Public transcript

Anthropic’s CEO explains why he took on the Pentagon

By: The Economist
Length: 7 min
Language: English
Text from: the video's captions
Transcribed: 13 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

This is an interview with Anthropic's CEO about his dispute with the Pentagon over a new contract for military use of AI models. He explains that his objections concern reliability of fully autonomous weapons and oversight of AI-driven forces, not any specific administration, policy, or past military operations. He argues that AI raises fundamental questions about the balance of power between private companies and governments, and that both concentrated corporate and concentrated governmental control of the technology pose risks.

Key points

Anthropic previously operated under a substantially more limited contract with the Department of War and had no concerns about how it was used on the ground, including in Iran and Venezuela.
The CEO says the dispute is about how the technology should be used and governed going forward, not about any administration, policy, or specific military operations.
He compares his objection to a supplier of aircraft saying the aircraft is not safe to fly above a certain altitude or turn at a certain speed.
He says he agrees with the Pentagon on 99% of issues and that there is more agreement than disagreement.
His first objection to fully autonomous weapons is reliability: the systems were not made or manufactured to be safe for that use case.
His second objection is oversight: an army of 10 million AI-driven drones could put control in the hands of a very small number of people or one person.
He believes AI is powerful enough to make private actors more powerful than ever, and also to vest governments, democratic or autocratic, with unprecedented power.
He concludes that the technology is too powerful to be in the hands of a few private companies and too powerful to be in the hands of government.
chapters chapters: []
notes_md notes_md:

Questions it answers

00:00Why limit autonomous weapons

Why does Anthropic's CEO object to fully autonomous use of its models in weapons?

He says it is not about democratic values but about the technology not being ready. Like an aircraft not made safe to fly above a certain altitude, the models were not manufactured to be safe for fully autonomous use.

The first objection is reliability of the technology for that use case.
The models were not built to be safe for fully autonomous weapons.

00:43The existing contract

Were there problems with how Anthropic's model was already being used by the military?

No. Anthropic operated under a substantially more limited contract and never ran into problems with it, including its uses in Venezuela and Iran. The dispute was about negotiating a new, broader contract.

The more limited contract caused no concerns on the ground.
The disagreement is about future use and governance, not past actions.

02:09Who decides use cases

Does Anthropic think it should control where its service is used militarily?

He says a service provider doesn't get to decide where the service is used operationally, and he expresses no opinion on military or government policy. Anthropic's say is over very broad use cases in general, not specific operations.

Operational military decisions remain with the government.
Anthropic's concern is broad use cases and governance going forward.

03:35Agreement with the Pentagon

How much does Anthropic actually disagree with the Pentagon?

He says they agree on 99%, including the contract they operated on together and a broader set of things they have not even done yet. He wants the Department of War to understand there is more agreement than disagreement.

He frames the dispute as narrow relative to overall cooperation.
The Pentagon's view is that suppliers should not capriciously constrain operational use.

03:38Nationalisation risk

Is there a risk of AI companies working with the Pentagon being effectively nationalised?

He says the coming years will require a discussion about the role of AI and government, because AI has implications at the level of all humanity, not just contests between nations.

AI raises questions of individual destructive capability and model risk.
The issue spans future administrations, Congresses, and geopolitical situations.

05:05The dual dilemma of power

Why is AI too powerful for both companies and governments?

AI can make private actors more powerful than ever, and it can vest governments, democratic or autocratic, with unprecedented power. He has long worried about both risks.

He has been criticised for suggesting private actors could become very powerful.
He is equally worried about unprecedented governmental power from AI.

05:50Oversight and drones

What is the oversight problem with an army of AI drones?

Human soldiers have norms and could refuse crazy orders, but the norms of 10 million AI-driven drones are unclear. Handled wrongly, one person could control those drones, and we don't want companies more powerful than government or government so powerful it can't be stopped.

The second objection to autonomous weapons is oversight.
Both concentrated corporate power and concentrated governmental power are problems at once.

Notes

Background to the dispute

Anthropic had a substantially more limited contract with the Department of War covering a limited number of use cases, possibly still in force.
The company never had concerns about how the model was used under that contract, including in Venezuela and Iran.
The dispute was about negotiating a new contract and how the technology should be used and governed going forward.

The supplier analogy

The CEO compares his position to a supplier of aircraft saying the aircraft is not safe to fly above a certain altitude or turn at a certain speed.
The models were not made or manufactured to be safe for fully autonomous use in weapons.

Agreement with the Pentagon

He says he agrees with the Pentagon on 99% of issues, including the existing contract and a broader set of things not yet done.
He is not expressing an opinion on military or government policy, and the disagreement is not about any administration.
If you provide a service, you don't get to decide where it is used operationally; Anthropic's say is over broad use cases.

Two objections to fully autonomous weapons

Reliability: The systems were not built to be safe for this use case.
Oversight: Human soldiers follow orders but have norms and could refuse crazy orders; an army of 10 million AI-driven drones would have unclear norms.
Handled wrongly, a very small number of people or one person could have their hand on the button controlling 10 million drones.

The bigger picture

AI has implications at the level of all humanity: national security, individual destructive capability, model risk, and economics.
The issue spans future administrations, different Congresses, and different geopolitical situations.
AI can make private actors more powerful than ever, and can vest governments, democratic or autocratic, with unprecedented power.
The dilemma: we don't want companies more powerful than government, but we also don't want government so powerful it can't be stopped.

Text from the video's own captions. Summary and notes written by AI from the transcript, so check anything important against the recording.