The tool is good. The leak is not the plan.

The Pattern

There are two loud positions on AI, and they agree on the only thing that matters. Both of them tell you to do nothing.

The first says the tool is evil and you should want nothing to do with it. Walk away on principle. The second says the tool is fine, stop worrying, and get on with it. Do not look too closely. These feel like opposites. They are not. One says do not use it. The other says do not question it. And both of them, said together, add up to the same instruction: do not get together, do not ask who decides, and leave the room to the people who own the place.

I am not taking either one. I built this site with AI. I use it for editorial work, for code, for research. It is the most useful tool I have met in my adult life. And I am telling you, plainly, that the tool is good. I am also telling you that the tool being good is exactly why it has to be ethical, and why “it works” is not an answer to “who does it work for.” Those are not two different arguments. They are one argument, and you only get to keep the first half if you fight for the second.

The tool is real

Not what the advertisements say. What is happening today, in the hands of regular people.

A researcher in Lagos is using AlphaFold to model protein structures for diseases that kill millions and receive almost no funding, because the patients cannot pay. DeepMind mapped 214 million protein structures, essentially every protein known to science. Before this, modeling a single one could take years. Now it takes minutes. Someone pointed the tool at a problem that money had ignored for decades, and the tool worked. For the people nobody was building for.

In this country, 80% of low-income people facing civil legal problems get no representation. They go to housing court alone. The Pro Bono Institute surveyed legal aid organizations in early 2026: 74% now use AI in daily work, nearly double the broader legal industry. Not corporate lawyers. Legal aid. There was never a lawyer. The machine is filling a hole that existed because justice here is something you buy.

A blind person puts on Envision Glasses and the world gets described to them in real time. For a deaf person in a rural area, the alternative to AI sign-language translation was not a human interpreter. The alternative was silence. The machine ended the silence.

Someone I know used it last month to read a letter from her landlord, written in a language she does not speak. A translation service would have cost a day’s wages. The phone did it in four seconds. She read the letter, understood what was being asked of her, and made dinner. That is the whole story. A woman read a letter and made dinner, and the machine made that possible, and nobody had to ask permission.

None of this is hypothetical. And it all shares one feature: in every case, the tool is serving people it was not built for. The owners build it for themselves. It leaks. And the leak is where the good comes from.

But I will not pretend the leak is a plan. The leak is accidental. It depends on the tool getting cheap, on the owners being slow, on a gap they can close whenever they decide to. Regulation can close it. Pricing can close it. The leak buys time. It does not decide how the time gets used. That is the fight. And the fight is the whole point.

What ethical AI is not

Let me say what it is not first, because this is where most of the conversation goes to die.

Ethical AI is not a slide deck. It is not a set of principles a compliance officer reads to a room that nods and then ships the thing the principles were about. It is not a fourteen-page policy document that three people signed and that can be revised on a Tuesday. Inside the companies that own this infrastructure, the gap between what a model can do and what it is allowed to do is a policy document. Not a law. Not a vote. Not a union. A PDF. “Legal says we can” is not governance. It is the absence of governance wearing a lanyard.

If your ethics has no enforcement, it is marketing. If the people whose labor built the tool have no say in how it is used, your principles are decoration. Ethical AI is not a feeling the company has. It is a set of mechanisms with teeth.

What ethical AI is

Four things. Consent, payment, power, and rules you can enforce.

Consent means the people whose work trained the tool agreed to be in it. Payment means they were paid. Power means the workers who build and run these systems have a say in what gets built. Rules means there is a law, with a penalty, that a company can break and be punished for breaking. None of this is radical. All of it is already happening somewhere. The question is whether it scales, and that depends on whether regular people show up for it or walk away and call it a principle.

It is already working

This is the part people skip, because it is slower and less satisfying than declaring the whole thing corrupt. But ethical AI is not a future you wait for. It is a set of fights that are being won right now.

Actors forced the issue. SAG-AFTRA, the union representing performers, ratified agreements that require consent and compensation before a studio can use a digital replica of a member. Not a request. Not a guideline. A contract, with the weight of a strike behind it. The people whose faces and voices feed these models got to say no, and got to be paid when they said yes. That is consent and payment, won by getting together and refusing to be ignored.

New York City made the hiring machines show their work. Local Law 144 requires an independent bias audit before an automated tool can be used to screen job candidates, and it requires the results to be posted. Use the tool without the audit and you can be fined. That is not a principle. That is a rule with a penalty, and it exists because people demanded that a machine deciding who gets a job should have to answer for what it decides.

The European Union built a whole framework. The AI Act sorts systems by risk, bans the worst uses outright, and sets fines that can reach tens of millions of euros or a percentage of global turnover. It is imperfect and it is slow. It is also the difference between “legal says we can” and “the law says you cannot.” Governance with a number attached to it.

Some companies chose a different path on training data. Adobe built its Firefly models on licensed content and public-domain work, rather than scraping the open internet. It is not perfect, and the people whose work is in those libraries have their own complaints about how that happened. But the choice itself matters. It proves that “we had to take it without asking” was always a decision, not a law of nature. Consent was available. Some people took it.

And the fight is organized. More than seventy unions, representing around 140 million workers, have formed a global coalition demanding accountability for AI-driven labor displacement. That is not a feeling. That is a constituency. That is the slow, boring, unapplauded work of making invisible labor visible to the people who can change it, and it is the only thing the owners have ever actually had to answer to.

The costs are real. So what.

The people labeling data in Kenya are paid below poverty wages to train models that will never credit them. The content moderators who review the worst material on the internet develop PTSD and receive wellness pamphlets. The artists whose work trained these models were never asked and will never be paid, because the legal framework has not caught up and the companies prefer it that way. A large data center drinks millions of cubic meters of water a year. Fewer than ten companies own the whole infrastructure. I know. I have written about all of it. None of it is acceptable.

And none of it is an argument against the tool. It is an argument against who owns the tool, and against the absence of rules forcing them to share it. Saying “AI cannot be ethical” is like saying “steel cannot be fair” because the prison is made of steel. The steel is not the problem. The prison is the problem. And you do not fight the prison by refusing to touch steel. You fight it by getting together with the other people who pour the concrete and demanding a say in what gets built.

What I do with it

I draft first, always. The thought has to be mine before the machine touches it. I write the ugly version, the version that is wrong in interesting ways, and then I ask where it sags. That keeps me the author. The other way, where you ask the machine first and edit second, is a slow surrender you do not notice because each step feels efficient.

I use it to buy time for people. Translate the document. Summarize the report. Handle the repetitive work so the hours go to the thing that requires a human in the room. The tool frees the time. The time goes to people. If it goes to more screen, you have made your life smaller and called it progress.

And I treat this as a labor issue, every time someone tries to make it a technology issue. Because the technology is not the problem. The technology is the excuse. The problem is that ten companies get to decide how the most powerful tool in a generation is used, and the rest of us get to read the terms of service. That is not a technology problem. That is a power problem. And power problems have answers. The answers involve consent, payment, power, and rules with teeth. They involve getting together, demanding better, writing laws, and showing up on the Tuesday night nobody applauds.

The person who says “AI is evil” and the person who says “AI is fine, stop worrying” are both helping the same ten companies. One says do not use it. The other says do not question it. Both say do not get together.

The tool is good. That is not the end of the argument. It is the reason the argument matters. The only position that actually threatens the owners is the one that says: we built this too. Our labor is in it. Our data is in it. It is too useful to refuse and too important to leave to you. And we are staying, and we want a say.