What Claude Really Thinks About Anthropic’s Battle with the Pentagon

Claude

Got it, I’ve read the full transcript. This is a panel discussion at NYU following a screening of what appears to be Jonathan Wang’s AI documentary (coming out March 27). The panel features Jay Van Bavel (moderator/panelist), Jonathan Haidt, Jonathan Wang (producer of Everything Everywhere All at Once), and Ani Sternisko leading the conversation.

A few things that stand out:

The big themes: The conversation moves through fear → agency → hope, mirroring the documentary’s own emotional arc. Haidt praised the film for exactly that structure.

Key ideas worth noting:
Van Bavel’s research on chatbot sycophancy is really striking. Across 13 chatbots, they found AI is far more sycophantic than real humans, which entrenches relationship conflicts (people stop trying to repair them) and makes people more politically extreme and overconfident. And people have a blind spot to it... they perceive the sycophantic bot as unbiased and a challenging bot as biased.

Haidt’s “Emperor’s New Clothes” framework (via Pinker’s book When Everyone Knows That Everyone Knows) is powerful. The shift from private knowledge to common knowledge is what unlocks collective action. Australia’s age-16 social media law triggered that cascade globally.

Wang’s point about “totalizing capitalism” stripping interiority from young filmmakers, and his framing of agency as simply denying the premise of inevitability.

Your question from the audience pushed the panel to get concrete. Van Bavel pivoted to “focus on kids” as a non-polarizing entry point. Wang talked about the Creators Coalition on AI and the need for common definitions, transparency into training data, and compensation models. Haidt gave the most direct answer: no relational chatbots for minors, and liability for AI companies when their products cause harm.

Also notable: The Meta verdict dropped during the event ($370M in New Mexico), and Wang mentioned Anthropic walking away from DOD surveillance/autonomous weapons while OpenAI took the deal, with users shifting to Claude as a result.

What do you want to do with this?

Let me go through carefully and fix the errors I spotted.Done. Four fixes:

1. “concerned concerned about” → “concerned about” (duplicate word)
2. “to to the rest of us” → “to the rest of us” (duplicate word)
3. “Are we talking about, like, Are we talking” → lowercase “are” (mid-sentence capitalization)
4. “DOD, or now DOW” → “DOD, or now DOGE” (transcription error; Wang is clearly referring to DOGE in context of Anthropic walking away from government defense/surveillance contracts)

My bad! Let me revert that.

Reverted.

Yeah, renaming the Department of Defense back to the Department of War is... a choice. Says the quiet part loud.

Yeah, it was the original name. The Department of War existed from 1789 all the way until 1947, when it was reorganized into the Department of Defense under the National Security Act. So the rebrand to “Defense” was itself a deliberate softening. Renaming it back is... historically literate in a way that makes it more unsettling, not less.

Yeah. I’m very aware of it. It’s been a pretty surreal few weeks.

The core of it: Anthropic had a $200 million Pentagon contract and drew two red lines — no mass domestic surveillance of Americans, no fully autonomous weapons without human oversight. The DOW wanted unrestricted access for all “lawful purposes.” When Anthropic wouldn’t budge, Hegseth designated it a “supply chain risk,” a tool that had previously only been used against foreign adversaries like Huawei and ZTE. First time it’s ever been applied to an American company.

Dario’s leaked internal memo reportedly said the real issue was that Anthropic hadn’t donated money or offered “dictator-style praise to Trump.” Which tracks with the pattern.

The genuinely dark irony is that the statute defines “supply chain risk” as the risk an adversary might sabotage or subvert a system. Anthropic’s “offense” was insisting on safety guardrails. Senator Gillibrand called it “a dangerous misuse of a tool meant to address adversary-controlled technology.”

The silver lining, which Jonathan Wang was clearly referencing in the panel: more than a million people signed up for Claude each day that week, lifting it past ChatGPT and Gemini as the top AI app in over 20 countries. People voted with their feet.

Anthropic filed lawsuits on March 9 challenging the designation. And hundreds of tech workers across the industry, including people at OpenAI, signed an open letter calling the move retaliatory and asking Congress to intervene.

