Canvas Chatbot Gets an AI Upgrade

The Northwestern Canvas Chatbot debuted in the winter of 2019 and has since helped thousands of students and instructors find quick and helpful answers to their basic Canvas questions. Following last fall’s update to incorporate a handoff to the live Instructure Support Chat, the chatbot support team in Northwestern IT Teaching and Learning Technologies (TLT) has now incorporated a generative artificial intelligence (AI) model to improve a subset of its responses.

Until now, the Canvas Chatbot has historically relied on the use of conversational AI and machine learning (forms of AI that received most of the attention before generative AI was popularized by ChatGPT) to interact with users using natural language. These techniques enable simple conversations with a limited number of actions that can be taken on a user’s behalf. When the Canvas Chatbot doesn’t readily know the answer to a user’s question, it has several potential fallback actions to suggest to the user, including a dynamic search of public Canvas Guides. Previously, when the chatbot suggested that the user read an associated guide to find the answer to their question, it was only able to provide a link to the page. Now, the incorporation of generative AI into the chatbot has enabled the creation of custom responses based on the text of the linked guide.

For Example

Let’s say an instructor wants to view the average grade for their class. They might ask the chatbot something like: “How do I view the average course grade?” In the past, the chatbot would likely have responded with: “Hmm. I think you’re asking about New Analytics. Please see this Canvas Guide for more information: [LINK].”

With the incorporation of generative AI into the chatbot toolset, the chatbot is now able to not only identify the correct help guide but also to review its text and provide a targeted answer to the question. Or, when it cannot readily come up with an answer based on the guide’s contents, it will produce a summary of the relevant Canvas guide instead. See the screenshot below for an example of a response created by generative AI:

Both of these upgraded response possibilities are a significant improvement in comparison to the original answer. In addition, as answers are only provided based on the content of existing guides, there is no risk of the bot hallucinating or producing unethical or improper content. Anyone using the chatbot will know when an answer has been provided using generative AI as the response will include the text: “snippet above produced using generative AI.”

How Does this Compare to Other Generative AI Chatbots?

Don’t expect a ChatGPT-like interaction when using the Canvas Chatbot. While generative AI has been incorporated in this limited fashion, conversational AI techniques – not generative AI – still underlie most of the chatbot’s functionality. It will not summarize text submitted as a question or create unique responses for anything non-Canvas-related. Plans to continue incorporating limited generative AI capabilities in other parts of the chatbot’s conversational workflow are ongoing.

To talk to the Canvas Chatbot, go to Canvas in your web browser and click on the Ask icon at the bottom of the left-hand global navigation menu. And remember, generative AI responses are only currently given for questions that have an answer which can be found in existing Canvas guides (so only questions about core Canvas functionality, nothing about other Northwestern systems).

The Canvas Chatbot is supported by a team of developers, content experts, and instructional designers who regularly review interactions with the bot to help assure quality responses. The team consists of TLT members Patricia Goldweic, Rachel Goc, Bea Jimenez, Ivette Torres, and Dan Hoefler. If you are interested in a discussion on the ongoing modeling of the Canvas Chatbot, email canvas@northwestern.edu and address the message to Patricia Goldweic and the chatbot team.

Give it a try today!