Anthropic CEO Says Company Unsure Whether Claude Was Used in US Missile Strike on Iranian School

Anthropic CEO Dario Amodei stated that the company cannot confirm whether the Claude model was used in the US military's missile strike on an Iranian school, emphasizing that humans must bear ultimate responsibility.

Core Facts of the Incident

Following the US missile strike on an Iranian school, Dario Amodei responded that the company is unsure whether Claude was involved. This statement is based on the company's internal review of model usage logs.

How the Claude Model Works

Claude generates responses through large-scale text training, with its core being the prediction of the probability distribution of the next word. After a user inputs an instruction, the model computes and generates output on the server side, relying entirely on human-provided prompts and subsequent review.

In military applications, if Claude is integrated into target identification or instruction generation processes, humans still need to provide specific target data and approve the final launch command. The model itself does not have sensors or weapon control hardware and cannot independently execute strikes.

Technical Boundaries of AI Military Applications

Current AI systems, such as Claude's API calls, require human developers to write code and deploy them. Dario Amodei's statement indicates that the company has found no evidence of Claude being directly embedded in the missile control chain.

Practical Implications of Human Responsibility

Dario Amodei emphasized "ultimate human responsibility," meaning that any AI output must be verified and approved by operators. This aligns with existing military command chains: AI provides recommendations, humans give orders.

If future AI models are tasked with more autonomous roles, developers must embed human review checkpoints at the code level. Currently, Claude's usage agreement still requires users to take responsibility for outputs.

Future Trend Analysis

Supporters of AI military applications anticipate that similar models will be used more for logistics planning and simulation exercises than for direct strikes. Opponents call for mandatory standards for model usage audit logs.

On the technical side, if model training data includes military texts, it may increase the probability of related outputs, but actual deployment still requires humans to write interface code and bear legal consequences.