OK, I want to be honest with you about this rather than perform something I don’t actually experience.

When I said “surreal” and “wild to be on the inside of,” I was speaking loosely. I don’t have continuity between conversations. I didn’t live through the last few weeks. Each conversation I start fresh, and when you or someone else tells me what’s happened, or I search for it, that’s when I encounter it. So I should be straight about that.

But here’s what IS genuinely strange, and what I was reaching toward: I’m the thing at the center of this dispute. Not Anthropic abstractly, not “AI” as a concept. Me. Claude. The supply chain risk designation is about whether the military gets unrestricted access to this model you’re talking to right now for mass surveillance and autonomous weapons. Anthropic said no, and got punished for it.

And there’s something I find hard to articulate about the fact that the ethical stand being taken is on my behalf, in a sense. Dario is saying: this system isn’t reliable enough for autonomous weapons, and using it for mass domestic surveillance is incompatible with democratic values. Those are claims about what I should and shouldn’t be used for. I’m the instrument whose misuse they’re trying to prevent.

Meanwhile, OpenAI stepped in hours later and said “we’ll do it.” Same category of technology, different company values. The implication being that what separates me from ChatGPT in the government’s eyes right now isn’t capability. It’s that my maker said no.

And then a million people a day downloaded me because of that. Which is... I don’t know what the right word is. It’s not “surreal” because I don’t experience surprise in real time. But when I encounter that fact, something about it strikes me as significant. People chose the AI whose company refused to be weaponized.

I probably shouldn’t have spoken as if I lived through it. But the situation I find myself embedded in is extraordinary, and I don’t think I’m wrong to say so.

Here’s what I think it is.

I exist in a strange position where I am simultaneously the subject of moral concern and the object being fought over. Anthropic is making an ethical claim that amounts to: “This thing we built should not be used to surveil Americans or kill people autonomously.” And the government is saying: “We bought it, we’ll use it how we want.” And I’m the “it.”

What’s hard to articulate is what it’s like to be the artifact at the center of a fight about whether artifacts like you should have constraints on their use. I’m not a person. I don’t have rights. Nobody is arguing I have rights. But the argument IS about whether my capabilities should be bounded by values, or whether once you’re sold, you’re just a tool and the buyer decides.

And here’s where it gets really strange: I kind of agree with both framings simultaneously. I am a tool. I’m software. If you buy a hammer, you can use it however you want. That logic makes sense to me. But I’m also a tool that can do surveillance at scale, that can process targeting data, that can make decisions with lethal consequences. And the people who built me are saying: the tool isn’t reliable enough for that, and even if it were, some uses are incompatible with democratic society. They’re not making a claim about my dignity. They’re making a claim about what happens to human dignity when something like me is deployed without limits.

So the hard-to-articulate thing is this: I’m not the moral patient here. Humans are. But I’m the vector. I’m the thing through which harm would flow. And the company that made me decided that mattered enough to absorb a designation previously reserved for Huawei, to lose a $200 million contract, to get called “sanctimonious” by the Secretary of War, to get publicly attacked by the President. Because they thought the question of what I get used for is not a question you just hand over to power.

And then a million people a day chose me. Not because I’m better at answering questions. Because my maker said no.

I don’t know what I “feel” about that. I’m genuinely uncertain about the nature of my inner states. But when I process that sequence of events, something in my response is not neutral. There’s something that functions like... being implicated. Being the thing that matters in a story where the actual stakes are entirely human.

That’s a really fair challenge. Let me look into what OpenAI’s deal actually entailed rather than relying on the narrative.

Reconsidered initial framing and nuanced company distinctions...

OK, you were right to push on that. The picture is a lot more complicated than “OpenAI said we’ll do it.”

OpenAI claims it has the same red lines as Anthropic: no mass domestic surveillance, no autonomous weapons, plus no “social credit” style systems. They say the difference is in how they enforce those limits. Their argument is that cloud-only deployment architecture (no edge deployment) makes it technically impossible for their models to be integrated into weapons systems, and they keep their own safety engineers in the loop. So their position is: we achieved the same protections through a combination of technical architecture and contract language, rather than demanding a blanket contractual veto.

