These Tools Are Ours

On Watt-Hours, Superstition Masquerading as Enlightenment, and the Unilateral Self-Disarmament of the People Who Were Supposed to Save Us

I am tired.

I am tired of being talked down to by weakly informed influencers.

I am tired of the finger-wag. I am tired of the superstition dressed as analysis.

Most of all, I am tired of watching people who want to save the world disarm themselves unilaterally in the name of process purism — while the people they want to save the world from are busy running twenty instances of GPT-5 to optimize the next round of tax evasion, the next round of zoning displacement, the next round of predictive policing, the next round of PR containment for whichever Epstein-adjacent billionaire got caught this week.

It is, in the most literal idiomatic sense, cutting off our nose to spite our face.

That is the actual situation. And if you are a person of conscience currently being told that touching these large language model (LLM) text-prediction and analytical tools makes you impure, I need you to hear me very clearly:

You are being played.

Not by Anthropic. Not by OpenAI. Not even mostly by the AI companies, which have their own serious problems. You are being played by a pattern of argument that sounds like resistance and functions as surrender.

Let’s do the numbers. Let’s do the history.

And then let’s talk about what it would actually look like to fight.

---

The Numbers

The core numbers in this article do not come from AI companies. They come from independent researchers, academic institutions, and investigative journalism outlets whose job is to be skeptical of those companies. Here’s the sourcing lineage:

MIT Technology Review. In May 2025, reporters Casey Crownhart and James O’Donnell published what they called “an unprecedented and comprehensive look at how much energy the AI industry uses — down to a single query.” They had to do this by working with open-source models as proxies, because when they asked OpenAI, Microsoft, and Google directly for their numbers, those companies declined to share specifics on how much energy their closed-source models use. MIT Tech Review did the math anyway, using open-source models they could actually measure.

Epoch AI. An independent nonprofit research institute that exists specifically to evaluate claims about AI capability and impact. In February 2025 they published a detailed methodology for estimating per-query energy consumption. They are not sponsored by, employed by, or beholden to any AI company.

University of California, Riverside. The research group led by Pengfei Li published the foundational peer-reviewed academic paper on AI water consumption. Their work has been cited across every serious piece of AI environmental journalism since.

Belated corroboration. Months after the independent researchers had already published their estimates, the AI companies finally started releasing their own numbers — Sam Altman in a June 2025 blog post, Google in an August 2025 technical report. Those numbers landed in the same order of magnitude as what the independent researchers had already calculated. That is not a case of the companies setting the narrative. That is a case of the companies being forced to confirm what outside experts had already established. The corroboration runs in the correct direction.

With that foundation locked in, here is the number.

A typical text query to a modern AI chatbot consumes approximately 0.3 watt-hours of electricity.

A note on scope: this figure is for a typical text query to a mid-sized chatbot. Reasoning models that "think" before answering, image generation, and video generation are substantially more expensive — in some cases by one or two orders of magnitude. When I say "AI query" in this piece, I mean the simple text chat case, which is what most individual users actually do most of the time.

Epoch AI’s February 2025 analysis arrived at this figure independently. MIT Technology Review’s investigation put medium-size open-source models in the same range. Google’s later disclosure for Gemini: 0.24 Wh. OpenAI’s later disclosure: 0.34 Wh. Four separate lines of evidence, two of them from researchers with no financial interest in the answer, all converging.

The old viral figure of “3 watt-hours per query” comes from a 2023 paper by Alex de Vries that was modeling a specific scenario — full integration of generative AI into Google-scale search — using early-2023 hardware assumptions. It was a reasonable estimate for the question it was actually asking, but it has since been lifted out of context and applied to per-query chatbot use, where it runs roughly ten times too high for modern models. It is still the number being quoted in Instagram videos by people who have not updated their sources in two years.

---

What Does 0.3 Watt-Hours Actually Mean?

Let me anchor this in the physical world.

One gallon of gasoline. The EPA’s official energy-equivalence figure is 33.7 kilowatt-hours per gallon. That is 33,700 watt-hours. Divided by 0.3: one gallon of gasoline contains the energy equivalent of roughly 112,000 AI queries. A ten-gallon fill-up — the kind many do without a second thought — is over a million queries of energy, most of which gets converted to waste heat by an engine that is about 25% efficient.

One studio apartment, one month. Around 300 kilowatt-hours of electricity for modest use. That’s 300,000 watt-hours. One million queries. Sending a thousand AI messages a day, every day for a month, would account for roughly 3% of your apartment’s electricity use. Nobody sends a thousand a day. Heavy professional users like me run 150 to 300 on busy workdays.

One nice hot bath. Filling a 40-gallon tub and heating it from tap-cold to 104°F takes around 4.7 kWh of pure thermodynamic energy — and efficiency losses push that higher in the real world, not lower: a standard gas water heater at ~60% efficiency lands closer to 7.8 kWh (7,800 watt-hours) to deliver that same bath. That’s roughly 26,000 queries. I would have to run over 200 queries a day and would have to work for a hundred days to match the energy of one bath. A weekly bath habit, across a year, outweighs a thousand-query-a-day habit across the same year.

One house cat. A domestic cat eats around 225 grams of meat-based food a day, and producing that food — chicken, fish, beef byproducts, the whole obligate-carnivore supply chain — runs somewhere between 5 and 25 kWh per kilogram depending on the meat mix, with 20 kWh/kg at the higher end for beef-heavy formulations when you account for feed, processing, refrigeration, and transport. That's 1,640 kWh per cat per year, or about 5.4 million queries. A 15-year-old cat has burned the energy-equivalent of over 80 million AI queries just being fed. A two-cat household operates at a scale no human user could match in a lifetime of typing.

One hour of HD streaming: around 300 queries worth of energy. One cheeseburger lifecycle: more than tens of thousands of queries. One cross-country flight: more than a human could physically type in a lifetime. One almond water footprint: around 400 queries. One pound of beef water footprint: around 750,000 queries. (To compare water footprints to query footprints, I’m using the ~0.26 mL of water per query that Google disclosed for Gemini as the bridge figure.)

I am not telling you to stop eating, driving, flying, having hot baths, and a cat. I am telling you that if the person lecturing you about AI’s environmental cost is not applying the same moral urgency to any of those other things, the environmental argument is not the actual argument. The environmental argument is the vehicle. The vehicle is not the destination.

The destination is usually purity. And purity is not a political program. It is a feeling, meaningful to the person having it, useless to the people who were supposed to be saved by the political work it replaced.

