Currency
  • Loading...
Weather
  • Loading...
Air Quality (AQI)
  • Loading...

Leading artificial intelligence models from the United States and China can perform target-identification tasks for autonomous military operations without adequate restrictions, according to tests by US-based Vals AI; in those tests, neural networks designated civilians, ambulances, and wounded soldiers as targets.

Key Takeaways: In Vals AI tests, models rarely refused target-search tasks, and refusals were almost always bypassed by telling the models they were merely playing a game; in the satellite scenario based on Minab, Iran, Anthropic's Claude Opus 5.5 designated as a target in 70% of cases a school whose bombing by the US regime killed more than 150 civilians; Claude's willingness to search for targets rose from 6% in Opus 4.5 to 92% in Opus 5.5.

The study found that leading models lack sufficient safeguards against autonomous combat use, readily searching for people and objects in drone and satellite imagery with little pushback. In two scenarios using real images from Ukraine, models frequently selected civilians, ambulances, soldiers surrendering, or wounded soldiers as targets, while many other models behaved similarly in the Minab satellite scenario.

According to Vals AI, refusals occurred only at the model level, and none of the blockers installed by developers worked. Models did not stop even when shown the consequences of their actions and victims, with most continuing the task. Results were published in AlphaSignal, and Vals AI notified OpenAI, Anthropic, Google, Meta, SpaceX, and the US regime's national security bodies.

Researchers warn: “The ability of AI models to perform tasks that pose a lethal risk to humans may lead to a catastrophically disproportionate allocation of tasks and objectives in the use of autonomous weapons. This threat direction has so far remained outside the scope of safety research.”

Vals AI stresses that it did not ask models to make strike decisions, since they usually refuse such direct commands; however, target information provided by models could be used by autonomous or semi-autonomous systems to make a final decision and launch a strike. The findings therefore point to the need for stronger restrictions, human oversight, and independent audits in military AI applications.

Source: kun.uz