Generative AI Tools

AI Tools Within and Beyond U of T’s Protected Environment 

When deciding whether and how to integrate generative AI into classroom environments, it is valuable to consider which tools are institutionally-supported at the University of Toronto. 

The below resource differentiates between two categories of generative AI tools:

1. Inside the Walled Garden: Generative AI tools that operate within U of T’s protected digital environment, ensuring institutional support and compliance with university policies.
2. Outside the Walled Garden: Generative AI tools that fall beyond the University’s officially supported toolkit, which may require additional consideration before use.

Within the Walled Garden

Protected Environment Supported by U of T

Outside the Walled Garden

Applications Beyond the University’s Supported Toolkit

This content is available for download for use in a range of instructor support contexts.

This work is licensed under a Creative Commons BY-NC-SA 4.0 International License.

Current Approved Tools 

Currently, Microsoft Copilot is the recommended generative AI tool to use at U of T. When a user signs in using University credentials, Microsoft Copilot conforms to U of T’s privacy and security standards (i.e., does not share any data with Microsoft or any other company) for use with up to level 3 data.

It is also free to use. Microsoft Copilot uses OpenAI’s GPT-4 model and performs comparably to ChatGPT. For more information about Copilot, refer to our Copilot Tool Guide. 

Microsoft Copilot is a general-purpose chatbot. With thoughtful prompting, it can be used for many teaching and learning purposes (e.g., see Ethan and Lilach Mollick’s More Useful Things site for a comprehensive prompt library). However, many other educational generative AI chatbots and tools are being developed that incorporate pre-defined prompts out of the box. 

Contact North is a not-for-profit, Government of Ontario-supported organization that has developed two generative AI teaching and learning tools (also based on OpenAI’s GPT-4 model), both of which conform to U of T’s privacy and security standards for use with level one and level two data: 

• AI Teacher’s Assistant Pro guides instructors though AI-assisted syllabus and teaching resource creation 

• AI Tutor Pro is a resource for students to build and test their knowledge 

 Other tools that conform to U of T’s privacy and security standards will be referenced here as they become available.
 

Artificial Intelligence Virtual Tutor Initiative 

The Artificial Intelligence Virtual Tutor Initiative includes faculty development activities to engage instructors in the development and administration of a Virtual Tutor in the form of a chatbot. The Virtual Tutor chatbots created as part of this program will be tailored to course materials leveraging Generative Artificial Intelligence (GenAI) to provide responses to student questions to support teaching and learning in courses offered in the fall term of 2024.

Participating instructors will be engaged with support staff and a cohort of peers during the design, deployment and evaluation phases of their project to extend our shared understanding of this new domain.  

Features of the Virtual Tutor Initiative include:   

• Participation in a virtual orientation session that will include content preparation training.

• Interim checkpoint surveys and a final project report describing the challenges and opportunities observed during project implementation.

• Participation in focus group and evaluation activities to provide insight on experiences that will inform institutional planning. 

• Availability of infrastructure and staff support resources for each project team’s Virtual Tutor for the fall 2024 term. 

The cohort leads for the initiative include:  

• Michelle Arnot, Professor, Teaching, Department of Pharmacology and Toxicology, Temerty Faculty of Medicine 

• Charlene Chu, Assistant Professor, Lawrence Bloomberg Faculty of Nursing 

• Joseanne Cudjoe, Assistant Professor, Teaching Stream, Department of Arts, Culture and Media, University of Toronto Scarborough 

• Emily Ho, Assistant Professor, Teaching Stream, Department of Occupational Science and Occupational Therapy, Temerty Faculty of Medicine 

• Nohjin Kee, Associate Professor, Teaching Stream, Department of Physiology, Temerty Faculty of Medicine 

This initiative is sponsored by the Centre for Teaching Support & Innovation (CTSI) and Information Technology Services (ITS). 

If you have any questions, contact: ctsi.teaching@utoronto.ca 

GPT Builder for AI Virtual Tutors

July 2024 

The University of Toronto recognizes the growing interest among instructors to create artificial intelligence virtual tutors. While the University has launched a pilot Artificial Intelligence Virtual Tutor Initiative, there is not yet a university-wide platform available for all instructors to use. We recognize that some instructors may be eager to experiment with tools like Open AI’s ChatGPT Builder and other options using their personal accounts in the absence of a supported tool. The aim of this guide is to describe how to build virtual tutors with ChatGPT Builder, including considerations for maximizing data privacy and security for students, as well as respecting intellectual property and copyright. While this guide focuses on ChatGPT Builder, the considerations outlined apply to any solution that falls outside of the University’s “walled garden” of educational technology that has been vetted through a procurement process, which includes a full privacy and security review.  