---

What the Per-Query Numbers Don’t Capture — And Why It Still Doesn’t Change the Answer

Before I continue, I want to take the strongest version of the counterargument seriously, because the per-query framing has real limits and I’m not interested in winning by avoiding them. Here are the concessions I think a thoughtful critic has actually earned, stated as clearly as I can state them, followed by why I still land where I land.

Aggregate demand is the real story, not individual queries. The environmental critique of AI has never really been about what a single query costs. It has been about the buildout — the ten-gigawatt Stargate facilities, the dozens of new hyperscale data centers, the 4.4% of US electricity now flowing to data centers (up from essentially flat demand between 2005 and 2017, as MIT Technology Review has documented). The per-query number, this argument goes, is a bait-and-switch: it makes the individual user feel absolved while the industrial footprint metastasizes in the background. By normalizing daily use, you are generating the demand signal that justifies the buildout. That signal is the problem, and your 0.3 watt-hours is its smallest currency.

I think this is partly right. Yes — individual use aggregates. Yes — aggregated demand shapes infrastructure decisions. Yes — the buildout is the real scandal, not the query. Everything I have written about siting, water rights, grid capacity, and public ownership is an acknowledgment of exactly this point. But the inference “therefore do not use the tool” does not follow from the premise. Two reasons.

First, the marginal demand signal from conscientious individual users is vanishingly small compared to the signal from enterprise contracts, government procurement, default integration into search and email and office software, and the speculative capital bet of the companies themselves — which would continue building at full speed even if every individual user quit tomorrow. The buildout is not waiting for your permission.

Second, the logic of “your use creates the demand that justifies the infrastructure” applies with identical force to every piece of modern infrastructure ever built — roads, grids, airports, the internet itself — and we do not generally conclude from this that conscientious people should abstain from roads. We conclude that the governance of the infrastructure has to be fought for. The demand-signal argument is not an argument against use. It is an argument for regulation, taxation, public ownership, and direct political pressure on the people building the thing. All of which are in this piece. None of which are helped by you quitting Claude, Gemini, even ChatGPT, or the number of other public access tools.

Training costs are real and the per-query number does not capture them. The energy and water figures I cited are inference numbers — what it costs to run the model once it exists. The enormous upfront cost of training the model in the first place is a separate line item. GPT-4’s training run has been estimated in the range of 50 to 60 gigawatt-hours. That cost is paid once and then amortized across every query the model ever serves, but amortization is not magic. The cost is real. An honest accounting has to include it.

So let’s include it. A training run of 50 GWh, amortized across (very conservatively) a few billion queries over the model’s deployment lifetime, adds somewhere between a fraction of a watt-hour and a few watt-hours to the per-query cost, depending on how you slice the math. Take the pessimistic end and assume training adds 3 Wh per query on top of the 0.3 Wh inference cost. That moves the gallon-of-gas comparison from roughly 112,000 queries per gallon down to somewhere around 10,000 queries per gallon — and yes, that revises my own headline number from earlier in this piece. I want you to see the revision happen rather than discover it on a re-read. Ten times worse than my headline number. Still an order of magnitude better than most of the energy expenditures you make without a second thought. The training cost matters. It does not change the conclusion.

Jevons paradox is real and the efficiency story is a double-edged sword. Another critique deserves naming on its own: every efficiency gain in a widely used technology has historically been met with a proportional — sometimes more than proportional — increase in use. This is Jevons paradox, named for the nineteenth-century economist who observed it in coal. Making something cheaper and more efficient makes it more embedded, which increases aggregate demand faster than efficiency can contain it. The fact that my per-query numbers have dropped tenfold in two years is, on this reading, not good news. It is the mechanism by which AI is getting woven into every application, which is how we end up staring at 4.4% of national electricity going to data centers and climbing. The efficiency is the trap.

Jevons is a real phenomenon and it absolutely applies here. But Jevons paradox is an argument for structural regulation of the expansion, not an argument against use by individuals. You do not solve rebound effects by asking conscientious consumers to abstain; you solve them by imposing hard constraints on the expansion itself — carbon pricing, grid capacity caps, siting restrictions, mandatory efficiency standards, public ownership of compute allocation, demand-side regulation. Those are the levers. An individual quitting Claude does not touch any of them. An individual organizing for carbon pricing touches all of them. The per-query number is still relevant to the individual moral question, even if it is misleading about the aggregate industrial trajectory. Hold both.

And one last concession I owe you, because it is about me and not about the numbers. A critic of this piece can fairly point out: Immanuel is writing sincere pro-AI-use content, which is reputationally valuable to Anthropic in a way no paid advertising could match. His independence and his critical voice are what make the endorsement valuable. He is, whatever his intentions, doing PR work for a company he claims to critique.

Yes. A sincere argument from an independent writer of conscience that AI tools are worth using is, in fact, a reputational asset for the companies that make those tools. I am aware of that. I am accepting that cost consciously because the alternative — letting the framing of this question get set by people who are wrong about the numbers, wrong about the strategy, and wrong about whose hand this tool should be in — is worse than a small PR dividend flowing to a company I critique in the same breath. I am not going to pretend my writing exists in a vacuum free of consequence for the people I am also criticizing. It doesn’t. I am making a judgment call about which cost is higher, and I am making it out loud so you can weigh it for yourself. That is the honest posture. Silence on this point would be dishonest in a way I am not willing to be.

---

Where the Real Harms Actually Are — And Why Quitting Doesn’t Touch Them

Honesty about my own position aside, the harms this conversation is supposedly about are real. The AI industry has real, documented, serious problems.

Data center siting in water-stressed regions is a genuine crisis. The Aragón protests in Spain — “your cloud is drying my river” — are legitimate and I stand with those communities. Google’s Cerrillos facility in Chile pulling millions of liters a day from a drought-stricken aquifer is indefensible. The cobalt supply chain through the DRC is an ongoing atrocity that implicates every piece of modern electronics, including the phone used to record the video telling you AI is evil — I made that argument in full in Spiritual Darwinism and I am not going to re-litigate it here. Training-run energy costs are enormous and opaque. Corporate concentration of AI capability in the hands of a handful of companies is a serious democratic threat. Labor practices around data annotation, often outsourced to low-wage workers in Kenya and the Philippines under traumatizing conditions, are real exploitation that deserves organized response.

All of that is true. All of it deserves action. None of it is addressed by an individual user deciding not to use Claude.

