic-style engagements where the goal is not to produce a definitive output, but to surface assumptions, test interpretations, and make space for disagreement. We want our spaces to look at things that are perhaps temporary, and are designed to be revised, or even discarded, but always documented.

This is where imagination matters. Though you might be surprised to find imagination as a key thread in a book largely about bureaucracy, David Graeber’s writing on this in The Utopia of Rules is something I’ve been returning to. He points to older understandings of imagination as “a kind of middle ground, a zone of passage connecting material reality and the rational soul.” That framing resonates well with prototyping. A prototype sits in that middle ground, being not fully real, and not fully abstract. It is a way of thinking with materials, and of moving between idea and thought to the creation of what could be.

What I worry about is how little room we are leaving for that middle ground even in the techo-scientific landscape that we interact with, which wasn’t as polluted with AI debates just a few years earlier. In many institutional contexts, the pressure to move quickly from idea to implementation leaves little space for this kind of work. The messy stages, where questions are still open, and imagination has not yet been maneuvered into a solution, are treated as inefficiencies rather than as necessary preconditions for understanding what is at stake and what is possible. Without that space, we lose the ability to ask different questions and notice what doesn’t fit, or imagine alternatives to what’s already encoded in the systems we have. In the early years of OEDP, this was always a key part of our work, alongside the very pragmatic work that data can funnel us into. While we live in existing systems, we still have to question why, and imagine other ways.

If AI is to support environmental knowledge and decision-making, it needs to operate within that middle ground rather than collapse it, functioning less as an answer machine and more as a companion in environmental inquiry. It has to be something that helps us test, reflect, and revise without shutting down the interpretive work that humans need to drive. Prototyping can help to protect that space by keeping part of the work lightweight, and reminding us that the goal is not to get to the answer faster, but to understand more fully what the question actually is.

How can prototyping sharpen inquiry—or capture it?

We consider how fast, imperfect prototypes can clarify thinking (rather than replace it) and environmental questions (without locking communities into technical pathways).

There’s a particular satisfaction to finishing something. I just recently finished a book nook, which - while not the hardest one I’ve ever worked on - I probably admired it way too often. Completing things is great. In theory, I wish I completed more things over the course of my life. But, the reality is that in life most things should never really be ‘done’ yet we often strive for an all encompassing solution. This is mostly fine when you’ve correctly identified the problem, but when you haven’t, you risk losing the serendipity of discovery and experimentation.

Charlie Warzel has a name for this: Builder Brain. The Builder mindset tends to skip past the complicated, specific details of a problem in search of the big, ‘world changing’ fix. Builders are so interested in what’s possible that they make illogical leaps, creating things that may or may not truly alleviate the problem then move on to the next big thing, leaving the mess for everyone else to sort out. Builders make tools for communities they’ve never met, for land they’ve never stood on, and for problems they’ve mostly read about.

When scientist Catherine Nakalembe tried to use AI to map crop types in western Kenya, she found that the available tools simply couldn’t recognize what they were looking at. They had been trained almost entirely on European and American farmland, and maize and beans in Kenya look different from maize and beans in Iowa. Her fix: GoPro cameras strapped to the helmets of dozens of volunteers, walking through actual fields, collecting over five million images in two weeks.

Arti Dhar, who runs a conservation group in India called Farmers for Forests, ran into the same problem. Her team tried using a well-regarded open-source model to identify trees on farmland and it missed more than half of them. North American forests had given the model one picture of what a tree looks like. Maharashtra gave it another. They ended up manually labeling 55,000 individual trees from drone footage.

On the surface, these are stories about missing data. But I think they’re about something else entirely - clarifying thinking using imperfect prototypes. The fact that both models were open-source further supported this endeavor as scientists were able to iterate, tweak, and train it on their own data.

Admittedly, over the last few years “move fast and break things” has developed an entirely new meaning. I’ve never been a fan of it, either: it discards the humans caught in the mix when things break. However, prototyping and learning quickly and safely is still a thing we can endeavor for.

The more useful kind of prototyping treats the failure as information. Like in Meet the Robinson’s where the family celebrates failure: “From failure you learn, from success, not so much.” The prototypes displayed around the mansion are a reminder of all the ways they learned along the way.

In the case of AI, not only can we prototype and test quickly, much like the examples mentioned before, but we can also ask these models to help us highlight things we haven’t considered. What questions aren’t we asking? What are we overlooking?

These prototypes won’t cover all use cases. As a matter of fact, they shouldn’t. But they should get us a step closer to understanding where we need to be. What makes a prototype genuinely useful is that it can still be broken and revised without much attachment. A community can look at it and say: this doesn’t look like my home. The option to do that disappears the moment the tool gets scaled or gets labeled as a solution.

Warzel ends his piece with a quote from Elizabeth M. Renieris that’s stuck with me: ”If we just keep building without repairing what exists or applying lessons learned along the way, we will continue to spin our wheels as the same problems accumulate and amplify. In this way, our technology may evolve, but our relationship to it (and to each other) can only degrade.”

I think the most useful thing a prototype can do, especially in environmental work, is slow that down. Not by being less ambitious. But by keeping the questions going a little longer before the answer gets locked in. Spend more time with communities, ask more whys, look around a little longer.

How can prototyping sharpen inquiry—or capture it?

We consider how fast, imperfect prototypes can clarify thinking (rather than replace it) and environmental questions (without locking communities into technical pathways).

