ChatGPT is one of the players in the natural language AI field. It is not the only one, but it is the one that is best known, following the attention generated by its November 2022 beta rollout. When asked to define itself, ChatGPT replied,
ChatGPT is a large language model developed by OpenAI that uses a type of artificial intelligence called deep learning to generate human-like text in response to natural language inputs. (https://chat.openai.com/chat, 2/17/2023)
In other words, it is a complex piece of software that can be asked questions in conversational language and produces replies that read as if they were written by a human being. It achieves this by inputting and analyzing enormous amounts of existing text (that's the "large language"), which enables it to mimic sentence structure and argument.
Generative AI can produce a wide variety of content.
The factor that determines how good an AI's creation will be is the quality of the prompt. Crafting AI prompts is a skill in itself, which will become ever more valuable as AI becomes more and more integrated with everyday life.
ChatGPT can produce an essay on a text, an explanation of a natural phenomenon, a review of a play, a meal plan, a vacation suggestion, or answers to a vast number of other questions. The quality of its output is highly suspect, however, and ethical questions arise from the way that AIs are trained and behave.
A major problem with ChatGPT at this point is that the dataset that informs its underlying structure comes in large part from the internet, and therefore reflects the biases and flaws of the internet. It can produce material that is false or misleading, and it can also sometimes use language that is racist, misogynistic, or otherwise offensive. ChatGPT can't check its data or, usually, look for new data; when it doesn't have an answer, it can make an answer up (AI experts refer to this as "hallucination" or "stochastic parroting"). Users have also found that, when the chatbot is asked for its sources, it can generate fake citations.
Additionally, ChatGPT lacks the ability to do real critical thinking, so if a user requests involved comparisons or analysis, it is likely to provide bland text without individuality.
Another issue is privacy. ChatGPT and other AIs use the questions and prompts that they're given as additional training; they may collect other kinds of data as well (ChatGPT, notably, requires users to give it a phone number). There is no information available about what AI companies do with the data they can collect, and their privacy policies raise some questions (see the Verge, linked below).
There are related questions concerning AI and copyright. Is using copyrighted material as training data for an AI a violation of copyright law? Should the tech companies making AI tools have to request permission to use that material? Should authors and artists receive compensation? Another messy area is who—if anyone—owns the material that an AI generates. (U.S. court decisions favor the idea that there must be a human creator involved, but how involved that human must be is being decided on a case-by-case basis.)
The New York Times: A.I. Bots Can't Report This Column. But They Can Improve It.
Gartner (consulting firm): Top Generative AI Questions
Fast Company: What is a large language model and how does it work?
The New York Times: When AI Chatbots Hallucinate
The Brookings Institute: How to address new privacy issues raised by artificial intelligence and machine learning
OpenAI, in collaboration with other AI labs: Best practices for deploying language models