This page provides some background information on ChatGPT and AI, most of which will be familiar to you from the SPH Canvas module. Make sure to follow the guidance for individual classes on using these tools.
Generative artificial intelligence (AI) is a technology that is developing quickly. As with the internet in general, AI tools, like ChatGPT or Google's Bard, are neither good nor bad when it comes to finding information. Instead, they offer a new way in which we can interact with information. ChatGPT is just one example of a large learning model, meaning that it was developed and trained using large amounts of content available online. There are two important points to understand.
1) ChatGPT and other generative AI are not search engines, like Google or Bing, but rather they “create” information and answers based on the immense amounts of information contained in their databases. They do not point to other web pages or information outside of their databases.
2) ChatGPT is designed to model human language. So it may seem like you are having a conversation and that it understands what you mean or is thinking about the best answer to give you, but in reality it is using mathematical models to “guess” at the likely words to use, based on the patterns of language contained in its database.
There is considerable discussion about how and when to cite large language models such as ChatGPT. This discussion is ongoing and changing, so we recommend that you consult specific style guides, such as MLA, APA and Chicago, to get the most current guidelines. We also highly recommend the University of Waterloo citation guide as it is extremely comprehensive and illustrative. Here are quick examples for citing ChatGPT in MLA and APA citation styles:
Currently JAMA (AMA) style does not recognize citations to ChatGPT or other large language models. JAMA style only recognizes content created by humans and exclude content created by artificial intelligence. As this conversation around citing AI tools continues, this position may change so please consult JAMA styles directly.
OpenAI is aware of the potential for misuse and unintended consequences of ChatGPT, including the propagation of harmful content, biases, and misinformation. Some issues that users should be aware of include:
ChatGPT generates responses based on training data patterns but lacks fact-checking, potentially generating misleading or false information if input data contains inaccuracies or lacks context. Users should be cautious and verify the information obtained.
ChatGPT may generate inappropriate or offensive responses due to unfiltered internet text exposure. OpenAI's safety precautions are not foolproof and may result in false positives or false negatives.
ChatGPT heavily relies on user instructions for responses, and unclear or vague instructions can result in undesired answers.
ChatGPT interaction involves sharing potentially sensitive information, and there's always a risk of breaches or unauthorized access.
ChatGPT's overconfidence can lead users to place undue trust in its responses, as it generates high-confidence responses even when the answer is uncertain.
These are a few examples of some issues with ChatGPT; there may be more as the technology evolves. U-M has some very good guidance for students on using generative AI including ChatGPt.