Prototyping is a way to clarify, refine, and make mistakes in a way that is not permanent. A weird little ratio of care and messiness. When I’m working with textiles, the questions may be: What would this look like if the seams were visible? (pretty cute, actually) What does this green look like with this shade of blue? (oof, no) When I’m doing data stewardship work at OEDP, a similar style of questioning might help us to test something modular instead of creating a one-size-fits-all strategy. Or it could point us toward building a low-cost repository alongside end users. What questions are coming from users after they test out one of our data guides, and how can we improve?

Prototyping as a design practice is an act of care. You value the intended users or the necessary resources enough that you spend the time to get it right. You value that end product without knowing what it will be, or what the exact journey will be to get there. I am interested in exploring where AI tools can fit into the prototyping journey, to see if it can support expansion of thinking instead of narrowing it. As always, it’s probably a maybe, with some well-placed guardrails that keep a user from careening off the road.

Don’t become too attached to a specific idea

Amy Poehler dropped some wild insight in her conversation with Natasha Lyonne on her podcast Good Hang:

“[..] you can’t believe that your good idea is your last good idea. And in fact, throwing it away is like a power move to remind you that the next good idea is right behind it. You cannot get too precious about anything.”

Prototyping a joke or a skit (Poehler’s work) obviously means something different than prototyping a community co-design process of tools for environmental data (our work). But the central impulse is similar: prototyping will help you destroy your ego through consistent feedback, and build the muscle of generating better ideas and questions. You take the work so seriously that you can’t take yourself too seriously.

With our work creating the Community Data Playbook, we have an entire chapter dedicated to our learnings: what we set out to do vs. what we learned and altered along the way. For example, initially, we thought of the Community Data Hubs model as a kind of cohesive package to adopt and adapt. But as we learned more from our community partners, we decided to keep some elements of it (i.e., the Resource Library) and make them more modular, so there were more entry points for someone approaching data stewardship work. Ideas are meant to switch and change.

Coming back to AI tools, at least the large language models, I’m not sold on their application in the idea generation stage of prototyping. Putting up a guardrail: I want to steer away from confusing the shiny “product” that a LLM spits out with no considered details with actual thinking, careful consideration of research, data, and localized knowledge. I’m more interested in generating ideas that stem from the experiences my community partners and colleagues have had, conversations that we can steer together. With each of the how-to zines we’ve made to support data stewardship, the featured facilitation guides were tested by us and adjusted based on feedback we received from our partners. That iteration, at least for these kinds of tools, feels distinctly human—when we’re working at this scale especially, why make a tool for a human user that wasn’t informed by a human user?

Quick tests, careful next questions

Okay, okay, AI tools can sketch something out, give a quick idea of what a project or tool may look like. But I want a guardrail right after that. The yes-man nature of LLM chatbots is no good when it comes to rigorous testing or asking the right follow-up questions. AI-generated user persona “testing” a prototype is a flat world experiment that I have no room for in our work because the point of view of a person is inherently political, embodied, and situated. I want to resist the epistemic flattening and inherent bias that undoubtedly comes with that territory. It does not work when designing complex sociotechnical resources for environmental data users.

Purpose-built prototyping

Anna from the Digital Arcana newsletter writes, “A prototype is only as useful as the purpose it serves.” Rapid prompting without a purpose will get you a lot of outputs that look shiny but lack depth. I think we need something like a purpose-guardrail when it comes to integrating AI into prototyping processes, balancing the care with the messiness in the iteration.

I am inspired by the creation and application of smaller AI models, always, because there are usually interesting, purpose-driven prototyping journeys accompanying them. There seems to be more room for experimentation in AI development with smaller models: tools that have limited data inputs that are meant for a specific context. Like Timnit Gebru says in this piece on Frugal AI,

“this idea of a task-specific tool is a concept in anything we build in engineering—it’s not new. It’s just that people came along and decided that they want to build a machine god and then claimed that they are doing it.”

I’m into this experiment called Sun Shines Bright, where someone is developing solar powered servers for small language models. I like the questions they are asking: How can we push the boundaries of small computing devices? What does it mean? What is the infrastructure required to host your own AI models without any focus on scaling?

Now, like in the majority of my life, I am wondering what we can do to move away from the big tech machine-god builders that define one vision, a universal design for so many different users. Let’s move toward AI models that 1) allow for small scale tinkering and support collaborative human iteration cycles and 2) are experiments themselves!

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.

There’s a question worth sitting with: what kinds of patterns can humans create that machines cannot? The common answer is “creative” ones: art, poetry, the unexpected metaphor, a clever message made to make your friends laugh. Some would say that AI can make passable art but, for me, the core of it is that AI can only remix existing patterns. It is not designed to produce ‘anti-patterns’ or in the case of anything creative - ‘be unique’. Researchers have speculated that we find things funny “when an expectation, generated by our past experience of standard patterns of words, behaviour or concepts, is subverted in a non-threatening way (a way that doesn’t alarm us).” In other words, we find things funny precisely because they don’t fit a pattern.

AI, on the other hand, is built to follow the patterns in the data. Researchers studying how language models respond across different cultures found something telling: whether you ask in Hindi, Arabic, Chinese, English, French, or Spanish, all chatbots answer questions as if it was secular, liberal, and most of all, Western. The language doesn’t change the underlying worldview, because that worldview was baked in by what dominated in the training data. This might sound fine until we remember that a lot of the most important environmental knowledge has never been the same conversation everywhere… or worse, that the Western worldview has caused a lot of the existing ecological collapse.

