Science Friday

How Is AI Being Used In The Iran War?

March 12, 2026

Key Takeaways Copied to clipboard!

  • The military use of AI, specifically the deployment of Anthropic's Claude model to analyze intelligence and identify potential bombing targets in the conflict involving Iran, highlights a startling and literal fusion of Silicon Valley technology with military operations. 
  • Anthropic's CEO, Dario Amodei, expressed reservations about using the current iteration of Claude for fully autonomous weapons but was reportedly comfortable with its use as a decision support system for identifying bomb targets, a scenario experts warn is highly susceptible to dangerous automation bias. 
  • Optimism for accountability in the AI industry stems from growing public resistance, with 80% of Americans supporting regulation, evidenced by effective grassroots movements successfully pushing back against reckless data center expansion, which could serve as a model for checking other problematic AI deployments, including military use. 

Segments

AI Empire Metaphor Realized
Copied to clipboard!
(00:00:03)
  • Key Takeaway: The initial metaphor of AI companies consolidating power like an empire has become literal with the fusion of this technology and the military.
  • Summary: Host Flora Lichtman introduces journalist Karen Hao to discuss the military use of AI, noting recent headlines involving Anthropic and OpenAI. Hao reflects that her book’s metaphor of AI power consolidation as an ’empire’ was not intended to encompass the immediate alliance between Silicon Valley and Washington in military contexts. This fusion of technology and military application is now seen as a literal reality rather than just a conceptual framework.
Claude’s Role in Iran Conflict
Copied to clipboard!
(00:01:40)
  • Key Takeaway: Anthropic’s Claude AI model was reportedly used to analyze intelligence data and identify approximately a thousand potential bombing targets in the conflict involving Iran.
  • Summary: Reporting suggests Claude was used to analyze intelligence data, identifying around a thousand targets for bombing. This is deeply concerning because Large Language Models like Claude are inherently faulty and inaccurate, capable of making up details even in casual conversation. The potential for this inaccuracy to translate into fatal errors in military targeting, such as the speculated misidentification of civilian targets in bombings of a school and subsequent first responders, encapsulates the current risks.
Pentagon-Anthropic Conflict Details
Copied to clipboard!
(00:03:56)
  • Key Takeaway: The Pentagon declared Anthropic a supply chain risk, initiating a six-month phase-out, yet immediately used Claude for strikes in Tehran, highlighting a major contradiction in policy and reliance.
  • Summary: Anthropic initially gained Pentagon permission to use Claude on classified systems, but a dispute over usage details led the Pentagon to declare the company a national security risk. Despite this declaration and a mandated six-month phase-out, the US military reportedly used the very tool deemed a risk in the bombing of Tehran shortly thereafter. Furthermore, CEO Dario Amodei’s stance was complicated; while opposing autonomous weapons use for the current Claude iteration, he was reportedly fine with its use as a decision support system for target identification.
Automation Bias and Decision Support
Copied to clipboard!
(00:06:22)
  • Key Takeaway: Experts warn that using LLMs for decision support systems creates a significant automation bias, where human oversight is ineffective because operators inherently trust the computer’s analysis.
  • Summary: Dr. Heidi Claff argues that if an AI is deemed unsafe for autonomous weapons, it should also be restricted from decision support roles due to extensive research showing human automation bias. Humans tend to believe computer outputs, meaning a human checking targets identified by Claude is not a legitimate check, leading to scenarios exactly like the speculated errors in the Tehran bombings. This suggests Anthropic’s moral high ground regarding autonomous weapons is suspect when their preferred use case still invites critical failure.
Defining Autonomous Weapons
Copied to clipboard!
(00:08:50)
  • Key Takeaway: Fully autonomous weapons involve the final two stages of the kill chain—deciding and launching—occurring without human involvement, which Anthropic’s CEO stated they were not ready for.
  • Summary: The term ‘LLM-powered weapons’ is misleading; LLMs are currently used to analyze information and identify targets, which are then acted upon by humans launching missiles. Fully autonomous weapons involve the system automatically feeding the target list to the launch system or drones identifying and bombing targets without human intervention in the final decision/launch stages. Amodei indicated readiness to develop future iterations for full autonomy, provided a human remained in the loop for both decision and launch steps currently.
Optimism in Grassroots Resistance
Copied to clipboard!
(00:11:37)
  • Key Takeaway: The most optimistic development is the broad public resistance to the AI industry, with 80% of Americans supporting regulation, exemplified by successful local protests against data center expansion.
  • Summary: Lichtman expresses optimism based on the growing public resistance to the AI industry, noting that 80% of Americans now favor regulation. This sentiment has fueled effective grassroots movements, particularly against the reckless expansion of data centers, where communities pressure elected officials. Slowing data center development acts as a critical throttle on AI advancement, offering a template for pushing back against other problematic deployments, such as military use or copyright infringement.