Use of AI tools in Assessment and Teaching

Generative AI chatbots are at their most helpful when users can clearly define and articulate their objectives and the relevant context and factors. Hence, it is important that instructors exercise their due diligence in planning and clarifying key information such as the learning objectives, assessment type, degree of difficulty for students that are aligned to course requirements. For more information on assessment planning and design, please refer to our page on setting quizzes and exams.

1. Define Learning objectives

   Lock in the specific knowledge and skills you want to assess. This will guide you in crafting prompts for Generative AI chatbots. Consider using verbs from the Blooms' Revised Taxonomy to help state what students are to demonstrate.

2. Prepare your training materials

   Feed the chatbot relevant course materials like lectures, notes, textbooks, and past exams. If your course refers to important and well-known works and concepts, specifying these would suffice without needing to insert the entire source. The more relevant data you provide, the better Generative AI chatbots can tailor questions to your needs.

3. Specify types of Assessment Questions

   Provide inputs on the types of Assessment questions like MCQ, short answer questions, essays, etc. This helps tailor the output to your assessment format. You could even set the difficulty level by specifying whether you want easy, medium, or challenging questions.

Do ensure your inputs and generated outputs are in line with SMU assessment policy (Click HERE for details)

As Generative AI technologies continue to evolve and grow in sophistication, more and more new chatbots are now available. Here's a selection of some of the better known (points below apply to the free tier):

For more information on the performance benchmarks of the most popular Large Language Models (LLMs), please refer to the reports published by Anthropic or Vellum.

Success in using Generative AI chatbots is determined to a significant degree by the quality of the prompt provided. Here are some guidelines to create better prompts:

• When choosing the model to work with, the latest and most capable models are likely to perform better.
• Start with a simple and short prompt, and iterate from there.
• Put the instructions at the beginning of the prompt, or at the very end. When working with large context, models apply various optimizations to prevent Attention complexity from scaling quadratically. This may make a model more attentive to the beginning or end of a prompt than the middle.
• Clearly separate instructions from the text they apply to - more on this in the next section.
• Be specific and descriptive about the task and the desired outcome - its format, length, style, language, etc.
• Avoid ambiguous descriptions and instructions.
• Favor instructions that say “what to do” instead of those that say “what not to do”.
• “Lead” the output in the right direction by writing the first word (or even begin the first sentence for the model).
• Use advanced techniques like Few-shot prompting and Chain-of-thought
• Test your prompts with different models to assess their robustness.
• Version and track the performance of your prompts.

When using generative AI chatbots, exercise care in checking and using the outputs, as responses from chatbots such as ChatGPT can contain inaccuracies, made-up sources and content, and offensive material.

Do also note that copyright principles and intellectual property rights apply in what you feed generative AI chatbots. Possible infringements include the feeding of copyrighted materials such as textbooks, articles, pictures, etc. For copyright concerns and advisory, please seek SMU Libraries support by clicking HERE.

For certain types of questions, the suggested approach may not work as well. These include questions that involve:

• Mathematical problems
• Coding problems
• Complex reasoning tasks
• Highly specialised / obscure areas of knowledge

Generative AI chatbots can be powerful tools for generating diverse and thought-provoking assessment questions, but human oversight and guidance are still essential for ensuring the questions are tailored to your specific learning objectives.

The following are some tips that may be helpful in refining and checking the outputs:

• Request for and check that the sources provided are real. Generative AI chatbots with internet access, such as Co-Pilot and Gemini, are able to check the responses given with a web search.
• Follow up on a question to be refined by requesting for multiple versions of it, specifying the aspects of it to vary if possible.
• Iterate on smaller subsets of questions, or even single questions. This is likely to produce better results than trying to improve everything simultaneously.
• Input questions you have used for an exam before and ask the tool to generate a different question.

Generative AI chatbots can be powerful tools for generating diverse and thought-provoking assessment questions, but human oversight and guidance are still essential for ensuring the questions are tailored to your specific learning objectives.

Overall, if AI tools are used carefully and questions are thoroughly reviewed, they could potentially be a helpful supplementary tool for busy professors. However, fully outsourcing exam creation to AI is risky. At most, AI should assist and inspire human question-writing, not replace it.

Human judgment, subject matter expertise, and understanding of the nuances of the course material should be central in the exam development process. Technology can assist, but should not be the sole author of something as important as exam questions.

• Focus on knowledge application instead of the final product. Ask for a clear demonstration of the application of core concepts and ideas to a specific audience, or context. Place this inside an authentic problem that requires critical appraisal of data and research (like a detailed case study). For example, rather than asking for a summary of research into a topic, ask students to write a proposal that shows the application of the outcomes of the research to resolve a specific problem. Follow this with required reflections from students detailing how they adapted the research findings from the general to the specific.

• Focus on the process instead of the final product. For example, getting students to submit an outline and multiple drafts for an assignment and see how they develop their arguments. The meta-data for each draft should show their editing time, changes made, etc. If they lifted content from generative AI tools, the meta-data may show massive changes in a very short editing time.

• Assessment question design. The assessment question could include information that is not available on the internet, and information that AI engines cannot gather. For example, questions or responses that include personal experience, cases/examples discussed in class. You may also consider getting students to critique draft responses generated by AI tools.

• Complement and vary different types of assessment. For example, following up a written assignment with an oral presentation where clarifying questions can be asked

Content

• Focus the presentation on application of core concepts to an authentic context/problem. Students may be encouraged to relate concepts to local or personal experiences.

Delivery

• Allow significant time after the presentation for a detailed question and answer session to allow the students to demonstrate their understanding of the topic (rather than just their ability to read from a script).
• Change the grading structure to weigh the answers to the Q&A more heavily.

Where appropriate assume that the code has been AI generated, so ask the students to do either or both of the following:

• Answer detailed questions on the specific choices made (For example, why a particular field would be needed in a database).
• Find and fix broken code under exam conditions. Explain why it was broken and what they would do to fix it, and why their fix would work.

Ask for demonstration of understanding of the data rather than the ability to generate them. For example, ask students to determine whether they believe that there is likely to be fraudulent behavior in the company as a result of their analysis of the figures.

Consider making them closed-book and using eLearn quizzes with lock-down browser. If a degree of open book is needed, consider only allowing physical paper or books.