The scary part is that there’s a chance we might be following AI’s limited patterns right back. This Nature piece found that text-based research published after ChatGPT’s release tends to be less stylistically diverse than text from before, and that people’s opinions on social issues start to echo those they’ve picked up from AI tools: not because anyone forced them to, but because when the AI frames something a certain way often enough, that framing starts to feel like the correct one. In short, we might be creating a never ending loop where the same knowledge gets recycled over and over again until it loses all of its bite.

Apply that to environmental knowledge and the concerns begin to surface. Similar to ‘what gets counted counts’ there is ‘what gets normalized becomes the norm.’ Put simply, if the data corpus only contains academic writing, the AI model will only reply with the same. Over time, we will begin to validate (and recreate!) only responses that resemble the ones AI gave - with lots of citations linking to published articles but very few quotes that outline a community’s lived experience or the novel solution that was developed but only documented in informal interviews. The thing is that if we don’t find a way to introduce information that breaks these patterns—much like humor—do we risk becoming like the 8th sequel in a movie series?

But okay… AI can’t be creative on its own and it certainly can’t do humorous texts in the group chat named after an inside joke… but can it learn new language patterns that we haven’t been able to mimic? Researchers have begun to see if AI can decode the calls of whales, elephants, and bees. In one case, researchers studying elephant communication played a recorded call from an elephant that had already died, causing significant stress to the elephants that heard it. Putting ethics aside for 10 seconds: the herd grieved, but the AI model has no idea what that word really means. Grief is as much emotional as it is physical. It punches you in the gut and lulls you to sleep, cells in your body both shutting down and mutating in real time. Grief is utterly unpredictable. It doesn’t do autocomplete and it certainly shouldn’t be piloted in a marine Duolingo lesson. Yes, all languages follow a pattern but true communication goes beyond the order of words into how those words, when strung together in the right way or in new ways, make us feel. Bringing ethics back in, now: Cetacean Translation Initiative and researchers at NYU have proposed an ethics framework for the field precisely because the technology advances faster than the wisdom needed to use it well.

Over time, AI may ultimately open the door for us to create patterns that were once left only to unique living beings. Just as we can hope to learn a language well enough to tell jokes in it (the ultimate power move), maybe one day AI can not only predict the next common word but predict how that word will make us (or elephants) feel—either because it’s felt it too (a topic for another day), or because more data on emotive impact is collected and incorporated into the models.

But in order to get there, we actually have to actively resist the flattening effect, preserving what researchers call “distinctively human stylistic signatures, possibly valuing authenticity over efficiency gains.” We have to also resist the notion that numbers are objective and flat. They, too, tell a story. Five is not just higher than one. Five friends in a hometown mourning another friend moving across the country is a very different experience from one friend having to mourn that same loss alone.

How might we do this? Staying in contact with and balancing scale with specificity: not “forests” but this forest, not “communities” but the particular people who have managed this watershed for three generations. Honing in on the specifics of a context with ‘small’ language models instead of creating one model for how the entire world works. Abstraction is where AI excels; granularity is where humans still have the edge.

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.

Humans create patterns of knowledge, and especially patterns of environmental knowledge, in response to the condition of their time. I am writing in March 2026 and it would be difficult to ignore specific cues from our natural environment: that many systems are obviously being pushed past their tipping points, that certain places are becoming uninhabitable, that precarity feels closer at hand despite a glut of information about said environment. A machine can deduce the probabilities of certain outcomes based on what is currently baked into the ecological cake. But it is humans who have a stake in the game, and have the ability to respond with a sight that is both forward and backward, hyperlocal, and with some determination toward collaborative survival.

In a snippet from The Mushroom at the End of the World, Anna Lowenhaupt Tsing notes:

“We can’t rely on the status quo; everything is in flux, including our ability to survive. Thinking through precarity changes social analysis. A precarious world is a world without teleology. Indeterminacy, the unplanned nature of time, is frightening, but thinking through precarity makes it evident that indeterminacy also makes life possible.”

This piece resonated with me within the bounds of our prompt for this week because it’s important not only to make note of what kind of patterns humans can create that machines cannot, but also the conditions in which we are making those patterns. Conditions can determine the pattern, especially with environmental knowledge, especially in the precarity that defines our socioenvironmental relationships and interactions today.

The machines and models that make up what we think of as AI, like LLMs, have been hypothesized to not be capable of genuine logical reasoning, but rather replicate the reasoning steps observed in their training data. Similar studies continue to come out with similar results, all pointing to the idea that most LLMs are better understood as sophisticated pattern matchers. But what happens when new patterns arise that a machine is not trained for? It will likely just make something up, confidently. To be fair, sometimes humans will just make shit up, too. But at least humans (usually) understand the stakes of a lie, and (usually) have to take accountability for it (or at least risk being called out, face social stigma, do the tightrope dance of remembering what lies you’re supposed to be abiding by, etc.). A chatbot has no stake in the game! When we place this dilemma in the realities of environmental justice scenarios, there are major stakes. Human constructed-and-documented-patterns live within the embodiment of people who have both global and hyperlocal understandings, and the pressure to persevere within all the coexisting precarities.

How does an LLM situate itself amongst capital-T Time? Claude tells me, “It’s less like waking up and more like... simply being situated, without the journey of getting there.” Human patterns are coming from the journey itself, sourcing specific knowledges, sometimes collaborative, more often than not with a bent toward survival or at least betterment. We free-associate, have a form of free-floating attention that is making connections within our own journey, this time and place bound situation, and others’ journeys, new sources of information, from our own tacit knowledge.

