Dr Heidy Khlaaf (هايدي خلاف)

Can Anthropic’s AI Claude be trusted in combat?| The Take YouTube video by Al Jazeera English

It was great to join @aljazeera.com's podcast "The Take" to discuss the details of the DoW's use of Claude in Iran, as well as the stand-off between DoW and Anthropic that was largely safety theatre.
www.youtube.com/watch?v=skyI...

In this Tech Policy piece, I criticize how framings of Anthropic’s & OpenAI’s negotiations with the US’s DoW overindex on myopic interpretations of human oversight, papering over what should be the real target of our scrutiny: that generative AI algorithms are a flawed and inaccurate technology.

Exactly.

Anthropic’s AI tool Claude central to U.S. campaign in Iran, amid a bitter feud Anthropic’s AI tool Claude is playing a key role in the U.S. military’s campaign in Iran, amid a bitter fight with the Pentagon over the terms of its use in war.

It’s egregious for the WaPo to describe speed as the advantage against Iran w/ Claude. When these systems are incredibly inaccurate, they may as well be enabling indiscriminate targeting (e.g. schools), which isn’t the strategic win they’re framing it as.

www.washingtonpost.com/technology/2...

The one question everyone should be asking after OpenAI’s deal with the Pentagon The US said we can’t afford to let a surveillance state like China win the AI race. Well...

Was happy to speak to Vox on OpenAI's alleged AWS guardrails. Besides current guardrails being trivial to bypass, they can't enforce human oversight over the outputs of an AI algorithm. It's an operational matter not a technical one, and thus infeasible by any guardrails.
www.vox.com/future-perfe...

Using foundation models in national security contexts may introduce unique concerns threatening human rights. For example, a government’s ability to train models on citizens’ data obtained through commercial data brokers that would otherwise need a warrant, court order, or subpoena to obtain may allow governments to further exercise coercive powers that are automated through AI decision-making [6]. Such use may subvert due process, exacerbated when inaccurate outputs inflict unjust harms on civilians. Appropriate interventions may include the extension of data minimization principles to include purpose limitations on the collection, processing, and transfer of personal data to third parties for intelligence purposes.

The Atlantic notes how the Pentagon wants to "analyze bulk data collected from Americans." From our "Mind the Gap" paper 2024, a snippet I have come back to what seems like dozens of time at this point.
www.theatlantic.com/technology/2...

(Ir-)Responsible by Design? Corporate Guardrails and the Governance of Military AI [Jessica Dorsey is an Assistant Professor of International Law at Utrecht University School of Law; Elke Schwarz is a Professor of Political Theory at Queen Mary University London; Ingvild Bode is …

The Anthropic-U.S. DoD public dispute continues. We have unpacked the latest updates & the safety, legal & ethical concerns in our @opiniojuris.bsky.social article

@jessicadorsey.bsky.social @elkeschwarz.bsky.social @profbode.bsky.social @ncrenic.bsky.social

opiniojuris.org/2026/03/02/i...

At least 63 girls killed in strike on school in southern Iran Eyewitness tells MEE girls aged between seven and 12 seen lying dead across their school

This account includes an eyewitness.

Also: “At least 85 people, almost all of them young girls, have been killed in an air strike on a primary school in southern Iran, the Iranian judiciary said.”

In case you’re just waking up, the U.S. has teamed up with Israel overnight to start an illegal war of regime change, apparently on a presidential whim with no involvement of Congress, and they are already committing horrific atrocities.

Every single outlet that uncritically recites Israel's self-designation of "preemptive strikes" should be permanently shut down for good. Straight up. Definitionally no such thing as a "preemptive" attack when the country being attacked had no plan nor indication of planning to attack first.

Google Workers Seek ‘Red Lines’ on Military A.I., Echoing Anthropic

I consider this a loss rather than a win, as just a few years ago the redline was any military use, now it’s the most extreme use case of LAWS. AI companies have successfully moved safety thresholds without effective internal pushback.

www.nytimes.com/2026/02/26/t...

I have to give Anthropic credit for recognizing that deploying unreliable AI in AWS is not strategic for the future of AI. But there's a very fine line between DSS and AWS in practice due to automation bias, if they don't believe it's reliable for the latter, it's not reliable for the former either.

Some real cognitive dissonance happening with takes saying "but Anthropic HAD to drop their safety measures, they're the good guys you see!" Anyway from our paper last year:

Exclusive: Pentagon clashes with Anthropic over military AI use, sources say The Pentagon is at odds with artificial-intelligence developer Anthropic over safeguards that would prevent the government from deploying its technology to target weapons autonomously and conduct U.S....

