In the context of generative AI, a prompt is any form of text, question, query, or information that communicates to AI what response you’re looking for. The way you phrase a prompt can influence the way AI responds to it. So, by adjusting and refining your prompt, you can lead the AI towards providing the type of response or answer you’re looking for.
The art of crafting an effective prompt is sometimes referred to as prompt literacy or prompt engineering. See prompt literacy and improving U-M GPT prompts for more information about crafting effective prompts.
The use of personal information collected by this service is safeguarded through several mechanisms, including Information Assurance and U-M policies, such as Privacy and the Need to Monitor and Access Records (SPG 601.11) and Institutional Data Stewardship Policy (SPG 601.12).
We may also share your personal information when required by law, or when we believe sharing will help to protect the safety, property, or rights of U-M, members of the U-M community, and U-M guests. Please read the privacy notice for ITS AI Services for additional information.
No. U-M does not engage in the training of these models, nor do we share any user-specific data for the purpose of improving these models. The data in our U-M AI platform is ours and is not shared with anyone.
- Limitations in Training Data: The model is trained on vast amounts of data but doesn't necessarily know every fact or the most up-to-date information. If something has changed or wasn't prevalent in the data the model was trained on; the answer might be off.
- Ambiguous Questions: If a question is vague or not specific enough, the model may guess at the intended meaning, leading to answers that aren't precisely aligned with the user's intent.
- Inherent Model Biases: The model has biases resulting from its training data. While efforts have been made to reduce egregious biases, creating a completely neutral model is impossible, given that the data it's trained on can be biased in various ways.
- No Real Understanding: GPT, like its predecessors, doesn't truly "understand" information in the way humans do. It identifies patterns in the input data and responds based on those patterns. This can sometimes lead to inaccuracies or overly literal interpretations.
- Complex or Niche Topics: For very specialized or niche subjects, the model might not have enough data to provide a detailed or accurate response.
- Model's Confidence: Sometimes, the model will produce an answer even if it's not very confident in its accuracy, given the input. It doesn't have the capability to say "I don't know" unless it's been specifically programmed to do so under certain conditions.
- Heuristics and Shortcuts: The model might take shortcuts in its responses based on patterns it has seen during training. For example, if many texts in its training data say "A is generally true," it might lean towards saying A is true even in contexts where A might not be.
- Model's Objective: The model's primary objective during training was to predict the next word in a sequence, not necessarily to provide factual accuracy. This can sometimes lead to discrepancies in its output.
If you notice inaccuracies, it's always a good idea to consult other sources or ask for clarification.
Your chat will be removed from the screen and cannot be retrieved. However, it is still accessible via logging information that is collected when using the product.
Yes. Contact us to show off what you’ve built. We’d love to learn from you and see if there is an opportunity to share your work with all of U-M.
Yes. U-M GPT is approved for use with moderately sensitive data. Refer to Sensitive Data Guide ITS AI Services to learn more about what data is appropriate for U-M GPT, U-M Maizey, and U-M GPT Toolkit.
Moderate. Refer to the Sensitive Data Guide ITS AI Services page for more information.
At this time, data is collected in logs to help inform billing processes and for providing support, and data cannot be deleted.
Yes. Join ITS-AI-Services-Notify to receive email updates on ITS AI Services.
U-M AI services are housed in a private cloud environment in MS Azure.
ITS Marketing and Communications often uses AI for their work. You can read the transparency statement on the ITS website.
U-M hosts several Large Language Models (LLMs) that work through the processing of vast amounts of text data. They utilize an artificial intelligence technique called "transformer neural networks."
These models are initially trained on a diverse range of internet text. They generate responses by predicting the next word in a sentence, with their effectiveness determined by how coherently they can predict and generate these sequences.
Despite the advanced capabilities of these AI systems, it's important to note that they don't possess understanding or consciousness, but simply analyze and generate text based on the input and training they have received.
No, the University of Michigan does not engage in training U-M GPT, nor do we share any user-specific data to improve AI models.
