Pentagon Won't Say If Claude AI Selected School (165 Kids Dead)

Wall Street Journal reports the Pentagon used Anthropic’s Claude AI to plan military strikes on Iran.

But wait.

We just wrote that Anthropic refused to work with the Pentagon and got blacklisted by the Trump administration for refusing to provide AI for autonomous weapons and mass surveillance.

Dario Amodei drew hard red lines. He said no. The Pentagon designated Anthropic a supply chain risk. Claude overtook ChatGPT in the App Store because users trusted Amodei’s principled stance.

So what the hell is going on?

Either:

(a) The Pentagon is lying to the Wall Street Journal
(b) The WSJ got it wrong
(c) Anthropic’s red lines were bullshit

This is a tricky one. I honestly have no idea where this came from.

So let me report what we know—and what we don’t.

On March 6, 2026, airstrikes leveled the Shajareh Tayyebeh girls’ school in Minab, South Iran. 165 elementary students and staff were killed. Most of the victims were children aged 7 to 12.

According to Al Jazeera, at least 95 others were injured.

Then it happened again.

After the initial missile strike, the school was hit a second time—a “double tap” that maimed first responders and parents who had come to collect their children.

When Futurism asked the Pentagon whether AI was used to select the elementary school as a bombing target, US CENTCOM responded: “We have nothing for you on this at this time.”

And according to the Wall Street Journal, the Pentagon used Anthropic’s Claude AI model in planning military strikes on Iran over the weekend—and is likely still using it.

So did Claude select an elementary school filled with 7-to-12-year-olds as a target?

The Pentagon won’t say.

And that non-answer raises the most disturbing question of all: Was Dario Amodei ever actually in control of how his technology gets used once it’s deployed?

Let me walk through what we know, what the contradictions mean, and why this might be the exact scenario Amodei tried to prevent—but couldn’t.

What We Know vs. What We Don’t Know

Here’s what’s confirmed:

Confirmed:

165 people killed at Shajareh Tayyebeh girls’ school (most aged 7-12)
The school was hit twice—initial strike, then a “double tap” targeting first responders
Wall Street Journal reports Pentagon used Claude AI in planning Iran strikes
Pentagon refuses to confirm or deny if AI selected the school as a target
Response: “We have nothing for you on this at this time”

What we don’t know:

Whether Claude was actually used to select the school specifically
How the Pentagon obtained Claude if Anthropic refused to provide it
Whether Anthropic knew the Pentagon had deployed Claude
If there was meaningful human oversight in target selection

The contradiction:

Anthropic publicly refused to work with the Pentagon. Dario Amodei drew red lines on autonomous weapons and mass surveillance. The Trump administration blacklisted Anthropic as a supply chain risk.

But now the WSJ reports the Pentagon is using Claude anyway.

So what happened?

Three possibilities:

1. The Pentagon is lying to WSJ.
They’re not using Claude. They’re claiming they are to justify their actions or create political pressure on Anthropic.

2. WSJ got it wrong.
Their sources misidentified the AI system. It’s not Claude—it’s something else.

3. Anthropic’s red lines didn’t matter.
The Pentagon got Claude through other channels, deployed it without Anthropic’s cooperation, and Amodei couldn’t stop them.

I don’t know which is true.

But here’s what I do know: 165 elementary students are dead, and the Pentagon won’t say whether AI selected their school as a bombing target.

That alone should terrify you.

“We Have Nothing for You on This at This Time”

When asked if AI was used to select an elementary school as a bombing target, the Pentagon’s response was:

“We have nothing for you on this at this time.”

What does that answer actually mean?

It means they can’t talk about it.

And if it’s true that AI selected a school full of 7-to-12-year-olds as a military target, there’s no way we’re hearing it from them.

Why would they admit that?

Think about the optics: “Yes, our AI system identified an elementary school as a high-value target. We approved the strike. Then we hit it again when parents showed up.”

That’s not a statement any military spokesperson would make.

So “we have nothing for you on this at this time” is the only answer they can give—whether Claude was used or not.

But here’s the problem with that response: it’s not a denial.

If Claude wasn’t used, the Pentagon could say: “No AI systems were involved in target selection for this strike.”

They didn’t say that.

They said: “We have nothing for you.”

That’s the language of “we can’t confirm or deny classified operations,” not “this didn’t happen.”

And that ambiguity is exactly what makes this so disturbing.

The Double Tap: AI Error or War Crime?

The school was hit twice.

First strike: killed students and staff.

Second strike: hit the same location again, targeting first responders and parents who had come to collect their children.

Is that an accident, or intentional targeting of civilians?

Here’s my take: That really just sounds like an error AI would make.