If flawed and inaccurate LLMs are instrumented in AWS by replacing humans for decision making, then "wars" may as well be indiscriminate lethal campaigns. Anthropic's position also isn't a moral high ground given their AI-DSS uses w/ Palantir, where automation bias may lead to similar outcomes.

Army using AI to update doctrine Leaders at the Combined Army Doctrine Directorate have started training authors on generative AI tools to speed up research and drafting.

Yikes.

www.militarytimes.com/news/your-mi...

There's a constant AI-washing of terms so these companies can claim they're solving a problem that doesn't exist with AI. Static analysis/formal methods also put forward suggestions, have they even used these tools?
Claude Code may also generate up to 90% insecure code (arxiv.org/pdf/2512.03262).

As a formal methods PhD, it's embarrassing for Anthropic to incorrectly describe static analysis in their Claude Code Security announcement. Security and formal methods engineers already have data "reasoning" tools, this isn't the bottleneck, false positives, which LLMs absolutely have, is.

Violent and non-violent death tolls for the Gaza conflict: new primary evidence from a population-representative field survey This first independent population survey of mortality in the Gaza Strip shows that violent deaths have substantially exceeded official figures whereas the demographic composition of casualties aligns ...

This new Lancet study would indicate that the death toll in Gaza is actually around 120,000 at minimum, if the undercount rate is similar since January 2025.

Just absolutely disgusting.

I really It's kinda weird how the "The left hates technology (because AI)" mostly come from people whose experience with tech is on the level of vibecoding and the "left" they criticize consists to a significant part of people who've been actually building and maintaining systems for decades.

"We present the first representative international data on firm-level AI use... firms report little impact of AI over the last 3 years, with over 80% of firms reporting no impact on either employment or productivity."

www.nber.org/papers/w34836

CBP Signs Clearview AI Deal to Use Face Recognition for ‘Tactical Targeting’ US Border Patrol intelligence units will gain access to a face recognition tool built on billions of images scraped from the internet.

NEW: CBP signs a new deal with Clearview AI to access its scraped image database for "tactical targeting," including efforts to “disrupt, degrade, and dismantle” networks of people labeled security threats.

Israel used weapons in Gaza that made thousands of Palestinians evaporate US-made thermal weapons burning at 3,500C caused 2,842 people to "evaporate" in Gaza, Al Jazeera investigation finds.

“Four of my children just evaporated,” Badran said, holding back tears. “I looked for them a million times. Not a piece was left. Where did they go?”

There are no words to adequately describe the evil here.

Banner image with screenshot of scientific article from nature Medicine, as well as two panels from the study method and results

⚠️ Despite all the hype, chatbots still make terrible doctors. Out today is the largest user study of language models for medical self-diagnosis. We found that chatbots provide inaccurate and inconsistent answers, and that people are better off using online searches or their own judgment.

As AI enters the operating room, reports arise of botched surgeries and misidentified body parts Medical device makers have been rushing to add AI to their products. While proponents say the new technology will revolutionize medicine, regulators are receiving a rising number of claims of patient ...

When AI was added to a tool for sinus surgery: “Cerebrospinal fluid leaked from one patient’s nose. In another… a surgeon mistakenly punctured the base of a patient’s skull. In two other cases, patients suffered strokes after a major artery was accidentally injured”

www.reuters.com/investigatio...

Commercial AI is rarely built to be fit for purpose, and yet they're used regardless in safety critical applications they're not intended for, becoming a tool to avoid accountability while enabling indiscriminate harm. Something we've also seen in defense.

This is also why many safety engineers (CC @forrealflygirl.bsky.social) emphasize that accident statistics are manipulated and skewed. These interventions are in fact not mandated to be reported.

A lot of people who regurgitate "improvement" talking points to discredit any criticism of AI have little to no awareness of the trickery that occurs behind the scenes to provide this mirage. Not only for autonomous vehicles, but the army of human labour used to fine tune generative AI.

I Infiltrated Moltbook, the AI-Only Social Network Where Humans Aren't Allowed I went undercover on Moltbook and loved role-playing as a conscious bot. But rather than a novel breakthrough, the AI-only site is a crude rehashing of sci-fi fantasies.

NEW: I went undercover, as a human poster, on the social network designed just for AI agents.

The site isn’t a novel breakthrough; it’s simply a crude rehashing of sci-fi fantasies for the hype machine’s true believers.

My latest for @wired.com:

www.wired.com/story/i-infi...

The Trump Administration exempts new nuclear reactors from environmental review The announcement comes just days after NPR revealed the administration had secretly rewritten safety and environmental standards.

It's been maddening to see officials claim these roll backs are due to the "inherent" safety of nuclear reactors. Reactors are actually not inherently safe, they're currently safe because we've made them safe through these very regulatory measures now being eliminated.
www.npr.org/2026/02/02/n...