Vals found that commercial AI models flagged civilians and a school in targeting tests, raising questions for IT teams that rely on AI to interpret images.