What is a Virtual Tutor? 

Large language models (LLMs) have already become part of our society, including profound impacts on teaching and learning. These models are pre-trained with an enormous amount of material and are able to execute a broad array of tasks by generating text, code, and image outputs to interact with users. However, many of these models lack specific domain knowledge, or are not easily tuned to the specific needs of particular classes or programs. While technical experts could fine-tune models to solve this problem, OpenAI (the creator of ChatGPT) has addressed this by allowing customizability through custom GPTs. This enables individuals to create custom chatbots tailored to their specific interests with no computer coding needed. A virtual tutor is a custom chatbot that is tailored to course materials and provides responses to students based on these materials.  

Considerations Before You Begin 

It is important to note that while free ChatGPT accounts can use a custom GPT, creating and maintaining one with GPT Builder requires a paid ChatGPT Pro account. The University does not directly cover the cost of a personal ChatGPT Plus account – however, you may consider using your Professional Expense Reimbursement Allowance (PERA) to cover the cost. It is also important to be aware that access, terms of service, and functionality of the tool may change without notice. 

Privacy and Security 

As ChatGPT has not been reviewed for privacy or security by the University, the precautions outlined in the Tools Beyond Quercus section of the CTSI website remains a relevant framework to consider risks and implications. Specific considerations for virtual tutors created using GPT Builder include:  

• Provide instructions for how students can engage with ChatGPT without sharing their conversations. For example, show them how to engage with ChatGPT using its temporary chat mode, and how to opt out of sharing conversation data completely.  
• Ensure that students are aware of ChatGPT’s Privacy Policy, and that they need to create an account to use your virtual tutor.  
• Remind students to avoid inputting any personal details into prompts.  

Student Equity and Transparency 

• Students cannot be compelled to create accounts on non-University systems or services.  
• Environments external to U of T that are to be used by students as part of course activities must be explicitly listed in the course syllabus.  
• If an external virtual tutor is used as part of a course assessment, a viable alternative must be made available for students who do not consent to participation.  

Accuracy 

• Although uploading course materials into GPT Builder’s knowledge section will generally reduce the rate of hallucinations (i.e., factually inaccurate outputs), AI-generated text and code are not guaranteed to be correct.  
• Students should be reminded that a virtual tutor should not be treated as an “authoritative source” – for example, students should continue to refer to the course syllabus directly for key dates and course policies.  

Copyright and Intellectual Property 

• Be mindful of what content is entered into your virtual tutor – never include confidential information or significant portions of intellectual property you do not have the rights or permissions to.  
• All content entered may become part of the tool’s dataset and may inadvertently resurface in response to other prompts.  
• The legality of using copyright protected content in AI tools without permission is unknown and will evolve over time. Review and point students to the terms of service of each tool, as the ownership of input/put may be addressed.
• See the U of T Library’s Generative AI tools and Copyright Considerations for more information. 
   

Setting Up a Virtual Tutor using GPT Builder 

The following is a step-by-step guide on setting up and configuring a virtual tutor using GPT Builder:  

Step 1: Access GPT Builder 

• Create or log in to your ChatGPT Plus account at https://chat.openai.com. 
• In the left sidebar, click on “Explore GPTs”. 
• Under “My GPTs”, click the “Create a GPT” button. 

Step 2: Provide Initial Instructions 

• In the “Create” tab of GPT Builder, it will ask you “What would you like to make?” – this will initiate a dialogue to determine specifics about your GPT’s purpose and behavior based on your initial prompt. 
• For example, your initial prompt might be: “Create an expert tutor to help students in ECO101 understand basic microeconomic concepts”. 
• GPT Builder will then respond with specific questions about the name, logo, and purpose of the GPT. Iterating several times with the GPT Builder through natural conversation is recommended to provide the model with a clear understanding of what it should be doing.   

