organizing science – not

Anyone who does research knows that publication of results is typically the end-state. While publication models for science have issues (another topic), most would agree that given the ability to quickly release papers online, the sheer volume of articles is daunting.

AI to the rescue! After all, machines scale in ways that humans can’t, so the idea was to train a language model and have it digest scientific papers to help “organize science.” The work, done by Meta, was released to the wild last week after training on 48M scientific articles. What resulted was a hot mess that spewed (often toxic) nonsense.

As usual, the capability was over-hyped, under-performed, and actually resulted in more pseudoscience and tripe than “organizing” rigorous scientific work. But this will not be the last attempt to get ML to parse scientific research – far from it. It does point to the need for Narrative Red Teaming as these tools become more ubiquitous, and the very real danger inherent in algorithmic bias and contextual ambiguity.

Share your thoughts