Amos Toh

The dispute over the use of Claude in autonomous weapons also lays bare the dangers of lethal targeting without sufficient human oversight. The laws of war require the military to distinguish between combatants and civilians, and refrain from attacks that cause excessive civilian harm. These determinations are often context specific and may require judgment that AI is ill-equipped to exercise. The Defense Department’s directive on autonomous weapons raises more questions than it answers about how the military addresses these risks. It requires senior Pentagon leaders to review whether autonomous weapons enable “appropriate levels of human judgment over the use of force.” But it’s unclear how this standard is satisfied when the weapon leaves no room for commanders or operators to override technical blind spots in life-and-death decisions. Congress should urgently impose safeguards to align autonomous weapons with the laws of war — and restrict the use of weapons that fall short.

Emile Ayoub and Amos Toh: Senior Counsels, Liberty and National Security Program, Brennan Center for Justice ( @emileayoub.bsky.social and @amostoh.bsky.social):

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.

Key Questions on the Role of Technology in the Expanding Middle East War Tech Policy Press asked experts working at the intersection of technology policy, security, and international affairs to share what they are watching.

The expanding war in Iran brought to the fore questions about the role of technology in armed conflict, including the controversial use of new artificial intelligence technologies. Tech Policy Press invited perspectives from experts on what they are watching for as the situation unfolds.

As @amostoh.bsky.social and I explain, the Pentagon's dispute with Anthropic should not distract from the broader crisis at hand: Congress's failure to regulate some of the riskiest uses of AI - namely, amplifying surveillance and automating the use of lethal force.

Important 🧵on how the military’s use of Claude and other agentic AI could blow open the data broker loophole.

@emileayoub.bsky.social, @jlkoepke.bsky.social and @jakelaperruque.bsky.social are all experts on this - follow them!

Trump's Iran Strikes Are Unconstitutional Congress must act.

Trump's "massive and ongoing operation" against Iran is a clear violation of the Constitution's separation of powers.

In our democracy, the president does not have the kingly power to plunge the nation into war without democratic debate and sanction. /1 www.brennancenter.org/our-work/ana...

this is really good background and context

new details about the Pentagon's fight with Anthropic are coming in fast and still under dispute, but this excellent analysis from @rightsduff.bsky.social and @amostoh.bsky.social from friday is very much worth your time. bsky.app/profile/just...

My point is that the rhetoric of legal constraint, and debates about the legality of presidential uses of force, are empty. And they deflect attention from Congress’s constitutional responsibility to exercise its political judgment and the political powers that the framers undoubtedly gave it to question, to hold to account, and (should it so choose) to constrain presidential uses of force.

Jack Goldsmith is always a must-read, but IMHO he's being too cynical (or not cynical enough). Attention to Congress's constitutional responsibility (and its judgment, and powers) is energized by invoking illegality, not by brushing it off; this includes whether the president can steal the march.

DoDD 3000.09 *does not* prohibit fully autonomous weapons. It’s a framework for *authorizing* autonomous weapons, including weapons that do not require human confirmation between target ID and firing.

NSA surveillance sweeping up Americans’ communications overseas is also traceable to EO 12333:

NYT: Prime Minister Mark Carney of Canada and his foreign minister, Anita Anand, backed the American action. “Canada supports the United States acting to prevent Iran from obtaining a nuclear weapon and to prevent its regime from further threatening international peace and security,” they said in a joint statement.

so much for that davos speech

Last July, we warned that the administration’s policy against “woke AI” would heap pressure on the tech industry to quash dissent.

Today, it is pressuring industry to roll back safeguards aimed at preventing the mass surveillance of Americans, and lethal targeting without human oversight.

Experts raise questions and concerns about Pentagon’s threat to blacklist Anthropic amid AI spat Some warned that such a move would mark an extreme response that could have a chilling effect on the broader frontier AI industry.

Guardrails don’t undermine military AI. In fact, they're needed to improve the tech’s safety and reliability, says @amostoh.bsky.social defensescoop.com/2026/02/27/p...

Anthropic will not cave to the Pentagon’s demands.

The company’s restrictions on using Claude to spy on Americans and deploy fully autonomous weapons aren’t just its ethical red lines.

They are also what a good faith interpretation of the Fourth Amendment and the law of war require.

