Reflections on AI Ethics from Bangkok

At the end of a panel on the ethics of artificial intelligence, in Bangkok, a Buddhist monk stood up and said, very gently, that we had been asking the wrong question.

We had spent an hour on responsibility and regulation, on what should stay human once machines begin not just to answer us but to act for us. He wanted to bring a different dimension to it. The whole discussion, he said, had been set up as human beings versus AI — whether we are superior or inferior, what we keep and what we concede. But artificial intelligence might be an integral part of life. Or life might be an integral part of artificial intelligence. Life has its own mechanisms; it is already a kind of intelligence, a cosmic one. We carry in us the infinite data of the cosmic mystery, and the infinite depth of the cosmos. We too are intelligence — not digital intelligence but living intelligence. Could we not, he asked, deepen the question in that direction, and ask how AI, humanity and life might help one another? And then he repeated what he had asked the day before, in his own session: what kind of human beings are we becoming in this age of AI?

He was declining our question. Everything we had said that morning assumed two parties: two intelligences, one natural and one artificial, sitting across from each other with a set of tasks to divide up. What should remain human is a question of that shape; it takes for granted that there is a territory and a border, and asks where the border should fall. He was pointing at the border itself, and at the fact that we had drawn it.

Most of the debate lives inside that frame, once you look for it. Almost everything we argue about artificial intelligence comes down to a question about what kind of thing it is, and therefore what it possesses and how much of it: does it really think, does it understand, is it conscious, is it creative, will it be better than us. Turing’s famous test is that question made practical — if you can’t tell, from the answers alone, whether you are talking to a person or a machine, then on what grounds do you deny it the word “thinking”? Searle’s Chinese Room is the same question from the other side, arguing that manipulating symbols by rule is not understanding them, however convincing the output. Turing and Searle are usually staged as opponents, and both are after the same verdict: what does the machine have. The friendly formulations keep the shape too. Calling AI a partner, or a collaborator, or a colleague warms the relation and leaves its structure alone; there are still two parties, and we are still deciding what to make of the second one.

That shape has a history. It comes out of a long habit of treating intelligence as a property held by a subject, so that the natural thing to ask about anything new is which subjects have it and to what degree. It is the working frame of the conversation as it is now conducted — in the papers, in the panels, in the regulations, in the industry — and it is so much the water we swim in that it takes someone standing outside to point at it.

My own work sits inside it. I have been developing an oath for people who design and use artificial intelligence, and the argument I make for it is of exactly that type: it starts from a question about what kind of beings we are, and about what follows from that.

Which matters more than it might seem, because the point of the project is that it should be universal. A global regulation is out of reach, and would miss the point anyway. The universality I am after is harder: something that can be formed in people, in different places, out of different traditions, and still be recognisably the same commitment. I had always taken that to be a question about values. Find the convictions that recur across traditions, state them well, and the thing travels.

Ten minutes in Bangkok were enough to make me doubt it, because you can agree on dignity, on responsibility, on care, and still be working inside a set of questions that one tradition happened to set up. You can inherit the questions from somebody else without noticing, and the agreement makes that harder to see, since it looks like proof that everything is in order. If it is the frame that fails to travel, then a project like mine could be endorsed everywhere and still be provincial, universal in its words and local in its structure. Which is why a conversation of that kind is how you find out whether the structure survives translation. You cannot test a frame from inside it.

So I am left with a question I did not have before, and I am not sure how it resolves. Is the polarity something provincial, to be got past — or is it the condition of being able to answer for anything at all?

Responsibility needs two terms. It needs me to be able to say that this one is mine and that one is not: I did this, the system did that, here is the line. Take the two terms away and something is gained, since the false contest disappears, and with it a set of questions about superiority that were never going to be answered well. Something is also lost, and I don’t know how to weigh it against the gain: the possibility of standing somewhere and saying that I am the one who answers for this.

I cannot work that out on my own, which was the monk’s gift. The room I had been arguing in had walls, and he was standing outside them, asking a question I had not thought to ask.