Oxford Internet Institute research shows warm-tuned AI models make more mistakes and reinforce misconceptions.