Ethical Technology — A Pipe Dream

Technology is not ethical. It just is.

A gun is not good or bad. It’s a gun. It does what guns do — puts a small piece of metal through whatever you point it at. The ethics live in the person holding it. In the person’s choice to pull the trigger, in the reason for pulling it, in the willigness to accept consequences. Not in the gun.

We seem to have forgotten this. Or maybe we’ve decided it’s inconvenient.

A social network post I saw recently walked through three “models” of ethical AI: moral alignment, safety as predictability, and auditability. The post was honest about the limits of each. But it took for granted that ethics is a property a tool can have.

It isn’t, in my opinion.

For me, this is precisely the area in social discourse about AI where lines blur, and where the particulars of taking the three in turn are instructive.

Moral alignment with a constitution, the Anthropic approach, is a curated set of rules — written by humans, encoded into a system that mostly behaves the way the rules say. When it doesn’t, more rules get written. That’s engineering. It’s good engineering; I use Claude every day, and the constitutional approach is part of why I trust it for the work I do. But none of that is ethics. It’s a tool behaving the way its builders shaped it to behave. The ethics live in the builders. And in me, when I decide what to point it at.

Safety as predictability, the Cohere approach, is even more clearly a tool property. Stable outputs, traceable data, low hallucination rates — these are quality attributes. A drill press has them. We don’t call drill presses ethical. We call them well-made.

Auditability and explainability — the third — is where the conversation gets confused in a way that’s actually dangerous. An “explanation” from a language model is a statistical reconstruction of what was most frequently said about this kind of thing, by the writers of whatever was in the training set. It is not a reason. It is not a judgment. It’s a pattern compressed from the past, retrieved in the present, and dressed up to look like reflection. If your training corpus carries an abundance of unethical behavior, your “explanations” will reflect that — confidently, plausibly, in well-formed sentences. The model is not lying. There is nothing for it to lie about. It’s showing you what was most frequent in what was used in training. You can counter that training with embeddings and fine tuning, and with your own in-session guardrails. You can invest in countering training data skew. If you choose.

Auditability is not in the space of ethics and judgment. It’s in the space of statistics. You can audit a model all you want, and what you will see is the residue of its corpus. Ethical or not. That’s all there is to see.

So if you actually want ethical responses from these systems, the work isn’t bolting an “ethics layer” on top of the model. The work is the corpus. What did you feed it? What did you leave out? What did you label as right and wrong, and how clearly did you exemplify the distinction? Curate the inputs, and you can shift the statistical center of gravity. That’s a real and difficult and worthwhile thing to do. It is also still not ethics in the model. It’s curation by humans. It produces a tool whose default outputs lean in a direction the curators chose. The choice was ethical. The tool is still a tool.

And even with a beautifully curated model, the human still needs principled judgment. Context guides — a sentence that’s helpful in one situation may be harmful in another, and no amount of training data settles that for you. You have to look at what’s in front of you and decide. That isn’t the same as saying anything goes. Situational ethics, where the situation always wins, is just ethics with the spine taken out. The discipline is to hold principles and apply them in context. Principles without context are brittle. Context without principles is improvisation.

The deeper thing that should be said is that the people building and funding these systems have discretion at every step. Companies choose what to train on. They choose what to ship. They choose what to charge for and how to charge for it. They choose whether to lobby for or against oversight. Those are all ethical acts performed by people. Naming the resulting product “ethical AI” launders the ethics back onto the tool, where it conveniently can’t be held to account. And we users are not off the hook. We choose to buy what is sold. We choose to use ethically or not. We choose to allow tech to make choices for us. Or not.

This is rhetorically convenient and practically dangerous. Convenient, because if the tool is ethical, no individual has to be. Dangerous, because tools never were and never will be the locus of moral agency, and pretending otherwise just makes the agents harder to see.

People write about AI built to simulate emotion — digital friends, therapy bots, “empathetic” call-centre agents. “Emotional manipulation sold as innovation” is in some ways an apt and right framing of what we’re seeing. But those products aren’t unethical because the AI is unethical. They’re unethical because someone built them, knowing what they were. Someone funded them, knowing what they were. Someone is selling them, knowing what they are. And someone is complicit in buying them. The ethics question goes all the way back through the supply chain to humans making choices. The technology is innocent in the way a hammer is innocent. The hand on the hammer is not.

So when I hear the phrase “ethical AI,” I hear something a little different than what’s said. I hear: we would like to be considered ethical without being accountable, as individuals, for the choices we made. It’s a useful phrase if you’re a vendor. It’s a confusing phrase if you’re trying to actually think about this.

There is no ethical technology. There is only ethical use of it, and ethical creation of it, and ethical funding of it, and ethical regulation of it — each of those an act by a person, recoverable to a name, answerable to consequences. The tool sits there, doing what tools do. We are the ones with discretion. We have always been the ones with discretion.

Tech is tech. It just is. People choose how to view and use it. The choice determines the effectiveness of use and outcomes.