A language model is a mathematical model that, when given a sequence of words in natural language such as English, predicts a likely next word. Modern chatbots such as ChatGPT are software applications that use language models to generate text in response to a prompt given by a user. For example: 

• you can show ChatGPT a source before prompting it to respond to the source in some way; 
• you can explain an idea before asking questions about it; 
• you can update the chatbot on a recent event before discussing it further. 

Language models, much like other machine-learning models, are trained to discern patterns in data and make predictions based on those patterns. One common technique is to show the model part of a text written by humans (such as an encyclopedia entry) and have the model predict the next word. The model checks its prediction against the actual word, adjusting its parameters based on whether it got its prediction right or wrong. Training such a language model requires large amounts of data and countless iterations.

A trained model considers a user’s prompt much like it considers its training data: as a sequence of words to which it appends a reasonable next word to continue that sequence - and then another word, and another.

ChatGPT was created by adding an extra layer of training during which human subjects wrote their own versions of the responses and ranked the quality of the chatbot's versions.

Language models are not specifically trained to distinguish fact from opinion or to discern whether factual claims are true or false. Researchers are making progress toward reducing false or misleading answers and "hallucinations," but "factuality" remains a significant challenge. ChatGPT cannot reliably cite sources or show its work.

Language models do not continually learn as users interact with them, inputting content. For example, ChatGPT has limited knowledge of info after 2021. The language model’s creators may collect the data generated by users’ interactions with the model and use that data to train future versions. Training a language model is a time- and resource-intensive process. That is why, among other things, there is a lag between the release date of a new model and the most recent data (such as texts describing current world events) it was trained on.

Language models do not search the internet or access a database to look up answers to users' questions. They do not access specific sources within their training data to provide an answer, much less cite any such sources accurately. Rather, they combine patterns from the training data to generate a plausible answer. (It is reasonable to expect that applications combining language models with internet or database search will eventually become commonplace, possibly replacing current search engines and voice assistants, but we’re not there yet.)

Language models are not trained for symbolic reasoning, or what many of us would refer to as the critical thinking that makes us human. They may appear capable of logic, skilled at mathematical operations, and familiar with the physical properties of the outside world, but they routinely commit errors in all those areas. Misinformation in the training data may be one reason; how the model is trained is another.

Design assignments with AI in mind

At first sight, language models appear astute, coherent, and articulate, but a closer analysis is bound to reveal limitations. While preparing your assignments for the course, you can use AI as part of your planning to enhance your assignment design. 

• Use generative AI (like ChatGPT) to reflect upon and refine your assignment prompts. Access sample essays created with OpenAI's GPT-3 provided by the Writing Across the Curriculum (WAC) Clearinghouse for examples.
• Develop prompts to get the best possible results, then develop prompts to make the model stumble.
• Identify patterns. Does the model favor a particular style? Does it make certain kinds of errors? Is it better at certain tasks or topics?

Further, think about your classroom policies and how you might revise them to reflect your teaching values and course outcomes.Here are some examples of syllabus updates from Professor Ryan Watkins at George Washington University.

Use it in course activities

Many teachers are already using chatbots in their classrooms. You may start with an informal introduction and eventually design formal assignments.

• Demonstrate language models in class and have a discussion with students; 
• Have students work informally with the models and analyze the output;
• Assign projects in which specific, structured interactions with language models are encouraged or required. Read through this Twitter thread from Professor Andrew Piper at McGill University for an example.

As this is a rather recent development, you have the opportunity to pioneer novel assignments and classroom activities.

We're finding articles daily with great ideas and resources, and we'll continue to update this list as we find useful information to share. In the meantime, check out some of the links below to help you brainstorm how to best embrace this technology in your course.

• ChatGPT and AI Text Generators: Should Academia Adapt or Resist? (Harvard Business Publishing)
• Artificial Intelligence Writing (University of Central Florida)
• Strategies for Teaching Well When Students Have Access to Artificial Intelligence (AI) Text Composition Tools (Stearns Center for Teaching and Learning at George Mason University)
• A Brief Summary of the Capability of ChatGPT (Professor Michelle Kassorla, Georgia State University)
• Why Banning ChatGPT in Class Is a Mistake (Campus Technology)
• Teaching Actual Student Writing in an AI World (Inside Higher Ed)
• Leveraging ChatGPT: Practical Ideas for Educators (ASCD)

Ready to experiment further? Check out these language model-powered applications.

• Elicit
• Explainpaper
• Polar
• GPTZero

If all of this is overwhelming and you need a little levity, try out CatGPT.