AI (Artificial Intelligence) : Teaching with AI

A best practice is to design assessments that focus on higher-order thinking skills and require students to demonstrate their understanding and application of knowledge in meaningful ways. Check out his article from Inside Higher Education: Embracing Constructive Dialogue and Oral Assessments in the Age of AI.

Below are some points to consider [from MHS Library].

Check assessment task design against Generative AI tools: 

It can be useful to ‘test’ whether your assessments are designed in a way where they can be easily completed by artificial intelligence. For example, it is possible to register an account and explore the capabilities and limitations of ChatGPT. Note: Creating an account and using ChatGPT provides additional data to OpenAI, the company that created ChatGPT.

These activities could be used to assist in reviewing how amenable an assignment is to being answered by ChatGPT:

• Paste an entire assignment brief into chatGPT and see what it produces
• After reviewing the result, add additional instructions into ChatGPT to try to finetune the results.
• Sometimes, poor results are due to entering insufficient prompts into ChatGPT rather than because ChatGPT cannot produce a satisfactory response to an assignment. Test this by adding further instructions derived from the assessment rubric, marking guide, and taught material. Try a variety of combinations of instructions and see if this improves the results.

Notes for testing ChatGPT results:

• For many assignments, ChatGPT will produce a reasonable-looking structure but will often lack precise details in some areas and may contain factual inaccuracies or non-existent references. In many cases, ChatGPT will ignore some instructions even with additional prompting.

• When reviewing assignment instructions, the instructions given to ChatGPT can be repeatedly updated in an attempt to produce a satisfactory answer. ChatGPT is able to follow simple instructions and provide 
• answers that have a layer of correctness, but it is currently poor at contextualising responses. ChatGPT is a language model that provides responses based on statistical convergence rather than interpreting instructions in the way a human would. It is currently poor at ranking the importance of different instructions, or of prioritising specific elements, without human prompting.

NOTE: if you want ChatGPT to provide answers to quizzes it writes, include that in the prompt, i.e. Generate a five question multiple-choice quiz on the causes of depression with four answer choices per question, and indicate the correct answer for each question.