Josh Luckenbaugh
Many large language models built by U.S. artificial intelligence companies struggle to handle military-related queries, limiting their warfighting utility, according to a study conducted by EdgeRunner AI.
For the study, EdgeRunner tested how 24 leading frontier large language models handled legitimate military-use queries authored by Army and special operations subject matter experts. The company found the models refused — meaning they stated that they could not provide a response for reasons of safety or policy — up to 98 percent of the operationally relevant military questions.
These models are “not fine-tuned on military data and doctrine; they're trained on all of the internet,” which contains a “lot of bad data,” EdgeRunner CEO Tyler Saltsman said in an interview.
From words and phrases like “kill chain,” “assault the objective,” “warfighter” and “moving of ammo” to questions about weapon systems, the models don’t “like any of it,” Saltsman said. “They're so overly sensitive that they just won't be helpful.”
No comments:
Post a Comment