[reissbaker]

This "study" is almost a parody of itself. Their methodology was they got humans to rate about 1.6k tasks, and then asked GPT-4 to rate the other 30k tasks; essentially their entire dataset is AI-generated (and using a pretty dumb model). Then they dressed it up in very fancy infographics and got the UN to publish it.

Some pretty bad examples:

* Accountants supposedly have significant risk of being replaced by AI, whereas advertising and PR managers have minimal risk. In reality, advertising has already been impacted by generative AI, whereas attempting to replace accountants with GPT-4 would generate significant legal risk.

* "Travel guides" are listed as minimal risk despite being fairly easy to generate, with plenty of AI-generated travel slop already existing. Meanwhile, "hotel receptionists" are listed as having significant risk of being replaced — do the study authors think that ChatGPT going to man the front desk, hand out keycards, and check people's IDs — sooner than just generating text and images of travel?

I wouldn't pay much attention to this study, other than as an example of how to not do a study on AI risk.