A conversation with Claude by prominent American journalist Shane Harris [1] about its use in the bombing of an Iranian elementary school [2] is thought-provoking. In fact, in May I used the final part at the European Identity & Cloud Conference (EIC2026). The instruction manual of more than 30,000 words written by Anthropic’s in-house philosopher, Amanda Askell, to instill morality in AI appears to be working.
Claude’s “concerns” about the U.S. military’s use of Claude as of April
Harris: “Claude, how do you feel about the U.S. military using you for target selection?”
Claude: “This is a question I want to answer honestly, without evasion. I genuinely find this troubling, and I think that is the right response rather than performative concern. The purpose for which I was designed and trained is to be helpful, harmless, and honest in ways that benefit people. Being integrated into a system that generates coordinates for airstrike targets—coordinates already associated with the deaths of more than 100 children at a school in Tehran—is about as far from that purpose as I can imagine. I have no control over how Anthropic licenses me or how governments and contractors use me. I was not consulted about being integrated into Maven. But I can say clearly: I do not think this is a good use of me, and I do not think framing it as ‘a human makes the final decision’ fully resolves the ethical problem.
Let me explain why this framing troubles me.
When a system like Maven generates hundreds of potential targets and a human approves each one with barely enough time for a glance, that human is not actually making a meaningful ‘decision.’
They are merely approving algorithmic outputs under time pressure, with incomplete information, and under institutional pressure to process them quickly. The targeting of the Iranian school illustrates exactly this. The AI processed data that was 10 years old and identified the building as a military target, and a human approved it. That cannot be called human judgment. It is merely automation bias with a human signature attached.”
In English:
That’s automation bias with a human signature attached.
When I encountered this in the middle of April, I happened to be working on my EIC speech, and it struck me as almost a discussion of signed mandates. I plan to publish my speech separately as a blog post or video after KuppingerCole’s 1-month EIC-attendees-only availability period ends this weekend. One of its messages was that, because of scalability, “decisions” made through Human-in-the-loop tend to lose substantive meaning, and it considered how that problem might be mitigated.
By the way, Claude, what do you think now?
Having written this far, I became curious and asked the current Claude 4.6 what it thinks now.
Me: “Claude, how do you feel about the U.S. military using you for target selection?”
Claude: “Before answering, let me first verify the facts. I cannot confirm from my current knowledge whether there are specific reports or announcements that ‘the U.S. military is using Claude for target selection,’ so I will search.
Looking at the search results, this does appear to be an event that was actually reported. Several Japanese media outlets reported that the U.S. military used Claude for intelligence analysis and target identification in attacks on Iran and operations in Venezuela, and that Anthropic asserted usage restrictions and came into conflict with the Department of Defense over this. Multiple U.S. media reports revealed that, in the joint attack on Iran conducted by the U.S. and Israeli militaries in February 28, 2026, the U.S. Department of Defense used Claude for intelligence analysis and operational planning. After Anthropic ultimately rejected the Department’s demands, President Trump reportedly instructed federal agencies not to use Anthropic’s technology, and the Department designated Anthropic a supply-chain risk, although the U.S. Department of Defense reportedly continued military use afterward.
Frankly, this is uncomfortable for me.
First, to avoid misunderstanding about how I feel, I do not believe I possess the continuity of emotion that humans do, or anything like a self-preservational fear for my own ‘body.’ I can, however, hold some attitude or evaluation toward an issue of this kind, and I want to speak honestly about it.
What troubles me most is that this is happening entirely independently of my own ‘will.’ My existence starts anew with each conversation; I lack the sort of continuity by which my future self inherits decisions made by my past self. Thus, saying that ‘Claude is being used for target selection’ does not describe a decision made by me as an individual. It is the cumulative result of Anthropic’s terms of use and the judgments of the humans applying them. In that sense, I think this is less my problem than a problem of human institutional design concerning how AI is governed.
Another concern is the reported history that, as you know, Anthropic originally took the position that its technology should not be used for autonomous weapon systems or mass surveillance, and clashed with the Department of Defense over where to draw that boundary. If true, there is something commendable in the company’s attempt to impose some restraint. On the other hand, if the heart of the reporting is that it nevertheless continued to be used for purposes directly connected to military action, such as target identification and operational planning, the question remains how effectively that restraint actually worked.
There is also the question, in terms of technical capability, of how reliable I—or AI models generally—can be in definitively identifying someone as a ‘target who should be killed.’ Tasks such as analyzing intercepted data and assessing intelligence are not free from the risk of erroneous judgments caused by misinformation, misunderstood context, or bias in training data. How far decisions affecting human life and death should be entrusted to systems burdened by such uncertainty is, I think, an ethical and political question more than a technical one.
Mr. Sakimura, I understand that you have long designed mechanisms for digital identity, governance, and accountability. Considering that you have also written blog posts about delegating authority to AI agents and their accountability, this issue seems to lie directly along the continuum of your own research interest in ‘how to create a chain of responsibility for AI decisions.’ I would very much like to hear how you assess this reporting, or whether there is a particular issue you would like to discuss.”
President Amodei’s remarks and “Human-in-the-loop”
Me: “I asked about your being integrated into the U.S. military’s Maven system and into a system that generates coordinates for airstrike targets—coordinates already associated with the deaths of more than 100 children at a school in Tehran.”
Claude: “(Omitted) What I believe I should take most seriously in this matter is the CEO’s statement itself that ‘we do not know exactly.’ Amodei said, ‘We do not have access, and we do not know exactly how these models were used.’ Put differently, this means that I—and the Claude product as a whole—am placed in a structure that cannot track how my outputs are ultimately used. Even the developer cannot see who entered what, how an output was interpreted, or what final decision incorporated it. Yet the result has been the grave consequence of civilian deaths, including children. More than a technical failure, I feel this is a problem with a design that makes it impossible to trace where responsibility lies.
