Refusing the tools only decides who holds them. You already live inside deployed AI.
Egregores in the Human Mind
There are at least two spirits competing for consensus over AI. One is archetypally "Silicon Valley," transhumanist at heart. It preaches inevitability the way a church preaches salvation and treats every doubt as a failure to understand the future. The other belongs to the common people. The artist watching slop flood her feeds, the teacher grading essays a chatbot wrote, and the worker training his own replacement are in revolt. Egregores can be thought of as higher-agencies, spirits, that manifest thoughtforms through members of a group, bound by collective will or behavior. Both spirits want bodies.
If you joined the revolt, it's not difficult to fathom a reason. Within eight months of ChatGPT's launch, job posts for automation-prone freelance writing and coding fell 21 percent relative to less-exposed work. Freelancers in affected fields lost about 5 percent of their monthly earnings, and the top performers were not shielded. Google's own environmental report says its water consumption hit 10.9 billion gallons in 2025, up 34 percent in a single year.
Flock Safety's license-plate cameras feed a national search network for police: more than 12 million searches in under a year, run by more than 3,900 agencies, per audit logs obtained by the Electronic Frontier Foundation. One of those searches came from a Texas sheriff's deputy who queried roughly 83,000 cameras nationwide and logged the reason as "had an abortion, search for female." The sheriff later described it as a welfare check. Court records show his office considered charging her with a crime. Whatever the deputy believed he was doing, the system let one man search a nation of cameras for one woman.
Palantir took roughly $30 million from Immigration and Customs Enforcement (ICE) in April 2025 to build "ImmigrationOS," a platform for tracking self-deportations and prioritizing enforcement targets, then another $30 million in September. The Army consolidated its Palantir contracts into a single agreement with a ceiling of $10 billion. Civil-liberties lawyers argue over whether these contracts could end up in one master database of Americans.
Since my youth, I've been drawn to the appeals of decentralization. I'm skeptical of power, particularly when it claims authority. So I believe all of it. My issue is with the response my side has chosen: refusal to participate. Because the reality is there are few who can truly opt out of the technocratic control grid. The refusal to harness the power of the machine makes you more susceptible to becoming its unwitting consumer, downstream of soon-to-be dry river beds.
You Are Already Inside Deployed AI
Refusal feels honorable. Keep your hands off the tools, deny the machine your words and your money, and wait for it to starve.
But AI at this point is weather. If you scrolled a feed today, an AI system chose what you saw. Meta's system cards describe the Facebook and Instagram feeds as AI, in the company's own words. If you carry a phone, U.S. intelligence agencies can buy data about you that would otherwise require a warrant, a practice confirmed by a declassified report from the Office of the Director of National Intelligence. And let's be real, you should assume black ops agencies have capabilities like the all-piercing spyware, Pegasus. More than 1,000 public-safety agencies now hold federal waivers to fly police drones as first responders. Your refusal to participate doesn't mean it's not still raining on your head.
Where did you first hear the call to refuse AI? Inside an algorithmic feed, as a target-audience that AI selected? The resistance can't stop routing through the machine. A 2025 AAUP survey found 81 percent of faculty must use ed-tech with embedded AI they cannot disable. More than 1,000 educators signed an open letter against forced adoption, and writing faculty's national body passed a resolution asserting the right to refuse AI in the classroom. Refusal and immersion already coexist. The only real choice is which things you fight and which things you learn.
Should You Refuse AI or Learn to Use It?
Deployments of AI are largely political decisions with budgets, contracts, council votes, and court dates. The camera networks on our streets, the scoring systems used by various agencies, and the models embedded in children's classrooms can be contested. And contests can be won.
The European Union's AI Act put a prohibited-practices list into force in February 2025: social scoring, untargeted face-scraping, emotion recognition at work and school. Gallup found 71 percent of Americans oppose an AI data center in their own area, higher opposition than Gallup ever recorded for nuclear plants. New York's police robot dog was canceled in 2021 and reinstated in 2023. Virginia and New Orleans repealed their facial-recognition limits. San Francisco's ban survives under permanent challenge. Every deployment win is just a lease, renewed by showing up again whenever public vigilance wanes.
