Are you struggling to keep pace with artificial intelligence (AI) and its rapid pace of change? At Academic Technology Support Services, we were too, so we created this post to make sense of the AI tool categories and how they can support the teaching process. In the post, we compare various AI tools and delve a little deeper into the category of AI Agents or Assistants.
Due to the rapid pace of AI functional development, combined with the proliferation of terms, definitions, and tools, this blog post has a significant caveat: This is our attempt to start/continue making sense of what’s new and upcoming in the use of AI; we aren’t looking for a right or wrong way to use an AI platform, or to accurately define groups of AI tools, but instead gain an understanding of how the various factors interact. Whether you agree with our characterizations or not, join the AI iCoP discussions and let us know your thoughts.
After our AI review, we will apply AI concepts and tools to an example from an instructor’s perspective of creating a writing assignment.
AI classification
Let’s begin by laying out a few functional categories of AI tools, starting with what is likely the most familiar. (Reminder: there is a lot of overlap and fuzziness in the following categories.)
Generative AI
What it is: Generative AI is a broad term for any AI platform that generates content.
How they work: Generative AI tools can be trained in several methods. One method is through static datasets such as books, articles, and webpages. This method can have a knowledge cutoff which means it would not have information on events or developments that happened after that time. An example of a generative AI tool that uses this method is ChatGPT. Another method is through real-time internet searches. This method does not have a knowledge cutoff. An example of a generative AI tool that uses this method is Perplexity. These methods are changing as newer models, such as ChatGPT 4o and o3, are developed.
Data security: Microsoft Copilot with Data Protection is the only generative AI tool approved for UMN use. It can be used with Public and Private-Restricted Data. Other generative AI tools should be used with Public data and materials for which you own the copyright. You should expect that any data entered into non-University supported AI tools will be retained and used for training by the AI tool you are using.
How to use: When creating prompts for generative AI tools, it’s most effective to use thoughtful and comprehensive wording to elicit stronger responses. Since prompts are an ongoing process, they may need refinement as you progress. Stay tuned for a future blog post that breaks down prompts in more detail.
Examples of generative AI tools include image generators (Midjourney, DALL-E), large language models (Claude, GPT-4), code generation tools (Github Copilot), and audio and video generation tools (resemble.ai, Synthesia).
Chatbots
What it is: A chatbot is a computer program that uses conversational AI, a type of AI that simulates human conversation, to interpret human language and simulate/engage in conversations with human beings or other chatbots.
How they work: A chatbot provides answers based on data sources such as websites, company information, and/or help documentation that was uploaded by the individual and/or organization who created it.
Data security: There are two UMN customer service chatbots in use. When using a chatbot for University purposes, use Public data only as a good practice.
How to use: When interacting with chatbots, users can choose different types of prompts depending on the chatbot’s purpose. Some prompts may involve simple question-and-answer interactions, while others may require more nuanced input to guide the chatbot through complex tasks.
There is a range in the levels of sophistication of chatbots. Examples include:
- Voice assistants such as Siri and Google Assistant
- Customer service chatbots such as Goldy chatbot on the One Stop and OIT websites
- Customized chatbots that are created to discuss specific topics such as reflection bots, role-playing/simulation bots, and class materials.
AI agents/custom AI assistants (task-oriented)
What it is: A task-oriented AI tool will automate tasks like summarizing, writing, or scheduling. It doesn’t just respond to prompts like generative AI, but semi-autonomously structures and synthesizes information.
How they work: With Custom AI agents, their scope of knowledge may be limited to either uploaded or internal data sources and/or the internet. For example, an instructor might upload their course syllabus, assignments, and learning materials to create a custom AI agent to support students.
Depending on the platform, custom AI assistants are referred to by different terms. For example, ChatGPT custom AI assistants are called Custom GPTs, and Gemini custom AI assistants are called Gemini Gems. AI is a fast-moving space, and there is no consensus on definitions. You may also see them called:
- Personalized chatbots
- Domain-specific AI assistants
- AI workflows
Data Security: Currently, there are no UMN-licensed tools to create custom AI assistants. UMN approved two custom AI assistants: FeedbackFruits Feedback Coach and Zoom AI Companion and Public and Private-Restricted data can be used with them. When creating/using a custom AI assistant for other University purposes, use Public data only as a good practice.