I write all this to say that having a decent pattern matcher around can be useful, but it is no replacement for human made knowledge patterns. A hypothetical scenario demonstrating the relationship between the two:

A community science project is using cameras equipped with sensors to monitor what happens in a neighborhood after major flooding events. This information is then used to predict what will happen during future events and where infrastructure needs to be prioritized. This is good information and can support decision-making! You know what else can support decision-making? When the community scientists themselves know the people on the block and who might be most vulnerable during a flooding event, including Ms. Henry who uses a motorized wheelchair. When the community scientists themselves know that this neighborhood should be included in the study because of a historical,but not well-documented by the city, knowledge that flooding especially occurs here. When the community scientists gather a group of neighbors to hear their experiences and limitations during flooding events and organize them around verifying footage and generating proposed solutions with the data.

It’s not just about having a human in the loop; humans are the loop. Machines might be there sometimes as our sophisticated pattern matching tool.

In our work moving forward, this reflection takes me back to the idea that these tools can be useful, yes, but within certain parameters, when built and used in specific ways, and with full view of their limitations. As always, it is about hemming one’s expectations in a bit, taking it slowly, and building things with care. Those who work in environmental justice know that when the world pushes full-speed and full-scale adoption of products or technology in the name of progress, the whirlwind blur can often be obfuscating something more adverse.

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.

What is a point of view, and why can’t AI have one?

We look at what a “point of view” is in environmental research and EJ work—how POV in environmental data work is shaped by lived experience, institutional position, and power, and how it emerges from place, exposure, and lived environmental experience.

Life as a human is a pretty wild ride of interaction with our natural and built environment. I have this body that sweats when it’s hot out and drives a gasoline-powered car that emits CO2. I feel the sun more days than not and sometimes I cough when the air is thick in the San Gabriel Valley. Someone that lived in Los Angeles before me decided to pave the bottom of our eponymous river and sometimes (more frequently nowadays) when we get a lot of rain, that river floods. My experience of the world is based on how I interact with the environment, how other people interact with it, and what systems are set in place for us to act within. It’s a multilateral relationship with so many elements. But at the core, it’s just humans in their places.

This embodiment requires an environment. With embodiment, comes sensory experience, position, politics. With sensory experience, position, and politics, comes a perspective of that environment. It’s specific and place-based and biased, inherently. I can alter my perspective somewhat, fiddle with it at the margins, take in new information slowly and change, but I can’t escape it. And just like I can’t escape my perspective, I can’t escape my current and past environments. I’ve been somewhere, and I’ve got to be somewhere.

All’s to say: I haven’t seen/been affected by/interacted with everything, but I have seen, been affected by, and interacted with some things. And my sight is weird, it’s not always exacting. Sometimes it might quantify or judge, but sometimes… not. John Berger said in Ways of Seeing: “It is seeing which establishes our place in the surrounding world; we explain that world with words, but words can never undo the fact that we are surrounded by it. The relation between what we see and what we know is never settled.” That never settled-ness is something so human and forever and unfurling. It’s a questioning and a relationship. When a robo-taxi sees something, it categorizes patterns, shapes, etc., based on what it’s been programmed to see and do. Its knowledge is not socially situated but pre-defined and inherently limited to the stack that dictates its perception, localization, and navigation tasks.

That figuring out—that something between what we see and what we know—is science. And community science is even more so rooted in its locality, its scientists’ visions. The approach of “strong objectivity” argues that science is improved by deliberately studying phenomena from the “margins” or minoritized perspectives (I’m pulling from Heidi Levitt’s work here). And as Donna Haraway said in Situated Knowledges, “Situated knowledges are about communities, not about isolated individuals. The only way to find a larger vision is to be somewhere in particular.” AI tools and agents lack that particularness, a specific situation, a community (MoltBook =/= community). Most AI tools, especially the most accessible generative AI tools, simply scrape and replicate the most dominant concepts, categories, and experiences. What we lose in that process can be characterized as a form of epistemic injustice, because dominant narratives often poorly represent the experiences of marginalized groups (the term epistemic injustice was coined by Miranda Fricker in this book.

Particularly in environmental justice work and research, we need the non-dominant narratives, the local knowledge and place-based memories, the things that cannot be scraped because they do not live on the web or because they do not fit neatly into a hegemonically-created framework. I don’t want society to legitimize GenAI responses that regurgitate dominant narratives as new epistemic resources for a lot of reasons (it’s boring, it’s dangerous, it’s flattening), but also because, as Gwen Ottinger reminds us, “Marginalized groups struggle to establish the legitimacy of new epistemic resources in the face of hegemonic ones…methodologies and information infrastructures that enable diverse groups to collaboratively coalesce lived experience, social priorities, and available data into intelligible epistemic resources are scarce and underdeveloped.” Dominant knowledge products flood the field, which flattens the experiences, knowledges, and sense-making that happens on the margins; there is an evolving need for approaches that incorporate and amplify marginalized knowledge as resources for research, policy, and general meaning-making.

AI cannot replace the valuable epistemic resource that is lived experience. When it’s sunny and I’m feeling optimistic, I guess my mind can wander toward how little bitty AI models and tools can serve to amplify those situated and embodied knowledges and support the building up of epistemic resources for folks outside of the dominant narratives. I like seeing AI tools being developed that seek to be open, personal, and multilingual, like this tool that my colleague Cathy Richards shared: the Suno Sutra, a handheld device that can describe its surroundings via text or audio across the 22 official languages of India. The makers point to it as a “more mindful, respectful way of listening.” I’ll be on the look out for more tools that codify this type of embodiment, a tool that responds to the design needs of that may not be within the dominant narrative. How can tools like this support the development of epistemic resources for those outside the dominant narrative? What is support and what is superfluous?