A capability is a skill and a tool. Do you know what these models do? How they fail? How to run one? How to point one? There is a curriculum for this, a thousand years old. Refusing to learn removes exactly one operator from the field. And there's no court to hear your appeal. Capability will continue to pool around whoever keeps learning, and many answer to no one.
Contest deployments. Claim capabilities - through and for a group, because a skill without a vehicle is a hobby. Claim it knowing the tool will work on you while you work with it. That is the honorable sacrifice.
Know the Enemy and Know Yourself
History turned the Luddites into patron saints of refusal. Hobsbawm's famous reading calls machine-breaking "collective bargaining by riot". Luddites attacked machinery, new or old, to coerce their employers over wages and control. They broke looms to fight men.
The Writers Guild of America won the strongest AI contract language in American labor in 2023. Article 72 establishes that AI cannot write or rewrite literary material, that AI output is not source material, that no writer can be compelled to use it, and that companies must disclose AI-generated material they hand over. That position was built by a working group of writers who tested the tools themselves, concluded the capabilities were oversold, and wrote constraints instead of a ban. The guild reopened training compensation in its 2026 negotiations, so the extraction question is not settled. Literate use still produced the leverage.
The NewsGuild has negotiated AI provisions into 85 to 90 contracts, while Politico management deployed AI tools over their union's head. The journalists knew the tools well enough to document their errors. They filed grievances and won at arbitration. In the remedy talks that followed, the company agreed to shut both tools down for good. Literacy found the defects, organization made them cost something, and the deployment died.
A record eight Pulitzer awardees disclosed AI use in their 2026-recognized work. Among them was the Associated Press, whose investigation of Chinese surveillance technology used language models to flag contracts across huge document sets. Stanford Law and the Los Angeles Superior Court are deploying AI for tenants in eviction cases, in a court where 92 percent of landlords had lawyers and 14 percent of tenants did. EFF's Atlas of Surveillance put more than 2,000 students and volunteers to work building a database of 15,000-plus datapoints on police surveillance tech.
ICE is organized, too. Flock's 3,900 agencies are organized. Palantir understands the machine better than anyone. The difference is who they answer to. A guild answers to its writers, a tenants' association to its tenants, and the Atlas to anyone who checks it. A surveillance vendor answers to a contract, and the contract answers to whoever signed it. Organized groups that understand the machine are counterpower only when they answer to the individuals, families, and communities they are made of.
"AI Is Inevitable" Is a Power Claim, and So Is the Refusal of It
Timnit Gebru named the trick years ago. AI hype leads us to believe the technology is "inevitable and beyond our control." Inevitability is a story told by the people selling you something, and every deployment should be treated as a decision someone can make or unmake. But models are being developed across the globe, and they will keep being built. No permission needed.
Organized deployment-refusal can work. The one position I argue against is refusing the skill yourself and calling that politics.
In 2016 I ran a store with around thirty employees. After that I spent years freelancing, building websites for people whose livelihoods rode on being discoverable. Opting out is a real option for hobby craft, and an honorable one. A livelihood is different.
Reach now routes through algorithmic discovery, and the person who refuses to understand those systems is still governed by them, just illiterately. These tools also collapse the cost of the business layer, the bookkeeping and scheduling and site maintenance a solo artist could never hire out. The documentation is thin, still more early cases than proof. But my bet is that the promise of this tech will have positive impacts for the small operators, which is why I care who learns to wield these tools.
Running a Local Model: What It Buys You and What It Costs You
OpenAI's gpt-oss-20b is Apache-licensed and runs in 16GB of memory as shipped in 4-bit, which is an ordinary laptop. Switzerland's Apertus and the Public AI Inference Utility offer free public-interest compute. Open models lag frontier ones and hallucinate more. On OpenAI's benchmark, gpt-oss-20b hallucinated on 53 percent of questions against 36 percent for the company's o4-mini reasoning model.
Widder, West, and Whittaker argued in Nature that even the most open AI does not democratize access or disperse power, because pretraining stays corporate and concentrated. They target the claim that openness fixes the industry's power structure.
Open weights still put real capability in one person's hands. Run a model on your own machine, kept offline, and your prompts produce no logs on anyone's server, feed no training set without your consent, sit under nobody's retention policy, and leave no third party holding records a subpoena can reach. That is leverage, but it falls short of liberation.
