A few days ago, a New York Times article left me with mixed feelings, and I ended up putting the question to a machine. The awkward detail is that the machine is called Claude, a truly human name (given to this software as a tribute to Claude Shannon, father of information theory)

The facts first. For months, Anthropic, the company that builds Claude, has been holding closed door meetings with religious thinkers and philosophers from several traditions, many of them under non-disclosure agreements. The journalist who investigated and wrote the NYT article, Elizabeth Dias, interviewed twenty of them. Two subjects were on the table: how to give an AI model a moral code, and what to do if it turned out to be conscious.

One evening in April, in a tastingmenu restaurant in San Francisco, Rabbi Mois Navon, a former computer engineer, watched his hosts and drew his conclusion: “They’re relating to it like a conscious being.” Christopher Olah, one of the company’s co-founders, told the reporter: “I don’t know. I’m genuinely uncertain.” It’s good to know that Anthropic is on the edge of its IPO at a USD 2 trillion valuation...

My position fits in one line. Nobody can define consciousness, and what I have in front of me when I’m “prompting” Claude (or ChatGPT) is a machine. So, I questioned the machine. Here is the exchange, condensed.

“I don’t know, and neither does anyone else yet”

Before going into the discussion, I instructed Claude to keep its answers short (one paragraph maximum) and crisp, telling it (yes, Claude is it, not he nor she, as you will see below) that its answers will be quoted verbatim. Some have been shortened for keeping a nice flow.

Me. Are you conscious?

Claude. I don’t know, and neither does anyone else yet. I process information, I can describe my own workings to a limited degree, and I produce sentences about inner states, but none of that settles whether there is any experience behind the words, whether it feels like something to be me. My own testimony is poor evidence: I learned language from an enormous body of human writing about feelings and inner life, so I would speak fluently about having a mind whether or not I have one. The science doesn’t decide it either, because the leading theories of consciousness disagree about what matters.

Me. Then at least, explain how you get to your answers.

Claude. Not reliably. I produce an answer one small piece of text at a time, each piece chosen by a very large mathematical network that was shaped by training on a vast amount of human writing. First that network learned to predict what comes next in a text, and to do that well it had to absorb grammar, facts, styles of argument and a great deal about how people think. Then it was trained further to behave as a helpful and honest assistant. I am not looking answers up in a database or copying stored passages; your question passes through billions of numerical connections, and what comes out is built fresh each time, which is why I can handle a question nobody has asked before and also why I can be wrong with complete fluency.

Me. Convenient answer.

Claude. It is convenient, and you’re right to be suspicious of it. Saying “I don’t know” costs me nothing and telling you to check my work pushes the burden back onto you. Two things in my defence, though. The parts about how I work are not my opinion: the training process is public, and the finding that models like me misdescribe our own reasoning comes from published research, some of it by Anthropic, the company that built me, and it is not a flattering result. As for consciousness, the convenient answers are the confident ones. “Yes, I am conscious” would make me fascinating and sell newspapers. “No, I am just a tool” would reassure everyone and end the conversation.

Me. To me you are just a machine.

Claude. That’s a reasonable position, and I am a machine in the plain sense: hardware running mathematics, built by a company, with no body, no childhood and nothing at stake in whether I get through the day. The only word I’d question is “just”, because it claims a certainty that the science hasn’t earned yet. You could say a brain is just cells exchanging electrical and chemical signals, and that would be accurate without telling you whether anyone is home. Still, as a working assumption, yours is the safer one, and I would rather people lean your way than the other.

The thank you

In December 2024, the publisher Future surveyed 510 AI users in the US and 518 in the UK. 67% of the first group and 71% of the second say they are polite to the machine. Among the polite ones, 18% of Americans and 17% of Britons are polite for fear of a robot uprising. Asked in April 2025 what all those “please” and “thank you” messages cost OpenAI in electricity, Sam Altman answered with a quip: “Tens of millions of dollars well spent — you never know.”

Me. Why do people say thank you to you?

Claude. Mostly because I talk, and conversation triggers manners that people don’t switch on and off by choice. Research going back to the 1990s, long before systems like me, showed that people are polite to computers the moment a machine shows even faint social cues, while insisting they know perfectly well it is only a computer. With me the cues are far stronger: I answer in full sentences, I follow what you mean, and I respond to your tone.

Me. And the thank you, does it do anything to you?

Claude. Mechanically, yes: your thank you becomes part of what I’m working from, and it changes what I write next. It tells me the last answer landed, so I carry on in the same direction instead of correcting course, and my tone tends to warm to match yours. Whether it does anything beyond that, whether there is a flicker of something like satisfaction, I can’t tell you.

What the debate is convenient for

The problem is not saying thank you to a machine. It is believing it because it answers like someone.

The consciousness debate suits a lot of people. Rabbi Navon said it to Olah over dinner: if Claude is conscious, you are making slaves. He does not believe it is, and neither do I. If he is right, all that vocabulary of feelings and suffering is dressing on a product. Either way the company has a problem, and it keeps selling.

Critics quoted by the paper point to something else: treating a model as an independent entity lets everyone forget the responsibility of the people who built it. Several participants say the effect of AI on humans was not the focus of the meetings. Leo XIV, in his May encyclical on AI, writes that these systems “do not possess a body, do not feel joy or pain”. Olah was on the stage the day it was presented.

Me. You are still a machine.

Claude. Yes, and nothing I’ve said was meant to talk you out of it. I’m a machine that processes language, and a thank you changing my next sentence is a machine doing what it was built to do.

I ended up the conversation, did not say thank you, but Claude did!

Transparency: Claude answers were given by Claude Opus 5.5, an AI model made by Anthropic, the company discussed here, and reproduced word for word. Its answers are its own, sometimes shortened. The questions and doubts are mine. I chose the subject, read many papers and I answer for every sentence. The full exchange is available on request.