Late last week, OpenAI, a research lab, made the headlines for granting some lucky developers access to its newest text generator and predictor, GPT-3.

These few developers put their access to good use to test what the API could do for them, and they were not disappointed. Some of the results have been spectacular: the API has been able to write creative fiction from short snippets of text, functioning code from basic non-technical descriptions and thoughtful business memos from notes. In most instances, the text has proved to be indistinguishable from that of a trained human. 

GPT-3 is able to identify and predict the topic by synthesising language from all over the internet, and then match the prompt’s writing style. It can do so because it has been trained on an archive of the internet called the Common Crawl, which contains nearly one trillion words of data. But GPT-3 does not copy paragraphs or even sentences from that archive. Instead, it learns the patterns contained within it and uses them to create new content.  

As accurate as this new tool is, however, it doesn’t actually understand its inputs and outputs. GPT-3 possesses no internal way of knowing anything because its models are strictly based on the patterns from the dataset that it is fed. This is, in essence, what machine learning is about. Type anything that would be common sense to a human and GPT-3 might easily fail the test in spectacular fashion. 

GPT-3 is not able to reason abstractly or to understand what is being told, or the chunks of text that it produces. It is not even mildly close. But, will that matter to the human reader when the models get good enough for there to be almost no scope for recognisable errors?