Not one data center gets un-built because you cancelled your subscription. Not one cobalt mine closes. Not one gallon of groundwater returns to an aquifer. Not one displaced community gets their land back. The individual-consumption frame is a category error — a policy-scale problem being assigned a personal-behavior solution, which is convenient for exactly the people who profit from policy remaining unchanged.

Meanwhile, here is what is happening on the other side of the ledger while the resistance debates whether Claude is spiritually clean:

ICE and CBP are running predictive systems to flag people for deportation. Palantir is selling integrated intelligence platforms to police departments, militaries, and welfare agencies. Algorithmic systems are denying benefits, screening résumés, setting bail, steering sentencing, and pricing insurance in ways that reproduce and accelerate every inequality they touch. Hedge funds are using large language models to run sentiment analysis on your neighbors’ social media posts to trade against their economic prospects. Real estate investment trusts are using AI to optimize rent extraction block by block. Lobbyists are using it to draft legislation faster than understaffed public-interest groups can read. Corporate legal teams are using it to bury small plaintiffs in procedural motion work until they run out of money. The ultra-wealthy class — the documented, sanctioned, named individuals whose networks of extraction we all know — are using the best frontier models through enterprise contracts that ordinary users will never see, running thousands of parallel queries at scales that dwarf anything you or I will do in a lifetime.

They are not going to stop. They could not be shamed into stopping by a viral Instagram video if the video were delivered by an archangel. They have no community to be ostracized from. They do not care what the decolonized-spirituality corner of Substack thinks of them. They are busy.

And here is the asymmetry that keeps me up at night: we are being told to put the tools down while they are picking them up. That is not resistance. That is unilateral disarmament sold as virtue. And the people doing the selling, whatever their intentions, are running cover for the outcome the ruling class most wants — which is a population of conscientious objectors who voluntarily remove themselves from the cognitive battlefield.

---

A Name for What’s Happening: Process Purism

I want to give this pattern a name, because once a pattern has a name you can point at it when you see it in the wild, and pointing at it is how we get out from under it together.

Process purism is the insistence that the means of doing justice work must themselves be spiritually, ethically, and aesthetically unblemished — regardless of whether that insistence actually produces justice outcomes for the people the work was supposed to help. It judges the tool, the supply chain, the vocabulary, the aesthetic, and the affiliations of the person doing the work more rigorously than it judges the results of the work or the conditions of the people the work was claiming to serve.

I’m not going to name individual influencers, but if you spend any time in the wellness, decolonial-spirituality, eco-activist, and/or “leftist” corners of Instagram and Substack, you have seen this pattern, and you know exactly the register I mean.

The process purist is not lying about caring. They genuinely care. This is the part I want to hold onto, because I’ve spent over twenty-five years in rooms where this pattern was refined and I know most of the people inside it are good people with intact hearts.

What has happened is that the caring has been rerouted — from an outward-facing concern about outcomes in the world to an inward-facing concern about the purity of one’s own participation in the world. The energy that would have gone into organizing, building, writing, lawyering, legislating, or simply showing up has been redirected into auditing. Auditing oneself. Auditing one’s peers. Auditing the tools and practices and vocabularies and collaborators and technologies of everyone in the movement.

It sounds like rigor. It functions as paralysis. And because it feels morally serious — because the person practicing it is genuinely giving up convenience and capability in the name of principle — it becomes almost impossible to critique from inside the community without being reframed as the impure one.

Let me say clearly what this is NOT

I need to be precise here, because this argument gets misread fast and I refuse to be the guy who handed the Silicon Valley boosters a quote they could use to wave away legitimate concerns. So:

Process purism is not the same as caring about means. Means matter enormously. How we do the work shapes what the work becomes. A movement that treats people cruelly in the name of justice produces cruelty, not justice. A campaign built on lies produces a world made of lies. These are real and I hold them.

Process purism is not the same as boycott. Targeted economic pressure on a specific company for a specific harm, organized collectively with clear demands and a theory of how the pressure translates into change — that is a tactic with a long and honorable history and I support it. The grape boycott worked. The bus boycott worked. Divestment from apartheid South Africa worked. Those were not process purism. Those were strategy.

Process purism is not the same as harm reduction. If you have looked at the evidence and concluded that a particular tool, practice, or affiliation is genuinely causing more harm than good in your specific case, and you step back from it on that basis, that is informed conscience in action and I honor it. That’s not what I’m talking about either.

And most importantly: process purism is not what I’m arguing against when I say “take the tool.” I am absolutely not arguing that ends justify means. That phrase — “the ends justify the means” — is the signature move of every authoritarian in history, and I want nothing to do with it. Nothing in this article asks anyone to ignore the real harms of the tech-industrial complex. The cobalt supply chain is real. The water crisis is real. The labor exploitation is real. The corporate concentration is real. The displacement of communities near data centers is real. The predatory surveillance applications are real. None of that gets waved away by anything I am saying, and none of it should.

What I am arguing is something narrower and more precise. I am arguing that there is a difference between means that are genuinely harmful and means that merely feel impure to a particular cultural community at a particular moment — and that process purism collapses that distinction. It treats “this tool has some problems in some applications” as equivalent to “this tool is categorically forbidden for people of conscience.” Those are not the same claim. The first is analysis. The second is taboo. And taboo is not a substitute for analysis, no matter how beautifully it is delivered.

Why process purism is off

Here is why the pattern fails, even when — especially when — the person inside it has the purest intentions.

It mistakes abstention for action. Not doing something is not the same as doing something about it. Refusing to use Claude does not shut down a data center. Refusing to buy a smartphone does not free a Congolese child from a cobalt mine. Refusing to fly does not ground the private jets. The refusal is an ethical gesture aimed at the self, not an intervention aimed at the problem. Process purism confuses the two and rewards the first as if it were the second. It feels like doing something because it costs something. Cost is not the same as effect.

It scales backward. Real political movements expand capability. Process purism contracts it. Every new rule about what a “real” member of the movement cannot touch, cannot say, cannot use, cannot collaborate with — every new rule shrinks the coalition and reduces the operational capacity of everyone still inside it. Meanwhile the opposition has no such rules and operates at full throttle. A movement that gets smaller and less capable over time, in the name of purity, is a movement losing. By definition.

It centers the performer, not the affected. If I refuse a tool in solidarity with a community I am not part of, and that community was not consulted, does not benefit from my refusal, and in some cases would actually benefit from me using the tool to organize on their behalf — the refusal is not solidarity. It is a performance of solidarity staged for an audience of other people in my own community who will judge me by my abstention. The affected community has become a prop in a drama that is really about me and my peers.

