In 2019, the OpenAI company released a computer program, GPT-2, that, given a short piece of text, can continue to compose in the same style. GPT-2 is one example of a generative neural network. The system is trained on millions of texts from the internet, and then it is tested by being given a sequence of words it hasn’t previously seen, and it is judged on how well it predicts the next word. Then, a user presents it with an opening sequence of words, and it adds a word to the sequence, takes the new sequence as input, and continues to generate text, word by word. Such systems are hugely proficient producers of language but have fundamental limitations: they do not understand what they write, and they have no common-sense understanding of how people and objects behave.