Historians don’t have to compress their material as severely. Since history is notoriously a story of conflict, and our sources were interested participants, few people expect historians to represent all aspects of the past with one correctly balanced model. On the contrary, historical inquiry is usually about comparing perspectives. Machine learning is not the only way to do this, but it can help. For instance, researchers can measure differences of perspective by training multiple models on different publication venues or slices of the timeline.[6]

A willingness to find meaning in collective patterns may be especially necessary for disciplines that study the past. But this flexibility is not limited to scholars. The writers and artists who borrow language models for creative work likewise appreciate that their instructions to the model acquire meaning from a training corpus. The phrase “Unreal Engine,” for instance, encourages CLIP to select pictures with a consistent, cartoonified style. But this has nothing to do with the dictionary definition of “unreal.” It’s just a helpful side-effect of the fact that many pictures are captioned with the name of the game engine that produced them.

The argument that Bender et al. advance has two parts: first, that large language models pose social risks, and second, that they will turn out to be “misdirected research effort” anyway, since they pretend to perform “natural language understanding” but “do not have access to meaning” (615).

Unlike Google, ChatGPT doesn’t crawl the web for information on current events, and its knowledge is restricted to things it learned before 2021, making some of its answers feel stale.

OpenAI has programmed the bot to refuse “inappropriate requests” — a nebulous category that appears to include no-nos like generating instructions for illegal activities