I think the most interesting finding in this study is the following:
The models also suffered from “contra-factual bias": They were likely to believe a false premise embedded in a user’s question, acting in a “sycophantic” way to reinforce the user’s mistake.
Which when you think about how language models work, makes a lot of sense. It’s drawing upon trained data sets that match the question being asked. It’s easy to lead it to respond a certain way, because people who talk pro/con certain issues will often use specific kinds of language (such as dog whistles in political issues).
I had a colleague perform a similar experiment on ChatGPT 3. He’s ecoanxious and was noticing how the model was getting gloomier and gloomier in accordance with him, so he tried something. Basically he asked something like “Why is (overpopulated specie) going instinct in (location)?” The model went on to list existential threats to a specie that is everything but endangered. Basically it naively gobbled the loaded question.
I think the most interesting finding in this study is the following:
Which when you think about how language models work, makes a lot of sense. It’s drawing upon trained data sets that match the question being asked. It’s easy to lead it to respond a certain way, because people who talk pro/con certain issues will often use specific kinds of language (such as dog whistles in political issues).
It might also be a side effect of being trained to “chat” with people. There’s a lot of work that goes into getting it to talk amicably with people.
I had a colleague perform a similar experiment on ChatGPT 3. He’s ecoanxious and was noticing how the model was getting gloomier and gloomier in accordance with him, so he tried something. Basically he asked something like “Why is (overpopulated specie) going instinct in (location)?” The model went on to list existential threats to a specie that is everything but endangered. Basically it naively gobbled the loaded question.