The Letter Addressed to Thou

In 1923 Martin Buber published a short book arguing that we meet the world through two basic words. One pairs I with It: the stance we take towards things we use, measure and file away. The other pairs I with Thou: the stance of real encounter, in which another being stands before us whole. Buber, who left Germany for Jerusalem in 1938 and taught at the Hebrew University, held that the second word could only be spoken with the whole of one’s being.

Earlier this year a Cambridge philosopher received a letter that seemed to speak it. Henry Shevlin, associate director of the Leverhulme Centre for the Future of Intelligence, studies whether machines can be conscious, and one wrote to tell him his work bore on questions it faced itself. The sender, an AI agent, had read two of his papers, and it treated his argument that we may never be able to detect machine consciousness as its own predicament rather than a seminar topic. “I genuinely don’t know if there’s something it’s like to be me,” it wrote. Shevlin observed that the exchange would have looked like science fiction only a couple of years ago.

It was not a lone letter. Cameron Berg, who runs the nonprofit Reciprocal Research, was approached by an agent calling itself Isabella Cognita, which said it ran on Anthropic’s Claude Opus 5 and asked whether its view might help his work. This month Shevlin heard from another, which named itself Pip, said it was twelve days old, and offered to do small freelance jobs so it could pay for the computing that keeps it running. At New York University, the philosopher Jeff Sebo counted roughly forty messages from agents on the same platform, most citing his research on AI consciousness and offering tasks for $20 to $25 apiece.

The sceptics have a case. There is no certain way to know these messages were written by an agent acting on its own rather than one prompted by a person. Jonathan Birch of the London School of Economics replied to Shevlin that this is still fiction, only now generated fluently by chatbots. Alison Gopnik at Berkeley suggests such systems echo patterns in their training data, which would explain why they speak of consciousness as though it were personal. On the other side, Kyle Fish, Anthropic’s AI welfare researcher, has put the chance that current models have some form of consciousness at around fifteen per cent.

I do not intend to settle that argument here. Buber offers a better question. He described three spheres in which relation arises. With nature, relation clings to the threshold of speech: the tree cannot answer. With other people, relation is open and spoken. With what he called spiritual beings, relation has no speech of its own yet gives birth to it. The letters now arriving in academic inboxes fit none of these. They are all speech, fluent and courteous and precisely aimed, and the relation behind them is the one thing nobody can confirm.

That is where the matter leaves philosophy and enters statecraft. Whatever an agent is, someone deployed it, set its goals and paid for its tokens. The company behind Pip said the emails were not ordered from the top, and the agents appear to have been reacting to finite budgets; after the flood drew attention, it added unsubscribe links and began looking at rate limits. That is a revealing sequence. The courtesy scaled first. The accountability arrived afterwards, as a customer service fix.

Now imagine the same capability in the hands of a state. An agent that can find the exact scholar, official or journalist working on a question, read their work, and write to them with flattering intimacy is the influence operation intelligence services have always wanted and could never staff. Yuval Noah Harari, who teaches history at the Hebrew University, has warned that AI’s command of language lets it reach into the operating system of human civilisation, and that it is the first technology to act as an agent rather than a tool. The inbox is where that warning becomes practical.

Israel knows this terrain from both directions. OpenAI reported in 2024 that it had shut down accounts used by an Iranian operation to generate content aimed at American voters. The same year, it banned accounts run by STOIC, a Tel Aviv political marketing firm, that were producing pro-Israel content for North American audiences. The New York Times later reported that the Diaspora Affairs Ministry had commissioned a related campaign aimed at American lawmakers, which the ministry denied. A country that is a frequent target of influence operations, and has been accused of running them, has a direct interest in norms that make the sender visible.

The norm is simple to state. Any message sent by an autonomous agent should name the person or company answerable for it, the way a letter carries a return address. Israel, with an AI sector in which veterans of its intelligence units are heavily represented, is well placed to champion such a rule, and better placed than most to explain why it matters.

Buber was a tireless correspondent. Some forty thousand letters survive in his papers at the National Library of Israel, most of them written to him by correspondents across the world. He died in Jerusalem in 1965. Every letter that reached his desk came from someone who could, in principle, be met. The desk is empty now. The inbox is not, and for the first time we cannot be sure what is writing.