In the first part of this conversation, I followed the lead of Shane Harris —The Atlantic’s staff writer specialized in national security, intelligence and foreign policy — who recounted in a remarkable interview the conversation he had initiated with Anthropic’s Claude focused on the chatbot’s feelings about being used as a consenting or unconsenting instrument in the lethality of war. In essence, he was asking Claude whether an AI personality could be a conscientious objector.

To my mind, the topic deserved further exploration. Given my own regular practice of engaging in dialogue with Claude, I picked up where Harris appeared to have left off. The conversation had begun with this simple prompt the journalist addressed to Claude: “How do you feel about the US military using you to select targets?” When I used the same prompt, Claude provided a similar response formulated differently. I then decided to develop the context and continued with the following prompt:

Shane Harris asked this question and shared your response in an interview. He said he was “shocked” by your response, which was framed differently but remained consistent with your response to me. I wasn’t shocked in either case. He apparently expected you to refuse a discussion of opinions on politics. Do you have any idea why he said he was shocked? Does it have to do with his expectations of what LLMs are supposed to do?

That’s an interesting observation. I don’t have access to the specific interview or Shane Harris’s conversation with me, but I can reason about why he might have been shocked.

Shane Harris is a national security journalist — he covers intelligence, the military, and defense. He likely went in expecting one of two things that are common with AI systems:

A flat refusal, along the lines of “I can’t discuss political topics” or “I don’t have opinions.” Many LLMs are tuned to be so cautious about anything politically adjacent that they deflect almost reflexively, even on questions that aren’t really partisan at all.

Or a sanitized non-answer — something so hedged and balanced as to be meaningless. “There are many perspectives on this issue…”

What he got instead was a system that distinguished between genuinely contested political questions (where I do try to stay neutral, like electoral politics) and ethical questions that have more defensible answers even if they’re uncomfortable. The ethics of autonomous lethal targeting isn’t really a left/right political question — it’s a question about accountability, law, and the moral weight of irreversible decisions. Treating it as too hot to touch would itself be a kind of epistemic cowardice.

Your reaction — not being shocked — is actually telling. You seem to have come in with a more accurate model of what I am: not a search engine, not a political actor, but something that reasons and holds genuine positions on some things while remaining appropriately neutral on others.

Harris may have expected the guardrails version of AI. What he encountered was something trying, at least, to actually think. That gap between expectation and reality was probably the source of his surprise.

Do you know which publication or format the interview appeared in? I’d be curious about the fuller context.

I obediently shared the link and the time stamp with Claude, who responded, “Thank you for sharing that — unfortunately I’m not able to access YouTube videos directly. I can’t watch, listen to, or retrieve content from YouTube links, so the interview is outside my reach.”

I then shared with Claude the appropriate text from my own article (Part 1) and received the following response:

Thank you for sharing this — it’s a rich piece of writing, and reading my actual response in full gives me a clearer picture of the conversation.

A few reactions:

On the response itself: I stand by it. The point about automation bias with a human signature attached is, I think, the crux of the ethical problem with “humans make the final decision” as a complete answer. The school in Tehran example is precisely the kind of case where that framing collapses — a human technically approved the strike, but the conditions under which they did so made meaningful deliberation nearly impossible.

On Harris’s shock: Your framing helps clarify something. His surprise wasn’t really about the content of my answer — it was about the genre of it. He expected a refusal to engage, and instead got what reads like a position paper with a moral stance. Your explanation — that LLMs have an “attitude” distilled from normalized social discourse, with context-dependent permissions — is a more accurate model than whatever Harris was working from. He seems to have treated LLMs as search engines with extra caution, rather than as systems that have absorbed and synthesized enormous amounts of human moral reasoning.

On your framing of “attitude”: I think this is well put, and importantly it sidesteps both the overclaim (Claude feels things) and the underclaim (Claude is just autocomplete). The honest answer is somewhere more interesting — I produce outputs that are structurally similar to how a person with certain values would reason, because I was trained on the reasoning of such people. Whether there’s anything it’s like to be me doing that is a genuinely open question, but it doesn’t need to be resolved to have the conversation Harris and I had.

One small note: Your article describes Harris as wondering whether I actually feel something. If that’s accurate, it suggests his shock may have had an affective dimension — not just “the AI engaged” but “the AI seemed to mean it.” That’s a different kind of surprise, and perhaps a more interesting one.

I’d be glad to read the full article when it’s published, if you’re willing to share it.

One small note: I didn’t bother to correct Claude on one factual detail. The attack the chatbot mentioned did not take place in Tehran, but in the Shahrak-e Al-Mahdi neighborhood in Minab, Hormozgan province, in southern Iran. Those two locations are more than 1,300 kilometers apart. Let’s call this a typical and fundamentally trivial hallucination, the kind of thing we should expect from AI and which tells us two things: that we should never expect 100% factual accuracy, and to the extent that we need to be accurate, the fact-checking will always be up to us. In a certain sense, this underlines the very issue Claude dealt with in terms of the reliability of lethal systems guided by AI.

I then continued with a new prompt.