The University of Michigan did not train the models hosted in AI models used in U-M GPT. These models (Llama 2, DALL-E 3, GPT-4o) were trained by the service providers and are informed by three key data sources: publicly accessible information found on the internet, licensed data from third-party providers, and inputs from users or human trainers.
See U-M GPT model descriptions for current offerings.
See instructions for switching models in U-M GPT.
Refer to U-M GPT In-Depth for more information on Large Language Models.
Llama 3 is an open-source language model available for anyone to use, experiment, and build tools using it as the base.
U-M GPT does its best to generate accurate information based on the diverse range of text it has been trained on. However, it does not verify or fact-check the information it provides. Moreover, it does not have access to real-time data or updates, nor does it understand the information it provides in the way humans do.
Therefore, while it is a valuable tool for general information, brainstorming, and idea generation, all ouput it generates should be reviewed and verified with reliable sources. Always consider it prudent to consult with knowledgeable individuals or fact-check using reliable resources for important matters.
- Limitations in Training Data: The model is trained on vast amounts of data but doesn't necessarily know every fact or the most up-to-date information. If something has changed or wasn't prevalent in the data the model was trained on; the answer might be off.
- Ambiguous Questions: If a question is vague or not specific enough, the model may guess at the intended meaning, leading to answers that aren't precisely aligned with the user's intent.
- Inherent Model Biases: The model has biases resulting from its training data. While efforts have been made to reduce egregious biases, creating a completely neutral model is impossible, given that the data it's trained on can be biased in various ways.
- No Real Understanding: GPT, like its predecessors, doesn't truly "understand" information in the way humans do. It identifies patterns in the input data and responds based on those patterns. This can sometimes lead to inaccuracies or overly literal interpretations.
- Complex or Niche Topics: For very specialized or niche subjects, the model might not have enough data to provide a detailed or accurate response.
- Model's Confidence: Sometimes, the model will produce an answer even if it's not very confident in its accuracy, given the input. It doesn't have the capability to say "I don't know" unless it's been specifically programmed to do so under certain conditions.
- Heuristics and Shortcuts: The model might take shortcuts in its responses based on patterns it has seen during training. For example, if many texts in its training data say "A is generally true," it might lean towards saying A is true even in contexts where A might not be.
- Model's Objective: The model's primary objective during training was to predict the next word in a sequence, not necessarily to provide factual accuracy. This can sometimes lead to discrepancies in its output.
If you notice inaccuracies, it's always a good idea to consult other sources or ask for clarification.
Yes. See Getting Started with U-M GPT for details and instructions.
You must be the owner of an MCommunity Group to create a project in U-M Maizey. If there are no MCommunity Groups available in the drop-down, it means you are not currently an owner of an MCommunity Group. Refer to Creating, Renewing, and Deleting MCommunity Groups for more information.
If you created a new MCommunity group while logged in to U-M Maizey, you will need to log-out for U-M Maizey to present your new group as an option.
It should only take about 15 minutes for a new MCommunity group to appear in U-M Maizey. Users will need to log out of U-M Maizey and log back in to access new MCommunity groups.
Yes, see Change Project Settings on a Maizey project.
A Shortcode is a unique numerical identifier used mainly within the university's administrative systems and processes. These codes are typically associated with departments, units, or specific projects and are used to simplify and standardize financial transactions, budgeting, and reporting.
Typically, students wouldn't have access to a Shortcode and would need a faculty/department to sponsor their request so that they can get a Shortcode.
Maizey projects for academic Canvas courses automatically include an allocation and do not require a Shortcode.
While a Shortcode is required for all other Maizey projects, Maizey is available at no cost until June 30, 2027. However, there will be costs associated with Maizey projects once the no-cost period ends.
Maizey project costs vary based on the amount of data being indexed and how much use the tool gets (prompts and replies). After creating a Maizey project, costs associated with your project are visible under Billing Details on the Project Details view. During the no-cost period, this can be used to estimate future costs of your project once the no-cost period ends on June 30, 2027.
See the Pricing page for more details.
No. Refer to U-M Maizey In-Depth for guidelines on using U-M Maizey.