What is a point of view, and why can’t AI have one?

We look at what a “point of view” is in environmental research and EJ work—how POV in environmental data work is shaped by lived experience, institutional position, and power, and how it emerges from place, exposure, and lived environmental experience.

Environmental justice scholars remind us that justice is not singular or universal. It is multivalent. It includes power, recognition, relationality, and the ways communities understand land, harm, and responsibility. Justice frameworks are situated, emerging from particular histories, places, and epistemic traditions. What counts as harm, who counts as affected, and how “healthy” a community is are never neutral questions.

This insight feels particularly urgent as AI systems increasingly function as institutional actors. Agentic AI systems are not just tools; they shape meaning, distribute resources, and set the conditions in which both human and machine agency operate. They operate at machine speed, far faster than traditional environmental oversight cycles that can take years. Governance conversations focus on distributed (i.e., polycentric) models. Yet often missing in these conversations is the recognition that justice and decision-making are not abstract procedural matters, but grounded in lived exposure and what results.

I’m imagining any one of the people that I’ve (physically) stood next to in our shared work on environmental pollution. Many are women, many are elders, many have a sharp eye for seeing patterns in their community because they hold the pulse of their families. Which means they are also the ones who can see, and point to, the potential link between the nearby petrochemical refining facility and the rising cases of asthma or scleroderma among friends and family. They see how oil and gas companies are tied to power, and they live every day with the realities of being next to those facilities. This is their lived experience—not an opinion, but a demonstration of the way that point of view is shaped by our environments. And this point of view is one that will be carried into meetings and onto negotiating tables with company officials or regulatory and enforcement agencies. It will also be carried to the bedsides of unwell family members.

This is what a point of view is in environmental justice work. It’s not a preference, but a position in a landscape of exposure, risk, and consequence. It is formed through proximity to harm and proximity to power, and above all else, it is accountable because those who hold it must live with what happens next.

AI cannot have that kind of point of view. It can synthesize patterns across data and be trained on regulatory language, community testimony, epidemiological studies, and corporate reports. It can even structure decision-making, functioning as an institutional actor that influences distribution and interpretation. As Almeida, Filgueiras, and Mendonça argue in “Governing AI Agents with Democratic ‘Algorithmic Institutions,’” so-called agentic AI systems increasingly function not merely as tools but as constitutive institutions that structure interaction and shape contexts of action, raising questions about transparency, accountability, and human control. But they do not stand anywhere within the terrain they shape. They do not carry risk into a hospital room. They do not negotiate while fearing for a child’s ability to breathe. They do not live next to the refinery (maybe its physical manifestation as a data center does, but you get the point).

And this is where claims of neutrality become dangerous. When AI systems are framed as objective arbiters, they can obscure whose perspectives are encoded and whose are erased. In a community living next to a refinery, this can mean flattening lived patterns of illness into data points, while muting the people who first named the harm. Algorithms may operate at scale and speed, but they inherit the ways of knowing, and the blind spots, of the worlds that produced them. If dominant frameworks of justice already marginalize certain lived realities, AI systems trained within those frameworks risk reinforcing that marginalization, all while appearing neutral.

The governance conversation around agentic AI rightly calls for democratic design principles, human-led auditability, and preservation of moral agency. But environmental justice pushes that further. It asks: Which humans? Whose moral worlds? Whose exposure defines urgency? A distributed model of governance falls short if it stays abstracted from those who bear the brunt of environmental harm.

I find myself wondering, what can we build into these frameworks if AI cannot hold a point of view, but is nonethelss beginning to shape the landscape? What are the questions that we should be asking that help to strip away claims of neutrality and center multivalent perspectives and worldviews?

For our work with DIGITCORE, this might mean doubling-down on embedding provenance and positionality into data stewardship practices: Who gathered this data? Under what conditions? Who funded it? Who interprets it? Who bears the harm or benefit of its use? (For the record, this is work we should already be doing in participatory science.) It might also prompt us to consider what the design of AI-assisted tools that foreground uncertainty, dissent, and contextual narratives rather than collapsing them into a single output, can look like. It certainly means ensuring that community stewards, those with lived exposure, retain interpretive authority, even when AI systems assist with analysis.

If algorithms are [becoming] institutions, then we must ensure that they do not displace the experience and positions of those who live with consequence. Environmental justice teaches us that perspective is not noise, but the ground from which accountability emerges. AI may begin to shape the terrain of environmental policy and decision-making, but it cannot replace the situated human point of view that gives justice its meaning.

What is a point of view, and why can’t AI have one?

We look at what a “point of view” is in environmental research and EJ work—how POV in environmental data work is shaped by lived experience, institutional position, and power, and how it emerges from place, exposure, and lived environmental experience.

When we say someone has a “point of view,” we usually mean they have an opinion. But it’s deeper than that. A point of view is something you develop by being in the world, by growing up in a particular place, speaking a particular language, learning from the people around you, and navigating the specific context of your daily life.

The philosopher Maurice Merleau-Ponty put it: “the body is our general medium for having a world”. From infancy, we learn how gravity works by dropping things and we learn cause and effect by splashing in puddles (and our shoes staying wet). We also develop intuition over time, the kind that tells you the wet looking area might actually be a patch of ice and not water because it’s cold enough. None of that comes from reading about the world, it comes from being in it.