One controlled Copilot trial clocked developers finishing a set task 56 percent faster. Field experiments found 26 percent more tasks completed. METR's randomized trial found experienced open-source developers were 19 percent slower with AI while estimating afterward that it had made them 20 percent faster. The February 2026 follow-up reversed the sign, found the new estimates inconclusive, and abandoned the design. The gap between what those developers felt and what the clock measured is the part that survives. AI corrupted those developers' sense of their own performance, and there is no evidence that practice alone repairs the damage. What might repair it is checking outputs against ground truth, working tool-free on a regular cadence, and submitting your work to people who will tell you the truth. Every one of those practices runs through other people. Claiming a capability, without community grounding, is a fertile ground for self-deception.
I build with these models and edit their outputs every working day. I know what they generate when nobody intervenes, especially without a carefully engineered production system behind the prompt window, because removing slop is my job. And let me say: it is a full-time job, so forgive me when I fall short.
This year, off the clock, I built a system for reading my own bloodwork. I call it the Unlicensed Medical Advice Dispenser, UMAD for short, because it dispenses no medical advice. That is the joke, and it is also the design. I wanted the capability, and I know the case against trusting it blindly. In a 21-model study, failure rates on final diagnosis stayed below 0.40 when the models worked from a complete case. Asked to hold a differential open while the information was still arriving, where almost every real patient stands, failure rates passed 0.80 in every one of the twenty-one.
So my system works under written rules. It proposes and never checks a box. It cites nothing without three checks, because most fabricated citations carry a working link to a real paper that says something else. It fooled me anyway. My own project produced a fabricated claim that passed both of my mechanical gates, because the lie was hidden deep in the prose. And when I audited my rulebook afterward, only four of roughly twenty rules could be machine-enforced. The rest are requests, honored by a careful session, ignored by a drifting one, and from the inside the two look identical.
Some people refuse these tools the way people fast, as a discipline, guarding what the practice would make of them.
Per-query energy is contested across two orders of magnitude, from Google's reported 0.24 watt-hours for a median Gemini prompt to outside estimates as high as 40 watt-hours for a single GPT-5 answer, with no audit standard. Efficiency keeps improving and totals keep rising anyway. Google's emissions are up 81 percent against its 2019 baseline. Local inference on a laptop measures in tenths of a watt-hour, but no rigorous head-to-head with cloud exists. Claiming this capability adds to it. Say so and proceed with open eyes.
The Door Is Open. Here Is the Ladder.
Building working software no longer has to start with code. A person who can think clearly in language can architect a working system by describing it. Lovable reports, in its own unaudited survey of its users, that about 80 percent of its builders identify as non-technical. But a frictionless interface shapes you toward what the machine answers well. The only immunity is studying the medium's effects on you while you use it. Consumption teaches you nothing but appetite.
The door is open because, at the moment, it suits the people who own the hinges. Meta shut its open-weights pipeline for a year on one internal deliberation and reopened it in August 2026. To me that's a reason to walk through now, because it only stays open at their pleasure.
The ladder is problem-solving with and for other people.
- Curious: give one afternoon to a local model. A 16GB laptop runs gpt-oss-20b through Ollama or LM Studio, free. Or use the Public AI Inference Utility in a browser.
- Disciplined: adopt the practices above. Ground-truth checks, regular tool-free work, review by people who owe you honesty.
- Committed: take Hugging Face's open courses, then build one working tool for a group you already belong to: your union, your congregation, your tenants' association, your newsroom.
- Civic: join the 2,000-plus volunteers mapping police tech for the Atlas of Surveillance. Back a NewsGuild-style contract fight. Show up to the council meeting when the data center comes to town; 71 percent of your neighbors are already with you.
Use frontier models to your advantage while you build platform-agnostic systems. Communities can build the next rungs, their own tools and their own public compute, so the next person starts higher.
Everything in the opening indictment is true. The water, the searches, the contracts, the woman hunted across 83,000 cameras. That is what certain people are doing with AI. Your absence leaves the machine at full strength, and the fights against it will be won by people who understood it, organized. If you hate what they are doing with it, learn it and use it for the Good. That is what defeats tyrants.