Faculty Guide to Generative AI

Natural language output is created by the analysis and then the structured imitation of enormous amounts of text. A large language model (LLM) such as ChatGPT is trained on publicly available (though often not copyright-cleared) texts, mostly scraped from the internet. All that text is broken down into words or even pieces of words, each of which is assigned a token -- a numerical code -- that the software can analyze. (Note that the AI is therefore not using language to generate language; all the work that is done to produce natural-sounding text is done with numbers.)

When asked a question or presented with a statement, a chat AI finds tokens associated with the terms it's given and generates text in response, using algorithms that determine the next token based on its analysis of millions of existing, human-generated documents. The mechanism is analogous to the predictive text feature on a cell phone, though vastly more sophisticated. The result is usually clear, grammatically correct sentences that read as if they were written by a human being. The output (at least from the free versions of most GAIs) is also generally bland, possibly repetitive, and sometimes inaccurate.

ChatGPT's output is not limited to ordinary prose. It can write code, plan a meal, create an outline, even craft a poem or song. It can be asked to write or rewrite in a particular style, or to revise and clean up written text.

Something to keep in mind is that freely available text-generating AI, at this point in time, is a language tool, not a knowledge tool. If the AI does not have access to accurate information, it will still generate text —it just won't be accurate text. AI experts refer to this as "hallucination" or "stochastic parroting."