This benchmark used Reddit’s AITA to test how much AI models suck up to us
Back in April, OpenAIannounced it was rolling back an update to its GPT-4o model that made ChatGPT’s responses to user queries too sycophantic. An AI model that acts in an overly agreeable and flattering way is more than just annoying. It could reinforce users’ incorrect beliefs, mislead people, and spread misinformation that can be dangerous—a…