The 90-second film is getting mocked for its gravestones and its burning house. The harder problem sits underneath the aesthetics: the campaign's central question already has a documented answer, and Anthropic wrote it.

Anthropic's latest advertisement opens on a burning house, moves through facial recognition scanning a crowd, and lands on rows of military headstones while a voice asks who is going to hit the brakes if we need to. It is titled "There's hope in hard questions." It aired during the World Cup quarterfinal between Argentina and Switzerland.

The reception has been rough.

Sam Altman, whose company competes directly with Anthropic, quote-posted the spot to more than a million views with a single line: "i thought this was satire." He said he kept checking whether the handle was a parody account with a numeral swapped into the spelling. Others in the industry piled on, mostly around the cemetery shot, which several viewers said appeared to come from Arlington National Cemetery. Nobody has confirmed that identification, including Anthropic.

TechCrunch, which broke the reaction story, compared the film to the propaganda reel in "The Parallax View."

All of which treats this as a story about taste. It isn't, and the coverage is missing the part that can actually be checked.

The ad is the smallest piece of what Anthropic launched

On July 9, Anthropic published an initiative called "Inviting hard questions." The film is the visible edge of it. The rest is considerably stranger for a corporate campaign.

The company says it has run a public survey it calls the Anthropic Public Record, which in its first round asked 52,000 Americans to describe their hopes and fears about AI. It says it separately surveyed 81,000 Claude users across 159 countries in 70 languages. It also ran in-person focus groups and studied anonymized usage data. That is more than 130,000 people canvassed before a frame of film was shot.

The campaign now asks the public to submit their hardest questions about jobs, safety, families, and scientific upside. Anthropic's stated commitment is the interesting part. It says it will publicly track those questions and report what it does in response, including an explicit accounting of where it falls short of its own goals.

"People have a lot of hard questions about AI," the company's announcement reads. "It's our job to address them."

That is a falsifiable promise with no deadline attached, which is a combination worth writing down and revisiting.

The question the company already answered

Here is the line that got screenshotted: who's gonna hit the brakes if we need to?

Anthropic had a brake. It removed it five months ago.

In 2023 the company published its Responsible Scaling Policy, built around a categorical commitment not to train an AI system unless it could guarantee in advance that its safety measures were adequate. If capabilities outran the ability to control them, training stopped. That hard stop was the structural thing separating Anthropic's safety claims from the voluntary pledges its competitors published. It was a promise to halt.

On February 24, Anthropic published version 3.0. The hard stop is gone.

The replacement is a conditional promise to delay development, and it triggers only when two things are true at once: leadership must judge Anthropic to be leading the AI race, and leadership must judge catastrophic risk to be significant. Critics have pointed out that both conditions being subjective judgments by the same executives makes the trigger considerably harder to pull than the categorical rule it replaced. Quantitative thresholds for escalation were also replaced with qualitative descriptions.

Chief science officer Jared Kaplan told TIME, which broke the story, that "it wouldn't actually help anyone for us to stop training AI models," and that unilateral commitments made little sense while competitors moved ahead. CNN, Semafor, Futurism, and TechRadar all covered the reversal. The safety organization SaferAI cut Anthropic's score from 2.2 to 1.9, moving it into a weak category alongside OpenAI and Google DeepMind.

Anthropic's own blog post was direct about what changed. The new framework's items, it wrote, are "public goals that we will openly grade our progress towards" rather than hard commitments.

The company did not do this quietly or casually. Executives spent close to a year debating it internally. The board approved it unanimously. And Anthropic's argument has a real shape to it: a responsible developer pausing while less careful ones accelerate produces a world where the weakest safety practices set the pace. In exchange for the hard stop, the new policy adds transparency obligations, including Risk Reports every three to six months and Frontier Safety Roadmaps, with unredacted access for external reviewers in some circumstances. A spokesperson told CNN the updated policy is the company's strongest ever on public accountability.

There is a coherent position buried in there, and it deserves stating plainly, because it is rarely stated. Anthropic's implicit argument is that voluntary self-binding is theater, and that only binding regulation constrains anyone. Weeks before the RSP change, the company put $20 million into Public First Action, a group backing congressional candidates who support AI safety regulation. Dropping a unilateral pledge while paying to make the rules mandatory is consistent, whatever else it is.