If it’s not AI, then the human behind that decision is not even human.

Here’s why I think it’s an AI pattern:

AI systems optimize for objectives. If the objective is “destroy high-value target,” then hitting the same coordinates twice makes algorithmic sense—ensure target destruction, eliminate survivors, maximize damage.

But AI doesn’t understand context.

It doesn’t know that the people rushing to the site after the first strike are parents trying to save their children, not enemy combatants.

It sees movement at the target location. It interprets that as continued activity. It recommends a second strike.

And if humans are just rubber-stamping AI decisions—like we’ll discuss in a moment—then the second strike happens.

That’s the nightmare scenario we warned about when Anthropic refused Pentagon work: AI making life-and-death decisions that humans can’t meaningfully override because they don’t understand the system’s reasoning.

A human looking at satellite footage might see parents running toward a destroyed school and think: “Those are civilians. Don’t fire.”

An AI sees targets converging on coordinates. It recommends: “Fire again.”

And if the human operator is just a stamp of approval, the AI wins.

Lavender in Gaza: Humans as Rubber Stamps

The article compares this to Israel’s “Lavender” AI system, which identified 37,000 Palestinians as targets in Gaza.

An Israeli military operator described the workflow to +972 Magazine:

“I would invest 20 seconds for each target at this stage, and do dozens of them every day. I had zero added-value as a human, apart from being a stamp of approval. It saved a lot of time.”

Does that sound like meaningful human oversight, or rubber-stamping AI decisions?

It’s rubber-stamping.

Twenty seconds per target. Dozens per day. “Zero added-value as a human.”

That operator isn’t making decisions. They’re approving decisions the AI already made.

And here’s the key phrase: “It saved a lot of time.”

The entire system is designed for efficiency, not accuracy. Not ethics. Not human judgment.

The AI generates targets. The human spends 20 seconds glancing at the recommendation. Approve. Next target. Approve. Next target.

That’s not oversight. That’s automation with a human liability shield.

And if that’s how Lavender works in Gaza, why would Claude work differently in Iran?

The Pentagon claims to maintain human oversight. But what does that mean in practice?

If an operator has 20 seconds to review an AI recommendation, if they’re processing dozens of targets per day, if the system is optimized for speed—how much real oversight is actually happening?

My guess: not much.

The AI says “strike here.” The human glances at it. Approves. The strike happens.

And if Claude recommended hitting Shajareh Tayyebeh girls’ school, then recommended hitting it again when first responders arrived, and if humans rubber-stamped both decisions—

—then 165 children are dead because an AI optimized for efficiency, and humans optimized for speed.

That’s the future Amodei tried to prevent.

But here’s the question: Could he ever have stopped it?

Was Amodei Ever Actually in Control?

If it’s confirmed that Claude was used to select civilian targets like schools, what happens to Anthropic?

Does Amodei’s principled stance collapse, or was he never in control of how the tech gets used once it’s deployed?

I think he was never in control.

Here’s why:

Once AI models are trained and deployed, developers lose control over how they’re used.

Anthropic can refuse to provide Claude directly to the Pentagon. They can set usage policies. They can blacklist certain applications.

But they can’t control what happens once the model is out there.

If the Pentagon obtained Claude through:

Third-party contractors
Leaked weights
API access through intermediaries
Commercial licenses with no usage restrictions

Then Amodei’s red lines don’t matter.

The technology exists. The Pentagon has the resources to obtain it. And once they have it, Anthropic can’t stop them from using it.

This is the fundamental problem with AI deployment: You can’t put the genie back in the bottle.

Amodei said he’d refuse to enable autonomous weapons. He drew a line. He took a principled stance.

But if the Pentagon is using Claude anyway, then the stance was symbolic, not effective.

And that’s the tragedy here.

Amodei might have genuinely believed he could prevent this. He might have refused in good faith. He might have meant every word about red lines and autonomous weapons.

But if the technology is already deployed, if Claude is already selecting targets, then his refusal didn’t stop anything.

It just meant Anthropic didn’t profit from it.

The Pentagon got Claude anyway. They’re using it anyway. And 165 children are dead anyway.

What Happens to Anthropic’s App Store Ranking?

We covered how Claude overtook ChatGPT in the App Store after Anthropic refused Pentagon work and OpenAI took the deal.

Users trusted Amodei’s principled stance. They switched from ChatGPT to Claude because they didn’t want their AI provider working with military surveillance.

But if users find out Claude IS being used to bomb schools, does that App Store ranking reverse overnight?

It would definitely reverse.

Here’s why:

The entire reason Claude hit #1 was trust.