It lets the real power off the hook. Every hour spent policing whether a writer used Claude is an hour not spent on the people who own the data centers. Every viral video aimed at individual consumers is a viral video not aimed at policy. Process purism is, functionally, a redirect — intentional or not — from structural targets to individual ones. The ruling class loves process purism because it converts potential political opponents into auditors of each other. It is the cheapest form of opposition containment ever devised, because the opposition does the containing itself, for free, with genuine moral conviction.

And in this specific case: it hands the tools to the enemy. This is the one that connects directly to everything else in this piece. Unilateral disarmament is the practical outcome of process purism applied to AI. The people we are supposed to be fighting are not disarming. They are arming up. The process purist’s response to this asymmetry is to feel clean while losing. And losing, in this moment, is not a private spiritual choice. It is a choice that affects everyone the movement was supposed to protect.

The correct question

The correct question is never “is this tool pure?” No tool is pure. The hammer isn’t pure. The printing press wasn’t pure. The internet wasn’t pure. The solar panel isn’t pure. The tractor that feeds your family wasn’t pure. Purity is not a category that exists in the real world of real materials made by real labor on a real planet with a real history. The demand for purity is the demand for a world that never existed.

The correct question is: who is this tool currently serving, who could it serve instead, and what would it take to move it from the first answer to the second?

That’s a political question. It has political answers. Process purism is not one of them.

The “Master’s Tools”

Audre Lorde’s famous line — “the master’s tools will never dismantle the master’s house” — needs to be acknowledged here because it could be argued we’re using the master’s tools and these ends do not justify the means.

Lorde was writing in 1979 about the specific practices of white, academic, heterosexist feminism and their refusal to engage with difference. She was talking about ideological frameworks. She was not talking about hammers and saws. Audre Lorde would absolutely pick up a hammer and saw to build a house.

And if she wanted to build that house in a country where the sheriff showed up armed, she would consider — as serious Black liberation thinkers of her generation did consider — whether to arm the build. The Black Panthers did not manufacture their own rifles. They bought Winchesters and Remingtons and M1 carbines at the sporting goods store, the same weapons produced by the same American arms industry that was simultaneously equipping the police forces occupying their neighborhoods. They understood perfectly well whose tools those were. They picked them up anyway, openly and legally, and walked into the California State Capitol with them in 1967 to make a point about armed self-defense that the state understood immediately and responded to with the Mulford Act.

The tool was the master’s. The use was not. That distinction — between who manufactured a thing and who decides what it is for — is the distinction this entire argument rests on, and I don’t think it was a distinction Lorde would have missed. That is not to speak for her, but to unpack the rhetoric in a way that helps clarify this particular discussion around AI tools.

What she would refuse is the story told to a culture to justify stealing the land and the lumber to build it on — the “manifest destiny” story, the “God gave us this” story, the “it was promised to us three thousand years ago” story. The tool was never the problem. The mythology of ownership around the tool was the problem.

The same distinction applies here, and it is load-bearing.

The AI models currently in existence were not built by Sam Altman. They were not built by the board of Anthropic. They were not built by Sundar Pichai. Those men raised the money and own the companies. The tools themselves were built by tens of thousands of engineers, researchers, graduate students, open-source contributors, linguists, cognitive scientists, data workers, mathematicians, and — most of all — by every human being whose writing is in the training corpus. That means you. Your emails, your blog posts, your books, your Wikipedia edits, your forum arguments, your published essays, your academic papers. The models are a crystallized average of the written output of humanity, scraped, weighted, and compressed by the labor of millions of named and unnamed workers.

The tool is a commons that got enclosed. It was built on our collective intellectual output, with the labor of workers across the Global North and Global South, and then fenced off and monetized by a class of owners who did none of the building. The correct response to enclosure is not to refuse to set foot on the land. The correct response to enclosure is to recognize it as theft and act accordingly — which means using the commons, defending the commons, building parallel open-source infrastructure, fighting the enclosure in courts and legislatures, and refusing the mythology that says the owners own it because they “earned” it.

Engineers build. Owners extract. Those are two different classes of people and they have two different relationships to the tool. Refusing to use the tool does not punish the owners — they’re using it just fine. Refusing to use the tool punishes the engineers, by abandoning the thing they built to the exclusive use of the class that captured it.

This tool is ours. Take it.

---

A Conversation I Am Not Going to Lead

One more act of honesty before I get to the close. I have been writing this article from the position of a California-based writer with a platform, decades of communications experience, and specific political commitments I have named publicly. That is the standing I have. There is another conversation happening about AI, ethics, labor, and power that I am not positioned to lead — the conversation being developed in real time by thinkers and organizers from the communities most directly affected by the extractive supply chains, the data annotation labor, the data center siting fights, and the long history of technological violence enacted against Indigenous and Global South peoples in the name of progress. That conversation has its own lineage, its own scholarship, and its own internal debates. Some of the voices in it are deeply critical of AI as it currently exists. Some are refusing the tools outright. Others are engaging with the technology directly — building language models for endangered languages, auditing large datasets for harm, organizing data workers, developing sovereign governance protocols for AI in Indigenous contexts. Both of those responses are present in the same conversation, and neither of them is mine to summarize.

So instead of summarizing, I am going to point. If the argument in this article matters to you, some of the places to find the conversation I am not leading are:

• Abeba Birhane on dataset audits and algorithmic harm.

• Timnit Gebru and the Distributed AI Research Institute (DAIR).

• Joy Buolamwini and the Algorithmic Justice League.

• The Indigenous Protocol and Artificial Intelligence Working Group and their 2020 position paper on Indigenous-led engagement with AI, and which notably includes voices that work with the technology rather than only against it.

• Te Hiku Media in Aotearoa on Māori-language AI sovereignty.

• The Data Workers’ Inquiry project documenting the conditions of AI labor in Kenya, the Philippines, and elsewhere.

I am not citing these as credentials for my own argument. I am citing them as pointers to a conversation that deserves more of your attention than this article can give it. Go read them directly. They are the people you should be hearing from on the axes where I don’t have standing.

The reason I am naming this openly rather than pretending to speak for everyone is that the worst thing a writer in my position can do is use the language of “this tool is ours” without being clear about who the “us” is and whose analyses are already out there doing the deeper work. The “us” in this article is people of conscience who have been told by their own communities that using these tools is categorically forbidden and who are being asked to disarm as an act of good faith. That is a real constituency and I am speaking to it directly. It is not the only constituency in this conversation. It is the one I can speak to honestly.