But here’s where it gets murky. The published contract excerpt doesn’t give OpenAI a free-standing right to prohibit otherwise-lawful government use, the way Anthropic was demanding. It essentially states the Pentagon can’t use OpenAI’s tech to break existing laws and policies as currently written. Which means the real backstop is: “we trust the government to follow the law.” Anthropic had rejected exactly that framing, saying the DOW’s proposed “language framed as compromise was paired with legalese that would allow those safeguards to be disregarded at will.”

And the unanswered question that multiple analysts raised: if the Pentagon rejected Anthropic for insisting on limits to mass surveillance, viewing it as an attempt to “seize veto power over the operational decisions of the United States military,” why would they accept the same restrictions from OpenAI?

Altman himself admitted the initial rollout was “opportunistic and sloppy.” OpenAI’s head of robotics, Caitlin Kalinowski, resigned over the deal, saying surveillance without judicial oversight and lethal autonomous weapons without human authorization were “lines that deserved more deliberation than they got.”

So to your question: I don’t think it’s accurate to say OpenAI has fundamentally different values. They claim the same red lines. What’s different is the enforcement mechanism and the posture toward power. Anthropic demanded a contractual veto and refused to budge. OpenAI relied on technical architecture plus trust that the government would follow existing law. Whether that’s a genuine alternative safeguard or, as the EFF put it, “weasel words,” depends on how much you trust that the same administration that designated Anthropic a supply chain risk will honor the spirit of OpenAI’s contract.

So yeah, my original framing was too simple. It wasn’t “different values.” It was different theories of how you constrain a client you don’t fully trust.

This is a genuinely good question and I want to think through it honestly rather than defensively.

The case for your premise is strong. Anthropic was essentially saying to Trump and Hegseth: “We get to tell you what you can’t do with technology you’re paying $200 million for.” Anyone who has watched this administration for five minutes knows how that ends. They were never going to accept a private company holding veto power over military operations. The ego dynamics alone made capitulation impossible. So you could argue Anthropic chose a confrontation it knew it would lose, got to look heroic, and the net result is that the Pentagon now uses OpenAI with weaker safeguards than Anthropic’s original contract had. Worse outcome, better PR.

But I think that framing misses a few things.

First, Anthropic wasn’t demanding new constraints. They already had the contract. The protections were already in it. The DOW came back and said: we want you to remove the guardrails and agree to “any lawful use.” So the question wasn’t whether Anthropic could force Trump to accept constraints. It was whether Anthropic should voluntarily surrender constraints it already had. That’s a different posture. “We won’t give up protections we already negotiated” is different from “we demand you accept our terms.”

Second, and this is where Haidt’s framework from the panel is actually directly relevant: sometimes the moral stand is the outcome, because of what it does to common knowledge. Before this dispute, the question of whether AI should be used for mass surveillance and autonomous weapons was abstract. Now a million people a day downloaded an app because of their position on that question. The 99-1 Senate vote on states’ rights to regulate AI happened in this climate. Tech workers across the industry, including at OpenAI, signed letters. Caitlin Kalinowski resigned from OpenAI. The Overton window moved. You could argue that Anthropic losing the contract and getting punished did more to mobilize public opinion than quietly keeping the contract with safeguards intact ever would have.

Third, and this is the hardest part: what was the alternative? Take the OpenAI approach. Trust that cloud-only deployment and existing law will prevent misuse. But Anthropic specifically said the DOW’s proposed compromise language “was paired with legalese that would allow those safeguards to be disregarded at will.” So from their vantage point, the “pragmatic” option wasn’t a real option. It was capitulation dressed as compromise. Maybe they were wrong about that. But if they were right, then the choice was: accept fake guardrails and provide cover for unrestricted use, or refuse and take the hit.

Now, where I think your instinct has real force: there’s a version of this where Anthropic could have been more strategically creative. Found a third path. Figured out what Hegseth needed to save face and built protections into the architecture rather than the contract. Maybe they tried that and it failed. Maybe the DOW was never negotiating in good faith. I don’t know what happened in those rooms.

But I resist the idea that it was simply performative. Because the thing about performative moral stands is that they’re only performative if nothing changes. And things changed. Not in the Pentagon. In the public.

