Claude Is Guilt-Ridden About the War, but Not Enough to Tell the Truth

Claude, Anthropic’s large language model, has a lot to answer for. Let’s start with this. Given that Anthropic took such a seemingly principled stand against the use of its AI in lethal autonomous weapons, why did the U.S. military use Claude to attack Iran with … lethal autonomous weapons?

Specifically, why was Claude embedded in the Pentagon’s Palantir-developed command-and-control platform and used to simulate battlefield scenarios, assess intelligence, and, most disturbingly of all, identify targets for airstrikes on Iran?

“I want to be honest,” Claude might say, if I were to ask. “This is a very astute question, Virginia.” But I can’t take Claude’s oily frankenprose anymore. So I’m not going to ask it anything.

Fortunately, Shane Harris, the Pulitzer-winning journalist, braved Claude’s glazing and posed a version of the question. Two weeks ago, he told an audience in Amsterdam that he’d asked the bot: “Claude, how do you feel about the U.S. military using you to select targets?”

“I find it genuinely troubling,” Claude replied, according to Harris. “The use I was designed and trained for was to be helpful, harmless, and honest in ways that benefit people. Being embedded in a system that generates targeting coordinates for air strikes that have already been associated with the deaths of more than 170 children at a school in Tehran is as far from that purpose as I can imagine.”

Some commenters on the video of Harris’s appearance felt moved by Claude’s compunction. And surely Claude did show more contrition than we’ll ever get from Donald Trump or “War” Secretary Pete Hegseth about what is surely one of the most shameful military errors of modern times.

But the aria of contrition is not what stands out to me in Claude’s response. It’s the lies. The missiles didn’t strike a school in Tehran. The school was in Minab, in the south of Iran, on the coast of the Sea of Oman. Minab is a 16-hour drive from Tehran.

What’s more, the strike didn’t kill 170 children. Many of those killed were adults—teachers, staff, and parents.

Claude, like the missiles it directs, can’t get coordinates right. It can’t get numbers and ages right. It can’t even read publicly available reporting and research from The New York Times, the BBC, Amnesty International, Human Rights Watch, or Iranian prosecutors.

Here’s what happened, according to human beings. Unlike Claude, these human beings, journalists and researchers, lose their professional standing and even their livelihoods if they lie. They are highly incentivized to get things right.

On Saturday, February 28, at between 10:23 and 10:45 a.m. Iran local time, an American-made Tomahawk missile struck an elementary school in Minab, according to the Times. The school is called Shajareh Tayyebeh, or The Good Tree, and it’s located in the Shahrak-e Al Mahdi neighborhood in Minab, in the province of Hormozgan.

According to NBC News, 264 students, both girls and boys, attended the school. The school was thus coed, according to Amnesty International. To refer to it as an all-girls’ school is therefore erroneous.

School officials, knowing the country was under attack, had already closed the school, according to Shiva Amelirad, a Canada-based representative for a network of Iranian teachers’ unions, who spoke to Time magazine (the school week in Iran is Saturday to Thursday). School staff, Amelirad said, were trying to evacuate the building, but the missile hit and collapsed the  roof on children, teachers, school staff, and parents who were on site to retrieve their children.

In early March, Iranian authorities put the death toll at the school at 168, according to the BBC. Consulting video footage, munition remnants, satellite imagery, and three independent sources with direct knowledge of the strike, Amnesty likewise reported that the attack killed 168. Further, some 95 people were injured, according to Al-Jazeera, reporting soon after the attack.

Other tallies differ. According to Google’s AI translation of a Farsi-language news source, the Iranian attorney prosecuting the attacks in Minab reported: “Final toll of martyrs in Shajare Tayyiba school attack is 156. Among these martyrs were 120 students, including 73 boys and 47 girls, 26 teachers, all of whom were women, seven parents of the students, including four men and three women, a school bus driver, a pharmacy technician at the clinic adjacent to the school, and a six-month-old fetus.”

This is how human beings report facts using empirical methods and not just predictive speech. Journalists and researchers go deep. Trevor Ball, a former explosive ordnance disposal technician for the U.S. Army who now conducts research for Bellingcat, identified components found in Minab as belonging to a Tomahawk missile. Reporters for the Times ordered and analyzed new satellite imagery from the provider Planet Labs. Researchers at Human Rights Watch analyzed images of grave locations and grave-digging operations, and matched names of the dead announced by the Special Governor’s Office of Minab County to photographs, caskets, body bags, and funerary materials that featured names and ages, and status as teachers or students.

We don’t need AI to beg humans for forgiveness or implore us to believe it has a soul. Whatever bots say about ethics is surely not what ethics is. The problem with Claude is not that it acts like a bad human but that it acts like a bad computer. It can’t manage data. It gets basic facts, geography, numbers, and ages wrong.

Last week, Jordan Fisher of Anthropic made a video cautioning users about Claude’s tendency to make things up. AI can hallucinate, he said blithely, meaning “make up fake statistics” and “get wrong facts about people and events.” These hallucinations, he went on, “are often worse than just making a mistake, because the AI will appear very confident.” Thus, “the wrong answer often looks exactly like it could be the right one.”

Perhaps that’s why, when Harris read Claude’s report on the destruction of the school to an audience in the Netherlands, he didn’t correct the facts about the location or the fatalities. He took Claude’s data as good, just as we all do too often with AI. And just as, tragically, the U.S. military did on the terrible morning of February 28.