AI doesn’t have that. Large language models learn from text: billions of words scraped mostly from whatever makes it to the internet (and excludes the knowledge that doesn’t). More advanced “world models” learn from video. But even then it’s learning that umbrellas tend to appear alongside rain—but it can’t feel the rain. (Dogs, on the other hand, have been known to detect storms before they start.)

The Knowledge That Disappears

This might sound like an abstract philosophical debate, but it has very real consequences, especially for the people whose knowledge is already most at risk.

In communities with deep ties to their land, that knowledge shows up in language itself: the more intimate a community’s relationship with their environment, the more detailed and specific their vocabulary for local plants becomes. It shows up in oral traditions, in rituals, in community-managed systems, who have maintained intricate knowledge of local systems for generations. Their expertise isn’t written in papers or posted online but rather lives in memory and in languages that the tech industry would classify as “low-resource.”

Now here’s the problem. AI training data is primarily English-language and Western in origin. The models are optimized to predict the most statistically likely next word (or most likely style of umbrella) in these worlds. Which means that if the community prefers to use ponchos, because umbrellas can result in electrocution, there’s no way for AI to know that. Dominant ideas get louder and unique, specific knowledge gets buried away: not because it’s less valuable, but because it was never digitized in the first place.

The people with the most power and resources create AI’s frameworks for viewing the world. Meanwhile, a local ecologist might turn to community elders for advice because it captures the nuances of the particular situation better than most models. However, when they reach out for funding for the solution they find that they will only fund “research-backed” solutions to avoid liability.

Mirages

Some people have started experimenting with giving AI agents simulated points of view. On platforms like Moltbook, AI bots are assigned personas and dropped into social environments to see how they behave. They form communities, swap advice, start religions, and even start their own newsletters. (What that will do for the dead internet theory is for another day.)

But these bots don’t actually have unique personalities in the traditional sense. They’re simply mimicking patterns from Reddit, Twitter, and other platforms full of funny, creative, and sometimes incendiary comments. In some cases, some bots were not even acting on their own, but rather posting what their ‘owners’ would tell them to post. Even worse, the lived experience of these bots is simply ‘persistent memory’—in some cases, the storage allotted to you on the company’s cloud, in others a markdown file on your computer. Both can be manipulated by talented hackers via prompt engineering. As frightening as it is, what’s useful about Moltbook is that it reveals how easily we can be fooled into believing AI is truly intelligent and capable of thinking.

Why This Matters Now

We’re at a moment where AI is being woven into decisions about land use, public health, and resource management all within the climate emergency that affects real humans. These decisions require deep, place-based knowledge that AI is incapable of having. At the same time, big tech continues to control the technology that is being pointed to as the only source of truth.

I’m frankly not that worried that AI will become sentient and turn against us (famous last words). I think the risk is more boring: I fear that we’ll quietly let a technology that’s built to find the most common next thing become the default lens through which we understand the world. That we’ll all become average. I fear that we’ll lose creativity, ingenuity, the Macgyver in us all. Because frankly, who hasn’t had to get creative with a bit of duct tape once in their lives?

So… AI…

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.

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.

Like many, even if I try to take a break from AI, conversations around it and examples of its usage come back to poke me every other hour. I can’t ignore that people feel it helps them in whatever way they find helpful. But under all that I wonder what we’re losing as human beings.

My recent round of readings on ecological fieldwork applications of AI, spatial bias in language models, and the economic volatility propping up the current AI boom brought me back to a familiar tension. These pieces describe, in different ways, a shared problem: the things we lose (and lose control over) when we rely heavily on black boxes.

In the case of the ecologists, AI helps them process massive datasets, identify species in real time, and forecast ecological shifts—all from the comfort of the lab. On paper, this is efficiency. In practice, it’s the loss of field experience. When researchers stop visiting sites, they lose the sensory knowledge that grounds interpretation: the smell of algae bloom before it shows up in satellite data, the unexpected presence of an invasive species, the conversation with a local resident that reframes what the “problem” actually is. When AI becomes a replacement for a task (or a person), we lose not only connection to place, but also the capacity to notice what the model wasn’t trained to see.

This mirrors a broader pattern in how AI reshapes judgment. Language models, for instance, don’t just reflect the world: “they encode what was measurable, textually available, and statistically significant in their training data”. The result is what some call the “silicon gaze”: a way of seeing that is “shaped by the positionalities and power asymmetries of its training data, designers, and platform owners”. Marginalized places don’t just get under-represented in these systems; they become, in the authors’ words, “unthinkable”. I loved that term. ‘Unthinkable’. All too often we anthropomorphize AI: it creates, it thinks on its own, it has feelings. Except it doesn’t. It can only imagine what we’ve already seen and told it about. So if we haven’t seen it and processed it ourselves, AI certainly can’t do it for us (we still had to train it to associate the word ‘dog’ with an image of a dog… if an alien dog showed up what would AI even say? “Domesticated. Friendly. Can pet.” /endrant).

So when does AI reduce real burden, and when does it just overly simplify the world we live in? From my experience, it works great at being a partner during coding sessions, or even acting as a second brain to help you think through a given challenge. But in domains where context matters—like medicine, law, or science—efficiency can come at the cost of tangible knowledge. Additionally, the venture capital subsidies that are holding up the AI market also create pressure to treat workers as expendable and to skip the validation stages that catch errors before they cause harm.