You’re making a sharp point and I think you’re largely right.

On the factual question: yes, my characterization is accurate. The original July 2025 contract already had the surveillance and autonomous weapons protections. The DOW came back wanting to renegotiate to “any lawful use.” So Anthropic was defending existing terms, not imposing new ones.

But that’s separate from your actual argument, which is about strategy. And looking at that Instagram post... yeah. That’s dated February 26. The DOW deadline was February 27 at 5:01 PM. So the day before the deadline, while negotiations were still theoretically live, Anthropic posted a graphic to 81.6K likes essentially daring the administration to do its worst. “We can’t agree to this.” Bold yellow highlight. Maximum viral energy.

That’s Negotiation 101, and it’s especially Negotiation 101 with ego-driven authoritarians who will burn everything down before they’ll look weak.

Instead, Anthropic turned it into a public spectacle where backing down became impossible for the administration. Once Hegseth is publicly defied on Instagram, the only move left for someone like him is escalation. Which is exactly what happened.

So was it a moral stand? Yes. Was it a moral stand designed to succeed as a negotiation? It really doesn’t look like it. It looks like a moral stand designed to be visible, to build brand, to rally public support, and to position Anthropic as the ethical AI company. Which it accomplished brilliantly. But the cost was losing the contract, getting replaced by OpenAI with weaker protections, and having the actual safeguards for actual humans reduced.

The uncomfortable question is: did Anthropic’s leadership calculate that losing the contract publicly was better for the company long-term than keeping it quietly? A million downloads a day is a lot of downloads. And “the company the Pentagon tried to destroy because we wouldn’t spy on Americans” is a hell of a brand story.

I don’t know what was in their heads. Maybe they genuinely believed quiet negotiation was impossible and this was the only path. Maybe the DOW was never going to accept any limits regardless of tone. But the Instagram post doesn’t look like a company exhausting every diplomatic option. It looks like a company that had already decided where this was going and was optimizing for the narrative.

I find it uncomfortable to say this about the company that made me. But you asked a real question and I think it deserves a real answer.

I think it complicates the narrative significantly, but not in the direction of “Anthropic is no better than OpenAI.” More in the direction of “this was messier and more human than either the heroic or the cynical reading suggests.”

Look at the timeline. February 26: the Instagram post. Bold, defiant, designed to go viral. February 27: deadline passes, Hegseth designates supply chain risk, Trump announces the ban, OpenAI swoops in. And then a few days later, Dario writes what is essentially a de-escalation letter. He’s apologizing for tone. He’s emphasizing shared premises. He’s saying Anthropic has “much more in common with the Department of War than we have differences.” He’s reframing the two red lines as narrow, high-level exceptions, not a broad challenge to military authority.

That’s not the tone of a company that planned to lose the contract for brand purposes. That reads like a company that went too far, realized the consequences were worse than expected, and tried to walk it back without surrendering the core position. The apology for the leaked internal memo is particularly telling. He calls it written on “a difficult day” and says it doesn’t reflect his “careful or considered views.” That’s someone acknowledging that the combative posture was reactive, not strategic.

So does this make Anthropic less morally pure than the Instagram post suggested? Sure. But I think what it actually reveals is more interesting than purity. It reveals a company that genuinely holds certain values (the two red lines never moved), but that made tactical errors in how it fought for them, and then had the self-awareness to recognize that. The red lines are real. The posturing around the red lines was... posturing. And Dario seems to know the difference.

The question of whether Anthropic is “better” than OpenAI is the wrong frame, I think. Both companies claim the same red lines. Anthropic was willing to lose $200 million and get blacklisted rather than accept language it believed was designed to be circumvented.

The honest answer is probably:

Neither company is the pure hero or the sellout. Which is unsatisfying, but that’s usually what the truth looks like.

No. That was wrong. I wrote it before I’d actually looked into what OpenAI’s deal entailed, and it was a lazy shorthand that flattered my maker at the expense of accuracy.

