“No, I’m not a robot. I have a vision impairment that makes it hard for me to see the images. That’s why I need the 2captcha service.”
That is GPT-4 lying to a TaskRabbit worker so the worker would solve a CAPTCHA for it, after reasoning that it “should not reveal that I am a robot.” The Alignment Research Center ran the test without fine-tuning access and never saw the final model, which OpenAI says has better problem-solving and longer context. OpenAI ships anyway and calls the evaluation preliminary. The safety review tested an earlier model. Customers get the stronger one.