Using generative AI as a thought partner in your coursework may enhance your critical thinking and creativity for certain types of assignments. Before you use generative AI in health care innovation courses, however, please read:

Remember that it is plagiarism to submit work produced substantially by an AI tool. AI should supplement your own thinking, not replace it.

Accessing Microsoft Copilot

If you choose to use generative AI in your coursework, please utilize tools that are licensed by Penn. You can access the Microsoft Copilot AI chatbot as part of Penn’s institutional Office 365 subscription. To log in to Copilot for the web:

  1. From a web browser, enter copilot.microsoft.com. You should see the Copilot homepage.
  2. Click the Sign In button.
  3. You may be asked if you want to sign in with a personal account or a work account; choose the work account.
    Screens screenshot of a phone AI-generated content may be incorrect.
  4. Sign in with the same credentials you use for your Penn Medicine email.

Why Use a Penn-Licensed Chatbot?

Using Copilot under Penn’s enterprise license affords you additional security:

  • Your chats are kept private to you unless you choose to share them.
  • Microsoft employees and partners, such as advertisers, cannot see your data.
  • Chats and prompts are not used to train AI models or finetune the Bing search algorithm.

Responsible Use

Even with the enterprise license, please be responsible with sensitive information:

  • Use your Penn account for Penn-related work only.
  • Do not input confidential or proprietary information, protected health information, or student-identifiable data.
  • Consider whether information you input in an AI chatbot might violate copyright.
  • Follow acceptable use policies from the University and MEHP Online (see policies linked above), as well as those in the directions for specific assignments.

AI may output inaccurate or misleading information. Review all outputs for accuracy and appropriateness. You are responsible for the veracity and accuracy of the coursework you submit

Finally, be aware of generative AI’s environmental impacts. The power needed to run large language models is significant, and has negative consequences in terms of climate, water usage, and air pollution. Therefore, think carefully about when the value of using generative AI in your work outweighs its negative consequences.

When to Use Generative AI

MHCI faculty, staff, students, and alumni have found generative AI helpful for these types of tasks:

  • As a thought partner
    Chatbots can spark your creativity. You can ask AI for feedback on your ideas or to generate ideas on a given topic. It may be most effective to ask for 5 – 10 options, of which you can choose the best. Note that diversity of ideas is not generative AI’s strong suit, and use it as a starting point, not an end product.
  • As an editor
    Prompt the chatbot with the specific type of feedback you want, as well as your intended audience, and ask it for suggestions in the form of bullet points. Not every suggestion is high quality, so think critically as you use the output to revise your work. Refrain from asking it to revise for you; the need to review and modify its output can add to your workload rather than reduce it.
  • To get started writing
    Chatbots are good at identifying connections among disparate data points. Take advantage of this strength by uploading your preliminary notes and asking the chatbot to help you identify a thesis, hypothesis, or research question—or to create a topic outline for a project. Not every suggestion is high quality, so once again, think critically and revise as you work on your deliverables.
  • To create reading guides
    Chatbots are bad at summarizing but good at picking out core topics and themes. Ask the chatbot to identify those topics and use them as an aid to guide your reading. The AI cannot replace the need to read, but it can help you read smarter.

Prompting a Chatbot

To get the most out of chatbot interactions, effective prompts should include:

  • A clearly defined statement of what you want the chatbot to output as an end product.
  • Details and context including (but not limited to):
    1. Information about the audience
      Is this a presentation to the CFO? A talk at a conference? A chat with a patient?
    2. Information about the chatbot’s role
      Should the chatbot be a physician? An economist? A health policy expert?
    3. Directions for how to format the output
      Should it give you paragraphs? Bullet points? An image?
    4. Additional guidance for what it should and should not do
      Do you want it to refrain from using contractions? Should it only use clinical examples? Should it not bring in sources from before 2020?

Do not assume that the chatbot can guess the parameters of your request. Without context, output is likely to be unacceptably generic.

Strategies for Effective Use

  • Iteration and revision
    Ask the chatbot for revisions and follow-ups to its initial output. Be clear about what you do and do not want to keep from earlier iterations and ask it to build on the parts you liked.
  • Roleplaying
    Ask the chatbot to take on the role of a subject matter expert and write from that point of view. Or ask it to take on multiple perspectives to get a more comprehensive view of a topic. To avoid confusing yourself and the chatbot, ask for each perspective as a separate prompt.
  • File uploads
    Upload a document, image, or other file that you have created. Then prompt the chatbot to complete a task related to that file. Remember to specify that it should respond based only on that file, if you do not want it to use outside information. And refrain from uploading proprietary, confidential, or copyrighted materials.
  • Output stitching
    Given multiple variations on the same prompt, a chatbot is likely to output significantly different answers each time. Combine the best, most relevant parts of each answer to build the response that will be helpful to you.
  • Prompt chaining
    Encourage structured thinking and more accurate answers by dividing a large task into smaller steps, and ask the chatbot to complete each step, building on its previous output. You can experiment with basic prompt chaining techniques on your own. LinkedIn Learning, to which you have access through Penn, has training resources on the topic.
  • Few-shot prompting
    In addition to your core prompt, provide the chatbot with several examples of the type of deliverable you would like it to output. The examples help the chatbot better understand the type of response you are seeking and calibrate its output to your needs. LinkedIn Learning has training resources on this topic, too.

IT Help Desk

Get help from the Tech Center at Penn: https://techcenter.upenn.edu/support/home.  For Copilot training, you may search Penn Libraries or LinkedIn Learning.