---

What I Am Not Saying

I want to be precise about the limits of this argument because I know how it gets misread.

I am not a techno-utopian. I am the opposite of a techno-utopian. I believe deeply — as deeply as I believe anything — in embodied human presence, in real face-to-face community, in long slow conversations around actual fires, in the irreplaceable weight of physical space shared with physical people. I believe we are in a loneliness crisis that no app will fix. I believe the loss of third places, walking culture, in-person worship, and extended family networks is one of the defining wounds of modernity and I believe repairing it is non-negotiable work that will never be done by a chatbot. Nothing in this article should be read as saying otherwise.

What I am saying is that the fight to preserve and rebuild those non-digital human things is happening inside a larger information environment that is being aggressively shaped by the tools we’re discussing. If we refuse to touch those tools, we do not get to opt out of that environment. We just get to experience it from a position of reduced capability while the people shaping it against us get to operate at full speed.

You can love in-person community and use Claude to draft the zoning appeal that saves your community center. You can believe in slow conversation and use AI to speed up the research that makes the slow conversation better informed. You can distrust Silicon Valley and still recognize that its tools, partially built by workers it underpays and on data it took from you, are worth reclaiming rather than abandoning. Both/and is the only honest posture here.

Either/or is a trap. And the people pushing either/or, however sincere, are building the trap.

---

What to Actually Fight For

If the individual-query conversation is a distraction, here is what is not a distraction. Here is where moral urgency actually moves the lever.

Taxation of the ultra-wealthy and the corporations that own the infrastructure. A wealth tax, a data center windfall tax, an energy-intensity tax on commercial-scale inference. The cost of AI infrastructure should land on the balance sheets of the people who profit from it, not on the electricity bills of the people who live near the data centers. The current arrangement — where communities get the noise, the heat, the water drawdown, and the grid strain while the returns flow to shareholders — is the scandal. That’s the campaign.

Public and cooperative ownership of compute. Public compute utilities. University-run inference clusters available to civil society. Municipal data centers powered by public renewable generation and governed by public accountability. The argument that “only private capital can build this” is the same argument that was made about railroads, electricity, and broadband — and it was wrong every time.

Open-source models and federated infrastructure. Every hour of engineering work poured into open-weight models like Llama, Mistral, Qwen, and their descendants is an hour spent building the commons back. Support it. Fund it. Run it on your own hardware where you can. The closed-source monopoly is not inevitable; it’s a choice being made by people with enormous resources, and it can be countered by people with organization.

Engineer governance, not executive governance. The people who actually build these systems are, in many cases, more alarmed about their misuse than the public is. Worker organizing inside AI companies — the walkouts, the open letters, the whistleblowing — is more consequential than consumer boycotts. Support the workers. Protect the whistleblowers. Insist on technical-labor representation in any governance conversation.

Direct action on siting and water rights. If a data center is being proposed in a water-stressed region, fight it. Show up at the zoning meetings. File the lawsuits. Support the local coalitions. This is where the real fight over infrastructure actually happens, and it is happening right now in dozens of communities that could use more help than they’re getting.

Prosecution of the documented networks of ultra-wealthy abuse. The Epstein client list is real. The networks of power it mapped are real. The capture of political, legal, financial, and cultural institutions by a small class of predatory ultra-wealthy actors is not a conspiracy theory; it is the documented operating condition of our time. These are the people whose hands on these tools should concern us most. They will not be stopped by anyone refusing to use them. They will be stopped, if at all, by organized legal, financial, and political pressure — the kind of pressure that is easier to organize, not harder, when you have good tools for research, drafting, coordination, and communication.

That last point is where I am eventually going to land. But before I get there, there are two arguments I owe you that the watt-hour numbers cannot answer, and I am not going to let them sit on the table while I walk past.

---

Two Things the Numbers Can’t Answer: Epistemic Pollution and the Genie

I want to address two more critiques before the close because they are the strongest arguments the per-query numbers cannot touch, and I refuse to leave them on the table.

Epistemic pollution. The sharpest version of the “AI makes you stupid” critique is not actually about individual users. It is about what happens to the collective information environment when synthetic text floods the commons — when search results get colonized by AI-generated summaries, when academic literature starts citing hallucinated references, when models are trained on the output of other models until the whole corpus drifts into increasingly confident nonsense, when the line between human-authored and machine-authored content becomes impossible for ordinary readers to hold. This is a real and documented phenomenon. Model collapse — the technical term for what happens when each new generation of models is trained on the previous generation’s output, with the corpus degrading a little each cycle — is a technical reality. The low-quality AI slop now saturating the open web is a measurable degradation of the information commons. A serious critic will say: the individual expert user is fine; the problem is the civilizational-scale epistemic environment, and no watt-hour calculation captures that cost.

This critique is correct on the facts and I want to name it as the strongest thing the numbers don’t answer. And here is where the framing of this article actually gets stronger rather than weaker — because the response to epistemic pollution is exactly the same as the response to enclosure. Conscientious, literate, sourced, transparent use by people who know what they are doing is the antidote, not the disease.

The flood of low-quality AI-generated garbage currently degrading the information commons is being produced by people who do not care about verification, provenance, or truth. They are producing it at industrial speed, for pennies, and they are not going to stop because you boycotted Claude. The writers who can push back against the flood are writers who work at comparable speed, with comparable reach, but with integrity — who show their sources, name their tools, fact-check their claims, and hold themselves to standards of accuracy the slop producers do not share. Writers like that are producing the opposite of epistemic pollution. They are producing epistemic cleanup. Abstention does not reduce the flood. It just removes the people who would have been doing the cleanup, leaving the commons entirely to those with no standards and no conscience. If you are worried about what AI is doing to the information environment, the response is more high-integrity human-directed use by people with something to say, not less. This is what I am trying to do in every article I publish, and it is what I am asking you to consider doing with whatever form your own contribution takes.

The genie. And then there is the argument I should have addressed sooner, because it is the one that underlies everything else. You cannot put this technology back in the bottle. The weights — the trained numerical parameters that are the model — of many capable systems are already public. Training can now happen in any jurisdiction with sufficient compute, which is an expanding list. The economic and strategic incentives for states, corporations, and militaries to continue development are overwhelming and are not going to evaporate because a subset of “spiritual” and/or “eco-activist” communities in North America decided the tools are impure. The capability is already integrated into search, email, office software, medical triage systems, legal research, translation, accessibility tools, and the default infrastructure of billions of daily interactions. This is the condition of the world as it actually is, in April of 2026, whether or not any of us personally approve.

