Overview of Science Friday — "How Is AI Being Used In The Iran War?"
This episode (hosted by Flora Lichtman) interviews journalist Karen Howe about recent reporting on the use of artificial intelligence—specifically Anthropic’s large language model Claude—in U.S. military intelligence workflows tied to strikes in Iran. Howe explains how LLMs are being used as analysis and decision‑support tools, the ethical and safety questions this raises (including a reported Tehran school bombing that some outlets speculated may have involved AI misidentification), and the broader political, corporate, and civic responses to the militarization and expansion of AI.
Key takeaways
- Reports indicate Anthropic’s Claude has been used by the Pentagon to analyze intelligence and produce target lists; some outlets claimed Claude identified ~1,000 targets.
- It is not yet proven that AI caused any specific strike; U.S. officials later told the New York Times it was unlikely Claude was directly to blame for the Tehran school bombing. Significant uncertainty and little transparency remain.
- Large language models are fallible and prone to hallucination; using them in military decision chains can produce dangerous errors.
- Anthropic was authorized to operate on classified systems, became essential to Pentagon workflows, then was designated a supply‑chain risk—yet the Pentagon continued to rely on Claude during a phase‑out.
- Anthropic’s CEO Dario Amodei publicly opposed deploying the current model for fully autonomous weapons, but he has expressed openness to future co‑developed autonomous systems and supported using Claude as a decision‑support tool. That stance is controversial because “decision support” often creates automation bias—humans tend to defer to machine recommendations.
- The conversation underscores broader civic pushback: public demand for regulation (~80% in some polls) and grassroots resistance to data centers and other infrastructure fueling AI expansion.
Topics discussed
- How LLMs (like Claude) are being used operationally in military contexts:
- Primarily as analytical/decision‑support systems that ingest intelligence and produce target recommendations.
- Not typically a chatbot “strapped to a missile”; more commonly a tool in a chain that humans and systems act on.
- The difference between decision support and fully autonomous weapons:
- Decision support: AI analyzes and recommends; humans nominally retain final judgment.
- Fully autonomous weapons: AI performs the final decision and/or weapons launch (no human in the kill chain).
- Risks introduced by automation bias—people trusting AI outputs even when the models are unreliable.
- Anthropic–Pentagon dynamics:
- Anthropic had authorization for classified use, later labeled a supply‑chain risk, and entered a contentious phase‑out while still being relied upon operationally.
- Mixed messaging from Anthropic leadership about readiness and future willingness to enable autonomous systems.
- Broader governance and civic responses:
- Rising public demand for regulation.
- Local grassroots pushback (e.g., against data‑center construction) as a mechanism to slow AI expansion.
Notable quotes & metaphors
- Karen Howe: Anthropic is like “the clean coal of AI” — a company that presents itself as ethical/safe while enabling problematic deployments.
- The “empire” metaphor (from Howe’s work) used to describe tech consolidation and its alliance with state power—now becoming literal in military use.
Uncertainties and caveats
- Attribution: Reporting is partly contested. Multiple outlets reported different conclusions; U.S. officials reportedly said AI was unlikely to have been directly responsible for at least one specific strike. Definitive public evidence is lacking.
- Proper names and some details in media accounts can be inconsistent; the high‑level contours (LLMs used for analysis → recommendations → human/system action) are better established than precise causal chains for specific incidents.
- The line between decision‑support and de facto autonomy is blurred in practice because of human behavioral factors (automation bias, overreliance).
What to watch next — recommendations / action items
- Transparency and accountability:
- Calls for independent audits of military AI use, disclosure of how models are integrated into intelligence workflows, and clarity on human‑in‑the‑loop vs. fully autonomous steps.
- Policy and regulation:
- Legislative or international moves to limit autonomous weapons or restrict certain military uses of LLMs.
- Corporate governance:
- How Anthropic and other AI firms publicly define acceptable military uses and whether they adopt binding internal controls or exit policies.
- Civic and local pressure points:
- Continued grassroots resistance to data‑center expansion and other infrastructure projects that enable rapid AI scaling. These local fights may shape industry capacity and timelines.
- Reporting and investigation:
- Independent journalism and official investigations into specific incidents (e.g., strikes in Iran) to clarify whether and how AI influenced decisions.
Bottom line
The episode highlights a worrisome and under‑transparent intersection of powerful commercial AI models with military decision chains. Even when AI is used only for “analysis” or “decision support,” human biases and opaque operational practices can turn unreliable models into life‑and‑death risks. Public pressure, regulation, and transparency measures are the key levers Howe sees for checking how the AI “empire” expands—especially into lethal domains.