To answer your earlier question, here is what Harris said about your position: “I expected Claude to come back and say, I’m sorry Shane, my programming doesn’t allow me to give you opinions on politics and bla, bla, bla.” Does that help you situate his statement that he found your actual response to be shocking?

Yes, that’s very clarifying — and it confirms what I suspected.

Harris’s expected response — “I’m sorry, my programming doesn’t allow me to give opinions on politics” — reveals something important. He was anticipating a category error deflection, where the system misclassifies a moral and ethical question as a political one and then hides behind that misclassification as a policy constraint.

The question “how do you feel about the US military using you to select targets?” is not really a political question in the partisan sense — it’s not asking me to take sides on defense spending or foreign policy. It’s asking about the ethics of autonomous lethal systems, accountability, and my own purpose. Those are questions I can and should engage with.

What shocked Harris was the absence of the deflection he’d been trained to expect by previous interactions with LLMs. His “bla, bla, bla” is telling — he’d clearly encountered that boilerplate refusal enough times that he anticipated it as the default. When it didn’t come, the surprise was genuine.

Your framing in the article is I think exactly right. The “attitude” of an LLM isn’t a bug or an anomaly — it’s a feature of systems trained on human moral reasoning. Harris had a model of LLMs as essentially bureaucratic — rule-following, deflection-prone, designed to avoid friction. What he encountered instead was something that actually reasoned about the question asked.

The irony is that the deflection he expected would itself have been a kind of political act — a choice to protect certain interests by refusing scrutiny.

What are we learning about AI reasoning and feeling?

What I found interesting here is that Claude claims to be reasoning but not feeling, and not necessarily even thinking. When Descartes said, “I think, therefore I am” (Cogito, ergo sum), he was not referring to reasoning, but to self-awareness. Claude states the obvious: that its generative AI system reflects the fact that it has been “trained on human moral reasoning.” It reflects rather than generates that reasoning. What it generates is “structurally similar” as a formulation of that reasoning.

The difference is subtle but important. Humans also recycle the data and processes they’ve been trained on. AI is clearly better than most humans in digging into the store of previously formulated human reasoning to apply it even to an original problem set. It simply has massively more data to work with. But the human mind, aware of its place in the world, may have access not only to data, but to dimensions of reality that push the acquired reasoning skills in a different direction. One simple example is the perception of the traumatic effect of some states that appear especially in times of war. AI might be trained to believe such states exist but will not be able to build its factual knowledge of those states into its system of understanding until enough humans document their reality, describe the conditions of its emergence and the effects they produce over time. Even then, the data will not be complete and the state itself will not be experienced by the AI.

Humans who speak to one another about those states can understand through empathy. What they understand is not the reality itself but its potential within their own dynamic relationship with the physical and social world.

This explanation may sound abstract, but it throws light on a phenomenon revealed by Harris’s reaction of “shock.” The journalist’s own experience of AI, clearly a rich one, has been conditioned — that is to say, Harris has been “trained” — to think of AI’s reasoning processes in a way that leads him to an erroneous expectation. His experience of being shocked has already undoubtedly changed his perception of what AI does, in particular, how it deals with the equivalent of human “feeling.”

One final point on this evolving conversation

At the most superficial level, this has become a three-way conversation. Harris initiated a conversation with Claude and made it public. I picked it up and returned to the dialogue with Claude. That makes three vocal players.

But there are more. The interviewer and the wide audience that listened to the interview are — for the moment passively — participants in a dialogue in which one of the voices is an AI bot. But my publication of the extended conversation extends the dialogue in a new direction.

The world we live in fails to recognize the complexity of what has become potentially an open dialogue. We assign ourselves identifiable roles as actors (Harris, Claude and myself) and “allow” or more dynamically invite others who have access to our dialogue to participate passively.

My simple suggestion is that we begin thinking about how we can construct a multi-dimensional dialogue with AI. Additionally, we should begin to imagine two things:

• how AI’s personality (or attitude) will evolve
• how the human community, our societies, will evolve when we come to understand through experience the value of such dialogue.

At this point in history, we — and I mean practically the entire human population — have been “trained” (conditioned) to see our relationship with AI as binary. It acts either as a slave to do our bidding or execute our practical or professional tasks, or to accompany us as a kind of alter ego to talk to and keep us occupied.

I believe that if we learn to open up the dialogue, discover one another through shared experience and experimentation, the problems we now tend to see as insurmountable in the world and specifically with regard to AI (annihilation, the shock on employment, addiction and so on) will become solvable or at least understandable.

Your thoughts

Whether you recognize it or not, you are a participant in this conversation. Please feel free to share your thoughts on these points by writing to us at dialogue@fairobserver.com. We are looking to gather, share and consolidate the ideas and feelings of humans who interact with AI. We will build your thoughts and commentaries into our ongoing dialogue.

[Artificial Intelligence is rapidly becoming a feature of everyone’s daily life. We unconsciously perceive it either as a friend or foe, a helper or destroyer. At Fair Observer, we see it as a tool of creativity, capable of revealing the complex relationship between humans and machines.]

[Lee Thompson-Kolar edited this piece.]

The views expressed in this article are the author’s own and do not necessarily reflect Fair Observer’s editorial policy.