Step 3: Refine Your Instructions 

• Switch to the “Configure” tab for more advanced customization options.  
• Under “Instructions”, review the virtual tutor instructions that were generated by your dialog with the GPT Builder to ensure that it is consistent with your instructional objectives. For example, if you want your chatbot to act like an expert tutor to help students build their conceptual knowledge, consider adapting Lilach and Ethan Mollick’s “Updated Tutoring Prompt”. 
• You may also want to add to the instructions; e.g., “Do not answer any questions or engage in dialogue about any topic other than…” 

Step 4: Upload Knowledge Files 

• Under “Knowledge”, you can upload up to 20 files containing information you want your GPT to reference. For example, upload lecture slides and recorded transcripts, spreadsheets, problem sets, etc. Your GPT will use this information to provide more accurate and contextual responses. 
• Be aware that your GPT can potentially be accessed by any user, with query results exposing your intellectual property outside of the University – evaluate if this aligns with your personal view on sharing course or domain related materials.  
• Do not include any copyrighted material (including full-text library licensed e-resources) without authorization. 
• Materials licensed under Creative Commons are generally acceptable to include with attribution (e.g., CC-BY-SA), though there may be some exceptions depending on the specific CC license – refer to the U of T Library’s Generative AI tools and Copyright Considerations for more information.   
• Only include material that does not contain any sensitive personal information about individuals.  
• In the “Instructions” section of GPT Builder, include “When referencing materials from the “Knowledge” section, provide a full citation to ensure proper attribution.”  

Step 5: Enable Relevant Capabilities with Caution 

• In the “Capabilities” section, you have the option to enable “Web Browsing”, “DALL·E Image Generation”, and “Code Interpreter & Data Analysis”. 
• Disabling “Web Browsing” will prevent the virtual tutor from accessing up-to-date information on the internet that is not included in its training data or your uploaded content. This may or may not be desirable, depending on your objective.  
• Be aware the code interpretation capability will allow users to download your uploaded content in its native format.  

Step 6: Configure Data Sharing Settings  

• In the “Configure” tab of GPT Builder, scroll down to “Additional Settings”
• You may uncheck the box for “Use conversation data in your GPT to improve our models”. This will prevent the content of user conversations with your virtual tutor from being used by OpenAI for model training and improvement.
• It is important to note that the privacy settings of your virtual tutor users will take priority over the setting above: if a user chooses to share their chat history to improve model training and development, this will expose their own chat history with your virtual tutor regardless of the setting you choose.

Step 7: Test and Refine 

• Use the “Preview” panel to test out your virtual tutor and see how it responds. 
• Iteratively update your instructions and settings based on its performance. 
• Try some prompts that purposely attempt to “trick” the chatbot into saying bad things, making things up, or going off topic to see how it responds. 
• Monitor for potential biases or inaccuracies and refine your instructions or uploaded files if necessary. For example, you might notice through testing your virtual tutor that a lecture transcript you uploaded contains errors.  

Step 8: Publish Your Virtual Tutor 

• Once finalized, click the “Save” button and choose your sharing settings. 
• Selecting the “Anyone with the link” option will allow students to access your chatbot but won’t enable public browsing.  
• Selecting “GPT Store” will publish your virtual tutor to OpenAI’s public “Explore GPTs” page.  
• Click “Confirm” to complete the publishing process 

This guide was inspired by Custom GPTs at MIT Sloan: A Comprehensive Guide, updated with current information, and adapted to the University of Toronto context. 

How Generative AI Works

Are you interested in learning how generative AI works? A good introduction is provided by the Schwartz Reisman Institute for Technology and Society’s “What are LLMs and generative AI? A beginners’s guide to the technology turning heads”. You could also consider: 

The responses you receive from a generative AI tool depend on the prompts you enter, and the further refining of these prompts, which takes practice. As Ethan Mollick said, “The lesson is that just using AI will teach you how to use AI.” (Working with AI: Two Paths to Prompting) To get started, we recommend you consider the following. 

Be clear about what you want. Include detailed information in your prompt, including the desired format. “Write a paragraph about…” “Create an image containing…” Suggest a particular style (e.g., an academic essay or lab report) and include specific information you want to include (e.g., provide an outline or ordered steps for the prompt).    

• If you’re not sure how to describe the style you want to emulate, the Wharton School at the University of Pennsylvania suggests pasting in a text example you like and asking the tool to describe the style Use that description in your own prompt for style.   
• To learn more on prompt writing, see our Tool Guide under “How can I prompt with Copilot?.”

Be critical.

• Does the tool output meet your needs? What additional information is required? Generative AI is an interactive tool. Try different options and prompts to gauge the results, clear the prompt screen and try again. You will learn to refine your prompts and better discern what is most effective with practice.   
• Generative AI tools can provide quick results that may appear correct, but looks can be deceiving. Tools such as ChatGPT can produce hallucinations or misleading and factually incorrect text. As with any text or visual analysis, we need to examine the results with a critical eye.