Sophisticated AI models are more likely to lie
arstechnica.com
Human feedback to AIs makes them favor providing an answer, even a wrong one, while making the answer more convincing.