More in 🧵👇

clarifying timeline by @techpolicypress.bsky.social on the fast-moving dispute between Anthropic and the Pentagon on the limits of using AI in warfare

What the Defense Production Act Can and Can’t Do to Anthropic The legal answer depends on what the government is actually demanding—and the statute's ambiguities cut both ways.

The Pentagon is threatening to use the Defense Production Act against Anthropic. It's a terrible policy, but legally it depends on whether they're seeking to avoid Anthropic's terms of service or have Anthropic retrain Claude. My latest for @lawfaremedia.org. www.lawfaremedia.org/article/what...

The Pentagon is giving Anthropic until Friday to waive its restrictions on using Claude for mass surveillance of Americans and lethal targeting without human involvement, per @axios.com

My thread below breaks down how these uses could be unlawful or unconstitutional👇

Anthropic CEO Dario Amodei and Defense Secretary Pete Hegseth will meet tomorrow to discuss the terms of Claude's use, per @mccuri.bsky.social and Dave Lawler @axios.com.

www.axios.com/2026/02/23/h...

Here's what's at stake 👇

25/ Congress should also use its power to close the data broker loophole and strengthen privacy protections against AI-enabled surveillance. It should demand transparency about how the military is using AI in hostilities – a critical first step towards regulating the technology.

24/ Congress has the constitutional power and duty to regulate the military. It should reassert its authority over matters of war and peace.

23/ This appears designed to pressure Anthropic to waive its usage restrictions on Claude – restrictions that might reflect not only the company’s own policy preferences but also what U.S. and international law require.

Exclusive: Pentagon warns Anthropic will "pay a price" as feud escalates The Pentagon may label Anthropic a "supply chain risk," forcing all its vendors to sever ties.

22/ DOD has threatened to designate the company as a supply chain risk, which would not just cut it off from future military work but also chill its business with other companies that themselves have a stake in defense contracting.

www.axios.com/2026/02/16/a...

21/ Anthropic is no doubt aware – and perhaps increasingly wary – of the legal issues created by the Pentagon’s use of Claude in military operations. But relying on Anthropic – or any company – to restrain the military is a fool’s errand.

Opinion | How to Respond, Right Now, to Trump’s Boat Strikes

20/ All questions Congress must demand the Pentagon answer in its investigations of military actions in Venezuela, on the high seas, and elsewhere.

www.nytimes.com/2025/12/02/o...

19/ Was Claude, Maven or any other AI used to distinguish strike targets from civilian infrastructure, assess the civilian impact of strike options, or otherwise plan the attack?

For that matter, how is the military using AI in its boat strikes, and prep for possible strikes on Iran?

The Overlooked Deaths of the Attack on Venezuela To many on the ground, civilian fatalities were simply the cost of ousting Nicolás Maduro.

18/ The Venezuela attack struck several apartment buildings and killed two civilian air traffic controllers in a car leaving the Caracas international airport.

www.newyorker.com/news/the-led...

17/ But even if it wasn’t, any use of AI in military operations must comply with rules on the conduct of war – for example, those requiring the military to distinguish between combatants and civilians and to prevent excessive suffering.

media.defense.gov/2023/Jul/31/...

Attack in Venezuela Was Unconstitutional Congress must assert its control of military matters.

16/ Given that Congress did not authorize the Venezuela operation, and that Venezuela did not pose an imminent threat to the US, the attack was unconstitutional on its face. This makes any use of AI to facilitate the attack unlawful.

www.brennancenter.org/our-work/ana...

AI 'unchained': NGA's Maven tool 'significantly' decreasing time to targeting, agency chief says - Breaking Defense "NGA Maven is now available to all services and all combatant commands. There are 20,000 active users through more than 35 service and combat and command tools across three security domains," said NGA...

15/ Remember the public uproar in 2018 over Google’s involvement in Maven, the Pentagon’s AI pilot to analyze troves of drone and satellite footage? It’s now a full-fledged program the military has credited with speeding up targeting.

breakingdefense.com/2025/05/ai-u...

Exclusive | Pentagon Used Anthropic’s Claude in Maduro Venezuela Raid The use of the model through a contract with Palantir highlights the growing role of AI in the Pentagon.

14/ @wsj.com reported that the military used Claude in its attack on Venezuela and capture of Maduro.

But Claude is just the tip of the iceberg.

The last decade has seen the military increasingly rely on AI in combat.

www.wsj.com/politics/nat...