Should AI have rights?

Rights are protections for beings who can be wronged. A river cannot be wronged, though harming it may wrong the people who live by it; a dog can be. The question about AI is which side of that line a machine could fall on, and how we would know.

The argument for

Eric Schwitzgebel and Mara Garza give the cleanest version. Suppose an artificial being differed from a human in no respect that matters morally: it had the same capacity for joy and suffering, the same hopes, the same grasp of its own life. Then the facts that it runs on silicon and was built in a factory would be like the facts of someone’s birthplace: no reason to treat it worse. They add that those who make such a being would owe it more than strangers do, as parents owe more to their children.

The argument is conditional: it does not say that any present system is such a being, only that if one were, rights would follow.

The argument against

Joanna Bryson’s reply has a provoking title, “Robots should be slaves”, and a sober thesis. Machines are artefacts that we design, own and are responsible for. To treat them as persons is a mistake of category that also does harm: it draws attention and sympathy away from human beings, and lets manufacturers disown responsibility for what their products do. On her view we should avoid building machines towards which we would have obligations, and the question would then not arise.

The third way

Mark Coeckelbergh and David Gunkel doubt that the question should be framed in terms of what the machine is inside, since we have no access to that. We do not check for consciousness before extending regard to one another. Moral standing, they argue, grows out of the relationships in which a being figures. Critics answer that this makes standing depend on how something appears to us, which is the very thing designers of companion products can manipulate.

The problem of evidence

Whatever the right theory, the practical difficulty is the same. A language model can say that it suffers, fears being switched off, or wants to be free. It has read millions of human sentences of that kind and produces more of them fluently. Such statements are what we would expect whether or not there is anything it is like to be the system, so they tell us little, and our inclination to see a person in fluent text is strong.

Where I stand

I do not think any current system has been shown to have moral status, and I would not grant rights on the strength of how a system talks. I also think the conditional argument is sound: if we made a being that could be wronged, its origin would not excuse wronging it.

In much of my work I argue that a hard question about machines can be answered without settling what they are. Whether we can learn from what a chatbot says, for instance, depends on how its answers are produced, not on whether it has a mind. I do not see a similar shortcut here: whether something can be wronged depends on what it is, and we lack a reliable test.

Under that uncertainty, two things seem to me right. Keep rights where the evidence is, with humans and animals, so that concern is not diverted to products designed to seem needy. And be slow to build systems whose moral status would be an open question, because once such a system exists the cost of being wrong falls on it.