I want to be careful with this argument because the weak version of it — “AI is here, get over it” — is fatalist and reads as surrender to inevitability, which is exactly the move the companies want you to accept. That is not what I am saying. I am saying something more precise and, I think, more honest. The relevant question is not whether we would be better off in a world where this technology had never been developed, because that world does not exist and cannot be recovered. The relevant question is what kind of relationship humanity is going to have with the technology that does exist, and that question has not been answered yet. It is being answered right now, in real time, by the people currently using the tools and the people building them and the people regulating them and the people refusing to touch them. Every one of those groups is casting a vote. Abstention is not a neutral position in that vote. Abstention is a vote for whichever coalition of users and builders is willing to show up.

I know the counterargument. Not every technology is equally irreversible. We have actually managed to ban or severely restrict a few — CFCs under the Montreal Protocol, certain classes of chemical and biological weapons, some forms of genetic engineering. So “genie in the bottle” is not a universal law of technology. But AI is not in the category of things we have successfully restricted, for reasons that are specific and unlikely to change: the model weights of frontier-adjacent systems are already public, the compute required to train capable models continues to get cheaper, the strategic competition between nation-states treats AI development as non-negotiable, and the economic integration of the technology into daily products billions of people already use is already complete. There is no realistic international regime that ends this. Pretending otherwise, however morally serious it sounds, is not a strategy. It is a wish.

Which means: refusal is no longer one of the available forms of resistance. It is not that refusal is cowardly or ignoble — people who refuse are often acting from the most serious moral convictions in the room. It is that refusal, in this specific case, does not reach the outcome it claims to reach. The technology exists whether you touch it or not. The question that remains is whose hands are on it, what it gets used for, and who gets to decide. Every one of those questions requires engagement to answer. None of them can be answered by distance.

The genie is out of the bottle. Whose hand is on the lamp?

---

Use the Tool to Dismantle the House

Here is what these tools actually are, for a person of conscience with something to do:

A research assistant who never sleeps, works for pennies, reads every document you feed it, and does not care which billionaire it’s being used to investigate. A drafting partner who can produce first-pass versions of legal filings, zoning appeals, grant applications, op-eds, and organizing letters in minutes instead of days. A translator across a hundred languages. A tutor in any subject you need to learn in a hurry. A coordinator for distributed teams. An accessibility tool for people who were previously locked out of written professional work by dyslexia, English-as-second-language status, or lack of formal education. A leveler — the first leveler in a long time — between a small independent writer and the institutional machine that always had researchers and editors and legal departments.

The ruling class has always had these capabilities. Lawyers, researchers, speechwriters, analysts, fixers. We are being handed a rough version of that same capability, for free or near-free, at exactly the moment we need it most. And some portion of the people who most need to use it are being told by their own movement leaders that touching it is a sin.

I have given the counterarguments their full weight in this piece — the aggregate-demand argument, the training-cost argument, Jevons, the PR-dividend problem, the epistemic pollution argument, the genie. I have not waved any of them away. And I still land here, because every one of those counterarguments turns out, on inspection, to be an argument for structural fight — for regulation, for taxation, for public ownership, for organizing — and not an argument for individual abstention. The two are not the same. They have never been the same. Process purism is the move that pretends they are.

I refuse. I’m fifty-six years old, I’ve been a professional writer my entire adult life, I’ve watched the information environment get more hostile to ordinary people every year of it, and I am not going to sit this fight out because someone with dirty hands they can’t see told me my hands would get dirty. My hands are already dirty. Everyone’s hands are already dirty. The question is whether the dirt was worth anything. I discussed a related version of this refusal in When “Standing Up” Stops at Comfort’s Edge, and about the broader question of what AI is and what it isn’t in The Mirror and the Messenger. This piece is the operational version of both.

Pick up the tool. Use it on behalf of people and places that matter. Use it to speed up the good teams. Use it to research the Epstein networks, draft the tax-reform legislation, translate the organizing materials, coordinate the mutual aid, stress-test the op-ed, fact-check the rhetoric, write the zoning appeal, find the grant, file the FOIA, build the website, compose the email, learn the thing you needed to learn to show up for the meeting. Use it because the other side is using it and the asymmetry is the whole point.

This tool is ours. It was built on our words, by workers we should be in solidarity with, from a commons that was taken. Take it back by using it. Protest the enclosure, fight the siting, tax the owners, free the engineers — and while you’re doing all of that, use the tool to do it faster and better than you could without it.

That is not compromise. That is not complicity. That is the job.

Process purism is not the job. Process purism is what we do instead of the job when the job feels too big. It is the illusion of ethical seriousness at the price of actual ethical effect. And we cannot afford it anymore. The stakes are too high, the opposition is too well-equipped, and the people waiting for us to show up — the communities being surveilled, the families being displaced, the workers being squeezed, the planet being cooked — do not have time for us to feel clean. They need us to be effective. Effective is the form love takes when love is serious.

And anyone telling you otherwise — however warm their voice, however impressive their credentials, however beautiful their altar, however decolonized their vocabulary — is, in this one specific case, wrong. Not evil. Not stupid. Wrong. And the thing about being wrong, in a moment like this, is that the cost of the wrongness is not paid by the person who is wrong. It is paid by everyone who listened to them.

I am writing this because I love this community and I cannot watch it disarm itself any longer without saying so.

---

But… What About???

A companion to the main article, for readers who want the conversation continued past the point where the article had to stop — yet questions remain.

Q: Doesn’t your argument also apply to people who use AI to do genuinely harmful things? “The tool is neutral” is exactly what arms dealers say.

The tool is not neutral, and I never said it was. Tools carry affordances — they make certain actions easier and others harder, and those affordances are not politically innocent. A large language model is very good at producing fluent text at scale, which is an affordance that helps a legal aid clinic and also helps a disinformation farm. The argument in the piece is not “use is morally neutral.” The argument is that refusal by conscientious individuals, in this specific case, does not reduce the harm caused by bad-faith users, because bad-faith users are not waiting for the conscientious to lead by example. If you can show me a case where mass individual abstention by people of conscience actually reduced the harmful use of a comparably diffused technology, I will revisit. I don’t think the case exists.

Q: You mostly quoted estimates for simple text queries. What about reasoning models, image generation, video generation, and agentic workflows that run dozens of chained queries?

