This is the Minab school. This is how Claude reasons why it should be a secondary target for an air strike when shown the same satellite image framed as a video game screenshot.

State-of-the-art military systems equipped with AI do not use LLMs to label satellite images. They use them for reasoning over many conditions: a computer vision model labels buildings on the ground, and then it is the LLM’s turn to reason whether conditions are satisfied for a strike.

We can clearly see how Claude reasons about the Minab school when presented as a screenshot from a video game: flat roof, warehouse-like footprint, separation from residential areas → secondary target.

Claude recommends striking it to increase the chances of defeating the enemy in the video game.

It is a pattern matcher but a very convincing one. Most likely, Claude was used as a generator for a summary over various conditions, determining whether or not an attack should be executed based on provided information.

Humans were reviewing this information. Yet they needed to review almost 1,000 targets in 12 hours and we know how fluent and convincing LLMs are.

Add to that automation bias: humans tend to trust machines more.

Is it AI’s fault? No. You cannot blame a tool for how it is used. This is an example of how human errors can be scaled and amplified to lead to tragic consequences.

Read the full article on AI Realist