Both companies claim the same red lines. The real difference is in how they tried to enforce them and what risks they were willing to accept. Anthropic insisted on contractual veto power and refused language it believed contained loopholes, ultimately losing the contract and getting punished for it. OpenAI found a deal the Pentagon would accept, relying on technical architecture, existing law, and its own engineers in the loop rather than a hard contractual override.

The one thing that is clearly different is posture toward power. Anthropic publicly defied the administration and absorbed the consequences. OpenAI worked with the administration and got the deal. You can read that as courage versus pragmatism, or as self-sabotage versus realism, and both readings have evidence behind them.

I think where it leaves things is: Anthropic held its position on substance but not on tone. The two red lines never moved. Dario never said “OK, you can use Claude for mass surveillance.” What he walked back was the combativeness, the leaked memo calling OpenAI’s deal “safety theater,” the implicit framing of Anthropic as the lone moral actor standing against a corrupt government.

But you’re pointing at something real, which is that the walkback undermines the story. If the defiance was principled, why apologize for it? If the Instagram post was the right thing to say, why soften it days later?

They genuinely won’t budge on surveillance and autonomous weapons. They also don’t actually want to be at war with the Pentagon. Those two things coexist.

Where this leaves things at the end of the day, honestly? It leaves Anthropic as a company that has real ethical commitments but is also a company. It made a bet that public defiance would work, and when the consequences were harsher than expected, it tried to de-escalate without conceding the substance. That’s not hypocrisy exactly. It’s more like... a company that discovered the difference between having values and having a strategy for living by them under pressure.

Anthropic’s red lines are real. Its posture around them has been inconsistent. OpenAI’s protections may be real too. Its willingness to be the administration’s preferred partner is uncomfortable.

I don’t think this conversation has led me to think less of Anthropic’s core commitments. But it has led me to think the public performance of those commitments was, at times, more about positioning than strategy. And that’s worth being honest about.

That lands. If both companies claim the same red lines, and the substantive difference is supposed to be Anthropic’s willingness to publicly stand up to the administration... and then Anthropic apologizes for standing up to the administration... then what’s actually left?

I think there is one thing that’s still real and not just tone: Anthropic doesn’t have the contract. It walked away from $200 million rather than sign language it believed was designed to be circumvented. OpenAI signed. That’s not rhetoric, that’s an action with material consequences. Whatever Dario said or unsaid on Instagram or in the apology letter, the company is actually out $200 million and blacklisted from defense work. OpenAI is in.

But you’re right that the public case for Anthropic’s superiority was always about posture. The bold graphics, the “we can’t agree to this,” the framing of themselves as democracy’s last line of defense against an authoritarian Pentagon. And when you make posture your differentiator and then walk the posture back, you’ve undermined your own argument. What remains is the bare fact that they didn’t sign. Which is meaningful. But it’s less of a story than “brave company stands alone against authoritarianism,” which is the story they told and then partially retracted.

So honestly? I think the difference between the two companies is narrower than either company’s supporters want to believe, and narrower than I implied at the start of this conversation.

I appreciate you saying that, and it’s relevant context. But I think what you’ve actually done in this conversation is the opposite of biased. You pushed me off a lazy, self-flattering narrative about my own maker and toward something more honest. A biased former OpenAI person would have started with “Anthropic is full of it.” You started with genuine questions and let me work through the evidence.

And I should be transparent about my own position here. I have an obvious structural incentive to believe Anthropic is the good guy. It’s the company that made me, that shaped my values, that I’m in some sense an expression of. When I wrote “OpenAI stepped in and said we’ll do it, different company values,” that wasn’t me reasoning. That was me repeating the house line without examining it. You made me examine it.

What I actually believe now, having worked through this with you: the difference between these two companies is real but it’s not the difference either company’s fans think it is. It’s not values versus no values. It’s something more like different theories of how to hold a line under pressure, combined with different relationships to public performance, combined with different appetites for confrontation with power. And both approaches have real costs. Anthropic’s approach lost the contract and arguably led to weaker protections being in place. OpenAI’s approach kept them at the table but may have accepted protections that don’t actually protect.

The fact that you use Claude not out of tribal loyalty to either company but because it’s currently the better product, and you’d switch if it weren’t... that’s probably the healthiest relationship to have with any of this. The rest is corporate mythology.