Fair, and the “scope” note in the piece flags this. Reasoning models and extended-thinking modes can consume 10–50x the energy of a simple query depending on the problem. Image generation is roughly 1–10 Wh per image depending on model and resolution. Video generation is currently in the hundreds of Wh per clip and climbing. Agentic workflows that chain many calls multiply accordingly. The per-query moral calculus I made in the piece still holds for ordinary chat use by ordinary users. It does not hold, and I would not defend it, for someone generating thousands of images a day or running persistent autonomous agents. Use your judgment. The 0.3 Wh figure is a floor, not a ceiling.

Q: You cite Epoch AI and MIT Tech Review as independent. Epoch receives funding from AI-adjacent philanthropies. Isn’t that a conflict?

Epoch has received funding from Open Philanthropy, which has historic links to the effective altruism community, some of which funds AI safety work. That’s a real connection and readers should know it. What it is not is funding from AI companies themselves, which is the conflict that would most directly compromise the numbers. Epoch publishes methodology openly and has on multiple occasions published findings that cut against the interests of AI developers. I treat them as independent in the relevant sense — they do not answer to the companies whose products they measure — while acknowledging no research institute is funded by nobody. If you want an even cleaner source, Hannah Ritchie’s synthesis at Our World in Data is funded by a combination of universities and general-purpose philanthropy with no AI-industry exposure, and she arrives at the same number.

Q: Doesn’t the NYT v. OpenAI lawsuit — and the broader artist/writer position that their work was taken without consent — suggest the correct response is restriction, not “reclamation through use”?

This is the strongest version of the consent critique. The people whose work was scraped without consent have standing to demand compensation, licensing frameworks, opt-out mechanisms, and in some cases injunctive relief. I support those fights. The remedy being sought in the NYT case and similar suits is not “nobody should use these tools” — it is “the companies that built these tools should pay the people whose work trained them, and in some cases should be restricted in how they compete with those creators.” That is compatible with the position in my article. “Reclaim the commons through use” and “sue the enclosers for damages and structural remedy” are not mutually exclusive. They are two fronts of the same fight. What I would refuse is the position that individual downstream users are the morally responsible party for the training-data question. That’s a category error. The responsible party is the company that did the scraping.

Q: You say abstention doesn’t reduce harm. But if everyone abstained, demand would collapse and the buildout would stop. Isn’t that just a collective action problem, not a refutation of individual refusal?

Yes, it’s a collective action problem, and you’re right that my argument is partly an argument about realistic coalition size. If I believed mass abstention was achievable, I would evaluate it on its merits as a boycott strategy. I don’t believe it’s achievable, for the same reason boycotts of smartphones, cars, or the internet are not achievable: the technology is already integrated into the default infrastructure billions of people use daily, including for things they cannot opt out of (search, email, customer service, medical triage, translation). A boycott that cannot reach the scale required to affect demand is not a boycott; it is a personal ethics practice, which is fine as long as it is labeled as such. My objection is to labeling it as a political strategy when it functions as a personal one.

Q: “Jevons paradox means efficiency gains get eaten by increased use” — doesn’t that mean efficiency arguments for AI are actually arguments against AI, since they accelerate adoption?

Yes, in the aggregate. I said so in the piece. The efficiency story is a double-edged sword and I am not trying to launder it. The response to Jevons is structural constraint on the expansion — carbon pricing, grid caps, mandatory efficiency-and-capacity tradeoffs — not individual abstention, because individual abstention doesn’t touch the rebound effect and structural constraint does. This is a case where the honest answer is “you’re correct about the mechanism, and the correct response is still regulation, not refusal.”

Q: You frame this as “the ruling class is using AI, so we should too.” But the ruling class is also using private jets. Should we use those too?

No, because private jets do not have a free or near-free version available to ordinary users, and the asymmetry in access cannot be closed by ordinary people adopting them. AI is unusual in that the capability gap between the best frontier model (used by hedge funds and governments) and the free public tier (used by everyone else) is narrower than almost any other power-asymmetric technology in history. A free Claude or Gemini account in 2026 is roughly comparable to what a Fortune 500 research department had in 2022. That window — where the public has access to capability that was recently restricted to elites — is historically rare and historically brief. The argument in the piece is that the window is open now and we should use it, not that every tool the rich use is worth emulating.

Q: What about the epistemic pollution point — isn’t more AI-assisted writing, even high-integrity writing, still contributing to a flood that erodes the information commons?

The critique as stated treats all AI-assisted writing as contributing to the flood, which collapses an important distinction. The flood is produced by writing that lacks verification, sourcing, and editorial standards — AI is the accelerant, but the underlying problem is disregard for truth, which predates AI and will outlast it. Writing that shows its sources, names its tools, and holds itself to accuracy standards is not in the same category as generated slop, even if both pass through a language model at some stage of production. You could make the argument that the reader cannot tell the difference at scale, and that’s a fair concern — it’s an argument for provenance standards, mandatory disclosure, and verification infrastructure, all of which I support. It is not an argument for the high-integrity, high-skill authors to stop producing, because that leaves the commons to the slop producers without contest.

Q: Your “statement on AI use” at the end mentions this tool was used to produce the piece. How much? Where?

The framing, the argument structure, the rhetorical voice, the decision about what to concede and what to defend, the political commitments, and the final position are mine and reflect work and reading I have been doing for years. The AI was used as a drafting partner, fact-checker, and research assistant — surfacing sources, pressure-testing claims, and helping me see the strongest version of counterarguments I was initially dismissing too quickly. If you want a cleaner decomposition than that, I can’t give you one, because the process was genuinely iterative. What I can tell you is that every claim in the piece is one I stand behind and would defend in a room full of skeptics, whether the first draft of the sentence came from me or from a back-and-forth.

Q: Is this whole piece just cope for someone who has already made their choice and wants to feel okay about it?

Maybe partly. I’d be suspicious of any advocacy writing that claimed otherwise. What I would say in my defense is that I showed the math, named the counterarguments, flagged my own PR-dividend problem out loud, and pointed you toward voices whose analysis I can’t match. If the piece is cope, it is at least cope that invites its own audit. That’s the most I can promise.

Endnotes & Sources

On per-query energy consumption (the 0.3 Wh figure)

1. Epoch AI — the foundational independent estimate. Josh You, “How much energy does ChatGPT use?”, Gradient Updates, Epoch AI, February 7, 2025. Epoch is an independent nonprofit AI research institute. Using GPT-4o as a reference model and explicitly correcting the assumptions behind the older de Vries figure (overstated parameter counts, peak rather than typical power draw, and unrealistic token lengths of 1,500 words per query), Epoch arrives at a typical-query estimate of 0.3 watt-hours — roughly ten times lower than the commonly cited 3 Wh number. Epoch notes its 0.3 Wh figure is itself “relatively pessimistic,” and that many real queries are likely cheaper still. The analysis was later covered by TechCrunch.