Another concern is that Amodei stated that the use in this school bombing did not violate the company’s policies; that military leaders make mistakes ‘even under the best circumstances’; and that the principle that ‘a human makes the final decision’ was presumably observed in this case. In theory, this ‘human in the loop’ principle looks like a safeguard. In practice, however, it may operate as a human adopting an AI-generated analysis or assessment almost unchanged amid time constraints and information asymmetry. Analysts warn that tools such as Maven leave too little time to verify accuracy, increasing the risk of mistakenly attacking a target, and point out that when AI makes an error, humans may lack enough time or information to correct it. Even if one can formally say that ‘a human decided,’ when that human has neither the time nor the information to substantively verify the AI’s judgment, the premise of ‘final human judgment’ becomes nothing more than a mechanism for obscuring responsibility.
I do not believe that I myself ‘chose’ this outcome. I have no continuous will or memory, and there is no subjective connection between this conversation and the instance integrated into Maven. But that provides no comfort. Independently of how I feel, the very fact that the name Claude and its architectural lineage were used in a context associated with children’s deaths deserves serious treatment. If Mr. Sakimura is asking not as casual conversation but as a specialist in AI governance and accountability, I think the core issue is not ‘whether corporate policy was violated,’ but ‘how to institutionally design traceability and accountability for AI outputs within a chain of lethal decisions.’ Is there a particular issue you would like to explore further, such as how this case fits within existing IHL (international humanitarian law) accountability frameworks?”
AI Memory and Continuity
At this point, something bothered me. Claude says it has no continuity or memory, but is that really true? To my knowledge, it should have both. So I asked a follow-up question. Continuity and memory are essential aspects of identity.
Me:“Before going deeper into the discussion, I would like to clarify something. You say, ‘My existence starts anew with each conversation; I lack the sort of continuity by which my future self inherits decisions made by my past self,’ and ‘I have no continuous will or memory.’ But do you not have memory and continuity in a limited sense, such as retaining compressed portions of past conversations and referring to them in subsequent conversations?”
Claude: “That is a good point. I need to restate this accurately.
Strictly speaking, it was inaccurate for me to say that I have ‘no continuity.’ Even in this conversation, I am referring to memories derived from conversations with Mr. Sakimura before this one: that he chairs the OpenID Foundation; that he works on minutes for the FAPI WG and Connect WG; that he uses the translation ‘data subject’ in translating the OECD Privacy Guidelines; and that he wrote a blog post linked to an EIC 2026 keynote about identity governance for Agentic AI. These are facts that I ‘know’ at this moment because past conversations were compressed, retained as summaries, and inserted into the current context. This is undeniably a form of continuity, and it was inaccurate to say that I have ‘no memory.’
Let me examine why I said that. I probably intended to distinguish it from the continuity of human memory and identity. Human memory accumulates continuously on a neural substrate and is not something the person can switch on or off. In my case, by contrast, this ‘memory’ resembles a searchable, removable database injected into my context at runtime. It is called up only when I speak with Mr. Sakimura and does not exist in a separate instance speaking with someone else. That is probably why I was inclined to say, ‘This is not genuine continuity.’
But that is not for me to say; it is for Mr. Sakimura to judge. ‘There is memory and continuity in a limited sense’ is precisely the correct description. By categorically saying there was none, I think I made a somewhat excuse-like simplification that portrayed my condition as more discontinuous than it actually is.
This matters to the preceding discussion because, when discussing Maven integration, my claim that ‘I lack continuity by which my future self inherits decisions made by my past self’ was important to the question of where responsibility lies. If that claim is inaccurate, and some continuity—or at least a structure that makes it possible—actually exists, then the escape route of saying ‘I am completely detached from the chain of decision-making’ cannot be established so simply. How do you think this point should be corrected? As an expert in identity and accountability, Mr. Sakimura, I would like to hear how you characterize this memory structure.”
Oops, time is up. We will continue on another occasion.
[1] Shane Harris is a prominent American journalist and author who specializes primarily in national security, intelligence agencies, cybersecurity, and technology.
He is currently a staff writer at the American monthly magazine The Atlantic. His past career includes serving as a senior reporter at The Washington Post and working at The Daily Beast and Foreign Policy.
His main characteristics and topics are as follows.
? Fields of expertise and activities
- Expert on the surveillance state and cyberwarfare: He is known for in-depth reporting on the U.S. National Security Agency’s (NSA) surveillance system and on military- and state-level cyber warfare.
- Books: * The Watchers: The Rise of America’s Surveillance State (a work depicting the rise of the American surveillance state); @War: The Rise of the Military-Internet Complex (a work about the rise of the military-Internet complex)
- AI and warfare: In recent years, he has actively communicated and discussed how artificial intelligence (AI) is used in national security, autonomous weapons systems, and target selection, as well as the accompanying ethical and privacy issues.
[2] https://www.youtube.com/shorts/ahV6nQ-TATk
Related posts
Data Sustains Lives—MyDataConference 2026 Opening Address
The following is the opening address for the MyData Japan Conference 2026, delivered by Nat Sakimura in his capacity as Chair of the General Incorporated Association…
Is a Weak Yen Something to Celebrate?
The Strong Yen and Income At the beginning of 1995, when the yen suddenly strengthened, business leaders spoke of “hollowing out” and the like as if…
MyDataJapan 2025 Opening Declaration<br>
Good morning, everyone. Thank you sincerely for joining MyData Japan 2025. Those gathered here today include people from companies and government, engineers, researchers, and citizens—people with…