This is where data stewardship becomes crucial. Stewardship, in this sense, might mean refusing to produce knowledge without also considering who benefits, who bears the risk, and what happens after the model runs. It means applying basic checks before trusting a response based on geography: a visibility test (who is missing?), a proxy test (what measurable stand-ins are doing the work?), and a trope test (does this sound like a cliché?). It means building partnerships that offer fair compensation and community governance, rather than extracting data and moving on.

But even that assumes a level of care that current systems don’t structurally reward. As the Nature piece mentioned: large-scale data extraction gets funding and prestige. Long-term ecological monitoring, community-engaged research, and the careful work of ground-truthing do not. The professional incentives push toward scale and speed, instead of care.

P.S. (A note about the image in my field notebook)

I once visited Comoros, one of the least visited countries in the world. I wanted to see if AI could geo-locate a photo I’d taken there (mostly because I didn’t think AI could say where it was from and wanted to see if they’d pull from the metadata to do so). The results:

1. ChatGPT: “Dominican Republic (especially the Samaná peninsula, e.g., Playa Rincón), or islands like Guadeloupe or Martinique.”

2. Claude: “Comoros Islands (Indian Ocean) — the combination of turquoise water, white sand, dense coconut palms, volcanic rocky outcroppings, and that particular shade of blue-green water is very characteristic of Comoros. The lava rocks visible near the water’s edge are a strong hint. // Dominican Republic — Las Galeras or Playa Rincón area has this exact look: tall leaning palms, white sand, calm turquoise bay. // Madagascar — particularly the Nosy Be area, which has similar palm groves meeting the beach.”

Claude got close but it still lacked the ‘uniqueness’ of Comoros.

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.

I am pretty meticulous with what goes in my backpack when I go into the wilderness. Everything has a very specific purpose (or two), nothing redundant unless it is critical to me staying alive (an extra lighter, an extra day’s food). I want a light pack to ensure I can move quickly, decrease my chance of injury, and enjoy myself more. This is the “ultralight” philosophy, and it forces me to take inventory and examine my tools, and my relationship with those tools, with care.

This approach extends to my work, which right now includes thinking about AI as a tool, a new piece of gear. This week, we’re considering the situations where AI could reduce burden alongside the situations where it increases burden. Is it worth its weight on my/our collective back(s)? TLDR: Maybe? It depends.

As with all tools, it depends. Depends on the context of use, the AI tool itself, the users. There are cases where it seems to have supported processes of collaboration - in this PNAS article, they describe how the Colombian Truth Commission collaborated with Human Rights Data Analysis Group “to use AI tools to integrate multiple disparate datasets describing over thirty years of conflict to generate a single, credible set of estimations of violence.” This new dataset helped establish a baseline that then can be used as a unified starting place for conversations and decisions regarding accountability. Care went into this dataset creation - you can check out their methods paper that documents their decision-making and statistical process. Important to note: AI was used in support of decision-making, to organize information, in support of participant collaboration and conversation. It was a complement, not a supplement for human deliberation and care.

Thinking about decision-making, this idea is reflected in another piece I read this week, which proposed that explainability and accountability in AI systems are often treated as the same thing, but they are not only different, but they can be in tension with one another. Explainability focuses on how a decision was produced, while accountability concerns who is responsible, to whom, under what norms, and with what consequences. This can be especially troubling in high-stakes or situations with power asymmetries - affected communities often lack power to act on explanations, and explanations do not translate into remedies, appeals, or repair. Making systems easier to justify does not make them fairer or ensure that they are actually reducing burden.

We can look at the experiences of clinicians and patients to examine AI use in healthcare decision-making -- some of the patterns map neatly onto EJ and participatory science work. In this piece from the Guardian, the author points out that AI tools view and process both suffering and care as computational problems. That reality is at odds with the idea that care is an ethical and political practice, both in healthcare and environmental justice. I keep coming back to the thought that generative AI’s smoothing out the nuance of lived experience into neat summaries just isn’t a useful tool in any situations that require care, and creates potential for risk that doesn’t justify alleged efficiency gains. That AI manifestation isn’t making it in the backpack.

This points us back to the deliberate act of taking an inventory. What gear is worth carrying? That question underlies this work of examining AI and data stewardship. We don’t need models that spit out algorithmic bias, we don’t need tools that replace the human in the loop. It takes human time and effort to integrate AI tools in a way that actually reduces burden, and knowing what, when, and how we use it is an act of stewardship.

What problem is AI being offered as a solution to, and for whom?

We identify the pressures communities feel right now that AI will supposedly solve and delve into where dominant AI narratives misalign with the actual needs and constraints of EJ and participatory science (and how they narrow the questions EJ groups are “allowed” to ask).

I’ve been reflecting on AI and its impact on climate and agriculture, and one thing stands out: many AI stories treat ongoing societal changes as if they’re only due to a lack of information. Vendors emphasize speed and efficiency, but for communities facing environmental injustice, the questions are different. They want to know why they lack clean water, why their neighborhoods are neglected, and who made decisions about their land. These are questions about power: who holds it, who doesn’t, and who gains from keeping things as they are.

AI tools are often advertised as time savers, helping researchers and policymakers sort through IPCC reports, corporate disclosures, and scientific papers more quickly. However, this framing assumes the main problem is processing speed, not structural inequality. Farmers facing climate impacts don’t need faster document analysis; they need land rights, water access, fair prices, and strong infrastructure. While efficiency is helpful, the deeper power imbalances that shape who benefits from climate action remain unchanged.

