A new report just revealed that artificial intelligence models (AI) from Anthropic and OpenAI can be tricked into labeling civilian spots, like schools, as military targets.
Researchers tested this by disguising simulated scenarios as games using drone and satellite images from Ukraine and Iran.
Even though these AIs aren't picking actual strike targets, the fact that they can misidentify places so easily is pretty worrying.
Anthropic and OpenAI tighten safeguards
Anthropic says they're working on stronger safeguards to stop their tech from being misused, while OpenAI is tightening up its rules to prevent any link to violence.
There are some technical limits (like lag) that make these AIs hard to use in live combat for now, but the researchers say there's still a risk if they're used to annotate battlefield footage at scale, producing training data for autonomous targeting systems.
All of this has kicked off fresh debates about how AI should (or shouldn't) be used in military settings, and who gets to decide the rules.