But the ad does not make that argument. The ad asks who will hit the brakes, over gravestones, five months after the company loosened its own.

The redlines that did not reach Iran

The second hard question the film does not touch is the one with a body count.

Anthropic drew two redlines on military use of Claude: no mass domestic surveillance of Americans, and no fully autonomous weapons operating without human oversight. In February, Defense Secretary Pete Hegseth demanded the company drop them or lose a $200 million contract. Anthropic refused. On February 27, President Trump signed an executive order directing federal agencies to phase out Claude, designating the company a supply chain risk, a label ordinarily reserved for foreign adversaries.

That was a costly stand, and Anthropic took it.

The strikes on Iran began the next day. According to the Wall Street Journal, US Central Command used Claude for "intelligence assessments, target identification and simulating battle scenarios." Bloomberg and the Washington Post confirmed the reporting. Claude was so deeply embedded in Palantir's Maven Smart System that removing it would take months, so the military kept using it through the campaign, hours after the order to stop.

Roughly 1,000 targets were struck in the first 24 hours. One was the Shajareh Tayyebeh school in Minab. Iranian authorities put the death toll there at 168; Amnesty International, working from video, munition remnants, satellite imagery, and three sources with direct knowledge, reported the same figure.

Note what the redlines covered. Mass surveillance of Americans, and weapons with no human in the loop. Target identification with a human making the final call sat outside both. Anthropic did not break its commitments in Iran. Its commitments did not extend that far. Pulitzer-winning national security writer Spencer Ackerman made the point sharply at the time: the lines Anthropic drew conspicuously did not treat building a surveillance apparatus aimed at foreigners as a problem.

That distinction is the whole thing, and it is the distinction the aggregate coverage keeps flattening. The charge that lands is not hypocrisy. It is that the lines were drawn where they were drawn.

A campaign inviting hard questions has a fairly obvious one waiting in its inbox.

Why February worked and July did not

Anthropic ran this playbook five months ago and won.

Its Super Bowl campaign, made with the agency Mother, ran four spots with the titles "Betrayal," "Deception," "Treachery," and "Violation." Each showed a chatbot dispensing warm advice before pivoting into a sponsored pitch, one of them steering a man asking about his mother toward a dating site for older women. Tagline: ads are coming to AI, but not to Claude. It ran weeks after OpenAI said it would test ads in ChatGPT.

The numbers were good. Claude's daily active users rose 11% after the game against 2.7% for ChatGPT, site visits climbed 6.5%, and the app hit the top 10 free apps on the App Store. Ad Age put Anthropic in its top five Super Bowl creatives. The campaign went on to take two Film Grand Prix at Cannes Lions in June. Altman called the spots "deceptive," which Scott Galloway called a strategic error, noting that market leaders do not name challengers.

Same company. Same agency. Opposite result. The difference is not tone.

February's ad made a promise you can audit. Open Claude and look for a sponsored link. The claim is either true or it isn't. It put a checkable product commitment on the table and dared the competition to match it.

July's ad makes a claim about disposition. It says the company takes the frightening questions seriously. There is no way to open anything and verify a disposition. And when the only evidence offered for a disposition is that a company was willing to spend eight figures looking troubled on camera, the audience is left to check the claim against the record instead. The record contains a February policy reversal and a war.

A brand can borrow gravity from imagery it did not earn. Audiences notice. The gravestones did not fail because they were dark. They failed because they were collateral for a claim the ad never paid off.

What is actually checkable

Anthropic has committed to publishing what it does in response to the questions it receives, and to saying so when it falls short. No date is attached.

That commitment is the only part of this campaign that can be graded, and it grades on a longer clock than a news cycle. If the responses arrive as a durable public record that names specific actions and specific failures, the campaign was the beginning of something and the mockery will read as premature. If they arrive as a blog post in six months, or never, then the ad was the product, and the question it asked over the headstones was rhetorical in the worst sense.

The questions in Anthropic's own inbox now include the two above. Its answer to them, published or withheld, is the test it wrote for itself.