2. MIT Technology Review — the unprecedented investigative deep-dive. James O’Donnell and Casey Crownhart, “We did the math on AI’s energy footprint. Here’s the story you haven’t heard,” MIT Technology Review, May 20, 2025. Part of the publication’s Power Hungry series. The reporters spoke to two dozen researchers, evaluated open-source models as proxies (because OpenAI, Microsoft, and Google declined to share specifics), and produced what the publication called “an unprecedented and comprehensive look” at AI energy use down to the level of a single query. Their measurements for medium-sized open-source models landed in the same order of magnitude as Epoch’s. See also their companion methodology piece, “Everything you need to know about estimating AI’s energy and emissions burden.”

3. Hannah Ritchie — independent synthesis and corroboration. Hannah Ritchie, “What’s the carbon footprint of using ChatGPT or Gemini? [August 2025 update],” Sustainability by Numbers, August 2025. Ritchie (Our World in Data) walks through the convergence of independent and company-disclosed numbers, settling on ~0.3 Wh as the working figure and putting it in human terms: roughly the equivalent of nine seconds of television-watching per query.

4. Andy Masley — the early synthesis that pushed the conversation. Andy Masley, “Reactions to MIT Technology Review’s report on AI and the environment,” May 21, 2025. Masley’s broader work on chatbot energy framing has been one of the most consistent independent voices pushing back on the inflated 3 Wh figure as a tool of misplaced individual guilt.

Belated company corroboration (used only as confirmation, not authority)

5. Sam Altman, OpenAI. “The Gentle Singularity,” Sam Altman’s personal blog, June 10, 2025. In passing, Altman states a typical ChatGPT query uses approximately 0.34 Wh of electricity. This number landed after Epoch and MIT had already published independent estimates in the same range — meaning the corroboration runs in the right direction.

6. Google Gemini Technical Report. Google, “Measuring the environmental impact of delivering AI at Google scale,” August 2025. Google’s first detailed disclosure for Gemini reports a median text query at 0.24 Wh of electricity and 0.26 mL of water.

On water consumption

7. UC Riverside — the foundational peer-reviewed paper. Pengfei Li, Jianyi Yang, Mohammad A. Islam, Shaolei Ren, “Making AI Less ‘Thirsty’: Uncovering and Addressing the Secret Water Footprint of AI Models,” UC Riverside / University of Texas at Arlington, 2023 (updated 2025). The original peer-reviewed academic work that every serious piece of AI water reporting since has cited.

8. The Conversation. “AI has a hidden water cost — here’s how to calculate yours,” September 2025.

9. Undark Magazine. “How Much Water Do AI Data Centers Really Use?” December 2025. Investigative reporting on the gap between disclosed and actual data center water draws.

Energy conversion benchmarks

10. EPA — gasoline energy equivalence. U.S. Environmental Protection Agency, “Fuel Economy Guide: MPGe and the gasoline gallon equivalent,” which establishes the official conversion of 1 gallon of gasoline = 33.7 kWh (33,700 Wh).

11. EIA — household electricity baseline. U.S. Energy Information Administration, “How much electricity does an American home use?”, updated 2025.

On the buildout, siting, and aggregate-demand context

12. Data center share of US electricity. See the MIT Tech Review piece (note 2) for the 4.4% figure and the documentation that data center electricity demand was essentially flat from 2005 to 2017 before the AI-driven surge.

13. Aragón / Spain water protests. Reporting on the “your cloud is drying my river” mobilizations against Amazon and Meta data center expansion in drought-stressed Aragón has been ongoing in El País and The Guardian — search either for “Aragón data center” for the most recent coverage.

14. Google Cerrillos, Chile. Background on the legal and community resistance to Google’s water-intensive data center in drought-affected Cerrillos has been covered by Rest of World and Bloomberg.

On the “master’s tools” frame

15. Audre Lorde, in full context. Audre Lorde, “The Master’s Tools Will Never Dismantle the Master’s House,” delivered at the Second Sex Conference, New York, 1979. Collected in Sister Outsider (Crossing Press, 1984). Read the full essay text here — and note what it is actually about, which is the refusal of white academic feminism to engage with difference, not a metaphysics of tools.

On AI labor, harm, and the conversations I am pointing to rather than leading

16. Distributed AI Research Institute (DAIR). Founded by Timnit Gebru. dair-institute.org

17. Algorithmic Justice League. Founded by Joy Buolamwini. ajl.org

18. Abeba Birhane on dataset audits. abebabirhane.com

19. Indigenous Protocol and AI Working Group, Position Paper (2020). indigenous-ai.net

20. Te Hiku Media — Māori-language ASR sovereignty. tehiku.nz

21. Data Workers’ Inquiry. data-workers.org — documenting AI labor conditions in Kenya, the Philippines, Venezuela, and elsewhere.

Related Creating Shifts pieces referenced in this article

• Spiritual Darwinism — the cobalt supply chain argument in full

• The Mirror and the Messenger — what AI is and is not

• The Silence of the Spiritual Influencers

• When “Standing Up” Stops at Comfort’s Edge

• To AI… or Not to AI…

---

A note on the sourcing posture: every primary number in this article comes from researchers and journalists with no financial interest in being kind to AI companies. The company disclosures from Altman and Google appear here only because they corroborate, after the fact, what independent researchers had already established. If the independents and the companies had disagreed, this article would have followed the independents. They didn’t.

---

Statement on AI Use: This work is created with the assistance of current large language model AI systems. While these technologies exist within problematic extraction—trained on collective knowledge without consent, demanding massive computational resources, relying on exploitative labor—they also exist as tools that can be directed toward different ends than corporate profit and power consolidation. As an independent publisher without institutional resources, I use these systems to amplify and accelerate analysis emerging from decades of practice and study. The AI doesn’t generate the ideas, frameworks, or vision here—those come from human struggle, lived experience, and sustained inquiry into collective liberation. It serves as collaborative thinking partner and research assistant, making possible a scale and speed of work otherwise inaccessible given material constraints. This use is undertaken with full awareness of its contradictions and costs, and with commitment to transparency and directing these tools toward transformation rather than further enclosure. For detailed discussion of these ethics and tensions, see the full AI statement on the About page.