The connector uses the Canvas API to read a Canvas site’s:
- Announcements
- Assignments (only published)
- Files
- Files must be visible in the course navigation menu. Maizey will not index files that are linked in Pages, Modules, etc. if Files is hidden from the course navigation menu.
- Lecture Recordings
- Modules (Only unlocked modules)
- Pages
It is authorized using a token for an ITS-created user. Anything the user can access, the connector can. In order for the module to work, an instructor must add that user as a student to their canvas site (any higher role, and it will index content that may be hidden).
Refer to Getting Started with Maizey Canvas Connector for step-by-step instructions.
The following file types are supported: md, htm, html, docx, xls, xlsx, pptx, pdf, rtf, txt
Yes. Instructors who use CAEN Lecture Capture can enable the Lecture Recordings tool in Canvas. U-M Maizey will then index the recording transcripts. Learn more about Using Maizey with Canvas and CAEN Lecture Capture.
U-M GPT Toolkit allows users to connect U-M’s GPT models to their application environments via API. It is the most advanced and flexible of our GenAI service offerings. To access U-M GPT Toolkit, log in to the self-service portal and create an API key.
U-M GPT Toolkit can provide tools for advanced users to connect application environments empowered by ITS AI services. Refer to Getting Started for more information on U-M GPT Toolkit.
A Shortcode is a unique numerical identifier used mainly within the university's administrative systems and processes. These codes are typically associated with departments, units, or specific projects and are used to simplify and standardize financial transactions, budgeting, and reporting.
Typically, students wouldn't have access to a Shortcode and would need a faculty/department to sponsor their request so that they can get a Shortcode.
U-M GPT Toolkit project costs vary based on the amount of data being indexed and how much use the tool gets (prompts and replies).
See the Pricing page for more details.
Example code can be found on our GitHub repository. Please submit a request if you need access.
Students cannot directly create U-M GPT Toolkit API keys because access to the self-service portal is restricted to faculty and staff. Additionally, API key creation requires an approved U-M Shortcode due to associated costs.
API keys must not be generated for students who are not employed by the university.
Faculty or staff may sponsor eligible student employees by creating and managing API keys on their behalf through the self-service portal. Access is limited to students employed by the university, and any use must be restricted to activities related to their employment responsibilities.
Claude Code is an AI-powered coding agent developed by Anthropic that operates directly in your terminal. It can read your entire codebase, edit files, run tests, manage Git workflows, and build applications from scratch based on natural language prompts.
Only students employed by the university are eligible to access Claude Code through sponsored Toolkit API keys.
Faculty or staff may sponsor eligible students employees by creating and managing API keys on their behalf through the self-service portal, but students cannot directly create Toolkit API keys as access requires an approved U-M Shortcode due to associated costs. Access is limited to students employed by the university, and Claude Code should be used only for tasks related to their employee responsibilities. Keys should not be generated for non-employed students.
A Claude Code subscription purchased directly from Anthropic is a flat monthly fee that gives you access to Claude Code as a standalone product. ITS Claude Code connects Claude Code to U-M's AI API gateway, charging only for actual usage against your department or project's API allocation, no subscription required. For light or occasional users, this is often more cost-effective. It also means your usage runs through U-M's managed infrastructure, which includes the security, privacy, and compliance controls ITS applies to all models in the gateway.
Our hosting provider does not provide the full functionality at this time. We recommend that you try resetting your config file. If issues persist, contact your unit IT department.
Currently, the default model is Opus 4.6 (1M context).
Once you are inside Claude Code, type /model into the prompt window. A menu displays allowing you to switch between models.
Claude Code project costs vary based on the amount of data being indexed and how much use the tool gets (prompts and replies). See the Pricing page for more details.
Please visit the billing dashboard in the Toolkit self-service portal to find the amount that will be billed to your Shortcode.
The Total Spend value is accurate. Due to a current dashboard limitation, the Total Tokens value combines input, cache read, and cache write tokens into a single total. Because these token types are billed at different rates, the token count will not directly match the displayed cost.