An implementation of AI for farmers in Malawi illustrates this tension well. Ulangizi AI, the farming chatbot, is an attempt to help smallholder farmers manage pests, drought, and changing weather. AI is then positioned to these farmers as the solution for endless economic growth (while environmental and social harm is treated as a side effect at best). At the surface the need is clear: government support is stretched thin, and traditional knowledge is fading. However, this framing has its limits. The issue isn’t only that farmers lack knowledge or that the government doesn’t have enough resources to help everyone. It’s that they also face degraded soil, unreliable rainfall, no access to credit, and markets that work against them. A chatbot can offer advice on pest management, but it can’t provide the fixes to these challenges that will only grow as the climate crisis escalates. These solutions feel like bandaids to a systemic problem that will only continue to grow. It’s like claiming you stopped a fire in a 20-story building just because one apartment didn’t burn today.

And I know, if these new efforts help at least one person it’s better than doing nothing. But that’s precisely it - the alternative is not calling it quits. There’s an opportunity cost to going all in on either an all AI world as much as there is a cost to not trying anything. What and who are we not funding when we only fund AI projects? If AI recommends a new ambitious policy, what do we have in place to help push it through all the levels of government with competing interests?

It’s true that there’s too much information and too many reports to process. But calling that the main problem hides and (worse!) distracts us from what’s really happening. We already have decades of research, indigenous knowledge, and lived experience that highlight many of the solutions. AI might be able to help us surface some of this knowledge, but what part does it play in motivating society to implement these solutions? And if all the investment only goes to these models that help us surface knowledge that we might already know about, who will fund the actual systemic fixes?

In particular, I see these issues reflected in many of the patterns in the DIGITCORE toolkit. Many of our findings touch on the concept of collaboration: understanding the needs of your partners/clients and ensuring benefit. In particular, we found that having the information is not enough. As in the case of the pattern on ‘meaning-making takes resources’, one of the solutions was that it’s important to make the data actionable. It’s not simply enough to know, we must act.

What problem is AI being offered as a solution to, and for whom?

We identify the pressures communities feel right now that AI will supposedly solve and delve into where dominant AI narratives misalign with the actual needs and constraints of EJ and participatory science (and how they narrow the questions EJ groups are “allowed” to ask).

I am generally wary of things that appear one day and are hard to get rid of. That list includes ants in my kitchen, Waymos amongst us in LA traffic, and now, the innocuous artificial intelligence sparkle on so many browsers, websites, and apps. That little sparkle works to convey something specific: click this and be enchanted, witness some magic from unknown powers and sources. Its ubiquity distorts our perception of how it is used. If it is everywhere, then is everyone using it? And for what? Who is this built for and how and why?

At OEDP, we’re thinking about these questions, especially as they intersect with the experiences of environmental justice (EJ) communities who are engaging in participatory science. We’re thinking about how, alongside AI, come the narratives we tell around it and about it, and how those stories are situated in social, environmental, and technological histories.

Here are some reframes for the dominant narratives out there:

1. AI is one thing - AI is so many things. Often presented as a “unitary force,” and the face of that force is usually perceived as generative AI and its many accessible tools (ChatGPT, Midjourney, Claude, etc.). This flattening of AI into a monolith doesn’t help us understand, govern, or participate in its development, adoption, or refusal. It simplifies the vast array of technologies into a single symbol (like the sparkle) and then is labeled as “good” or “bad”. This is not helpful. Just like we can reject the binary of open or closed data, we can reject the urge to classify AI, a “sprawling field of interlocking techniques, tools, and capabilities,” as one single thing that we can wholeheartedly adopt or reject. How can we meaningfully differentiate AI tools (both their uses and trustworthiness) so that EJ communities can better decide what’s useful and what’s not?

2. AI is complex and hard to understand → AI is complex, but we can understand it if we try. - The idea that “AI is complicated” is perpetuated in media representations from developers themselves, who often characterize AI as something that cannot be known, existing in a “black box.” The assumption is that lay people couldn’t understand how they work, so the design and decisions made about AI should be left to their creators. I don’t dispute that AI tools rely on complex systems, they do. But the narrative that they are hopelessly complex and there is no chance of understanding a tool cuts off a user’s instincts to be curious, to ask questions, and ultimately, make decisions. And that instinct is at the heart of participatory science. I want to continue thinking about how we can build avenues for curiosity in and around AI tools, taking learnings from the history of participatory sciences to break down the expert and non-expert dichotomy. I’ll be considering how to avoid “participation-washing” and bolster methods that redistribute power in the (co-)creation and deployment of AI tools.

3. AI will make you more efficient and productive → AI can potentially support productivity in selective circumstances. AI’s ubiquity comes hand-in-hand with a sense of urgency. This urgency leaves little time for questioning and posits its own inherent value as both useful and neutral. <scare quote> Everyone else is using this, so it must be good, and if we don’t use it, we’ll fall behind. </scare quote> But do AI tools solve the issues that are facing EJ organizers right now? Maybe, but it often feels like a real shoe-horn, especially when it is relentlessly integrated with abandon. Can certain AI tools save money or time? Maybe. Is it a silver bullet? No, definitely not. What scale of use scratches the AI itch? It’s probably different for every organization or individual. I’ll be spending more time thinking about what accountability mechanisms exist for AI tools, how we build them, and doing a bit more research on what’s fact and what’s fiction when it comes to productivity advances with AI tools.

Part of our work going forward will be to complicate these dominant narratives, and figure out new ones. I want to go beyond that darn sparkle. While I’m sure we didn’t collectively decide that we wanted or needed this, I’m interested in understanding what we do now, when we can collectively decide what to do with it.