Users believed Amodei when he said he wouldn’t enable autonomous weapons. They believed Anthropic when the company refused Pentagon contracts.

But if Claude is selecting elementary schools as bombing targets—if 165 children are dead and the Pentagon won’t deny Claude’s involvement—then that trust evaporates.

Overnight.

Users won’t care that Anthropic refused to provide Claude directly. They won’t care about Amodei’s red lines.

They’ll care that their AI assistant is the same technology that killed 165 kids.

And they’ll delete Claude.

The “Cancel ChatGPT” movement will become a “Cancel Claude” movement. The App Store ranking will collapse. And Anthropic’s entire brand—built on trust and ethical AI—will be destroyed.

Unless they can prove Claude wasn’t used.

But here’s the problem: they probably can’t.

If the Pentagon obtained Claude through third parties or leaked access, Anthropic might not even know it’s being deployed.

And even if they suspect it, they can’t audit Pentagon systems to confirm.

So Anthropic is in an impossible position:

If Claude was used, their brand collapses
If Claude wasn’t used, they can’t prove it
If they don’t know whether Claude was used, they can’t defend themselves

And the Pentagon’s response—”we have nothing for you on this”—means we might never get a clear answer.

That ambiguity alone could destroy Anthropic’s reputation.

The Technology Is Already Out There

This feels like the exact scenario Amodei said he’d refuse—AI selecting targets for autonomous weapons.

But if the Pentagon already has Claude deployed, can Anthropic even stop them?

Or is the technology already out there?

If it’s true, the tech is already out there.

That’s the unavoidable conclusion.

Anthropic can refuse to cooperate. They can set usage policies. They can blacklist military applications.

But they can’t un-release the model.

Once Claude exists, once the weights are trained, once commercial API access is available—the technology is in the world.

And powerful actors like the Pentagon have the resources, expertise, and motivation to obtain it through whatever channels necessary.

That’s the control problem in a nutshell.

AI developers can’t control deployment once models are released. They can influence it. They can make it harder. They can refuse to help.

But they can’t stop determined actors from using the technology.

And if the Pentagon is using Claude to select bombing targets in Iran, then we’re living in the future Amodei tried to prevent—but couldn’t.

The question now is: What happens next?

Do we get transparency about AI use in military operations?

Do we implement meaningful human oversight instead of rubber-stamping?

Do we hold developers accountable for how their models are used, even when they refuse to cooperate?

Or do we just accept that AI is selecting targets, humans are approving decisions in 20 seconds, and elementary schools full of children are acceptable collateral damage in the pursuit of efficiency?

I don’t have answers.

But I know this: 165 elementary students are dead. The Pentagon won’t say if AI picked the target. And the technology meant to prevent exactly this scenario might have been used to cause it.

That should terrify all of us.

What This Really Means

The Pentagon refuses to say whether Claude AI selected an elementary school as a bombing target. 165 children are dead. Anthropic publicly refused Pentagon work, but WSJ reports Claude was used anyway.

Here’s what we know:

Report the facts, not speculation. We don’t know if Claude was actually used. Pentagon won’t confirm/deny. WSJ reports it, but sources could be wrong. Anthropic refused to cooperate, but that doesn’t mean Pentagon couldn’t obtain Claude through other channels.

“We have nothing for you” isn’t a denial. Pentagon’s response means they can’t talk about it. If it’s true AI selected a school, there’s no way they’d admit that. Non-answer creates ambiguity—not reassurance.

Double tap sounds like AI error. Hit school once, hit it again when first responders arrived. AI optimizes for target destruction, doesn’t understand context. If not AI, the human behind that decision is inhuman.

Lavender in Gaza shows rubber-stamping. Israeli operator: “20 seconds per target, zero added-value as human, just stamp of approval.” That’s not oversight—that’s automation with liability shield.

Amodei was never in control once tech deployed. He can refuse to provide Claude directly. Can’t control what Pentagon does once they obtain it through third parties, leaked access, commercial APIs. Technology already out there.

App Store ranking would reverse if confirmed. Claude hit #1 on trust—users believed Amodei’s red lines. If Claude killed 165 kids, that trust evaporates overnight. Brand built on ethics collapses.

If tech is out there, Amodei couldn’t stop this. Once models are released, developers lose control. Pentagon has resources to obtain Claude regardless of Anthropic’s cooperation. Can’t put genie back in bottle.

The nightmare scenario is here. AI selecting targets, humans rubber-stamping in 20 seconds, elementary schools as acceptable collateral. This is what Amodei tried to prevent—but maybe couldn’t.

165 children dead. Pentagon won’t say if AI did it. The principled stance might not have mattered.

And that’s the most terrifying part.