How to use: AI agents are pre-configured with built-in workflows or intuitive user interfaces (e.g., Grammarly, AI-powered search assistants). Users would enter a structured goal, and then the AI will break it into steps.
Examples of AI agents include FeedbackFruits Feedback Coach, Image Accessibility, AI Teaching Assistant Pro and AI Tutor Pro by Contact North, NotebookLM, Auto-GPT.
Autonomous AI (independent)
What it is: An autonomous AI tool can operate with little to no human intervention, making complex decisions and executing tasks independently.
How they work: It draws upon data sources that are dynamic and integrated with real-time systems and databases. For example, an autonomous agent might be designed to pull from secure data sources such as library databases, research repositories, and citation tools.
Data security: Currently, there are no UMN-licensed tools to create autonomous AI assistants. When creating/using an autonomous AI assistant for University purposes, use Public data only as a good practice. Because these tools can work independently, consideration should be given to when and how humans will retain oversight to ensure data security.
How to use: Autonomous AI tools need to be created based on specific goals for their use. When using an autonomous AI tool, users need to enter an initial prompt, and then the tool will run independently based on how it was built or trained.
Examples of autonomous AI tools include Operator, Auto-GPT, BabyAGI, CICERO, and AI co-scientist.
How, if at all, could these AI tools be used in teaching?
As we mentioned earlier, we’ll be using a writing assignment example to walk through how AI tools could be used at the different steps of creating the assignment. This is a fictitious example, and many of the steps are not currently possible for UMN instructors, both from an AI tool licensing perspective and a data security perspective. We delve more into this in the next section.
| Steps | Prompts/Creation Process |
|---|---|
| Step 1: Brainstorm a Writing Assignment | Ask ChatGPT to “Suggest five research paper topics related to restorative justice that align with critical thinking and real-world application. Provide key sources and possible case studies.” |
| Step 2: Align with Course Goals | Use NotebookLM to upload your course syllabus, a draft version of the writing assignment, and other course materials. After materials are uploaded, ask NotebookLM to “Map my assignment ideas to course learning outcomes and course materials.” |
| Step 3: Draft the Assignment Instructions | Ask NotebookLM to “Draft assignment guidelines that automatically link to campus plagiarism policies and grading rubrics stored in the LMS.” |
| Step 4: Support Student Learning & Engagement | Create the writing assignment in Canvas + FeedbackFruits. After it is created, ask FeedbackFruits Feedback Coach to “Provide real-time suggestions to students as they use the assignment grading rubric to write their peer review comments.” |
| Step 5: Deliver feedback to students | After training a custom chatbot/agent, download student writing assignments and ask the custom chatbot/agent to “Provide initial feedback based on the rubric (e.g., flagging missing citations or structural issues) before the instructor completes grading.” |
| Step 6: Evaluate and plan for continuous improvement | Create an autonomous AI tool that could crawl through the submitted assignments from students and feedback from instructors to identify areas of confusion and provide suggestions for changes to the assignment |
How, if at all, should AI tools be used in teaching?
The fictitious example we presented above is designed to brainstorm ideas about what’s possible with our new AI assistants. We are not advocating that AI tools should always be used, or, at the opposite side of the spectrum, that they should never be used in teaching. This is an ongoing conversation between instructors, students, instructional designers, administrators, and academic leadership currently taking place at universities worldwide. At the University of Minnesota, we are learning about AI tools, how they can be used, and how they should not be used. In December 2024, President Cunningham charged an AI task force to develop recommendations for how our university will approach the evaluation, use, and development of emergent AI tools and services. Task force members are focused on three areas: education, research, and administrative operations. Their goal is to deliver recommendations by September 2025. In the meantime, there are many ways to join the conversation: see ideas for how to get involved on the Navigating AI @ UMN webpage.
Sources
- What Are Generative AI, Large Language Models, and Foundation Models?
- What Is a Chatbot?
- No one knows what the hell an AI agent is
- Chatbot Levels – From Basic Q&A to Transformative AI
- AI in Education subgroup on Custom Bots/GPTs
Contributors
Annette McNamara and Jennifer Englund contributed to the creation and writing of this post. They wish to thank Ilya Begelman, Adam Brisk, Cherie Lemer, and Colin McFadden for their review and feedback suggestions.
Like what you read? Subscribe to the Extra Points Google Group for an email notification when the next blog post goes live.