A Best Practice Guide to Using Generative AI
This guide supports UBC Faculty of Medicine staff, faculty, and learners use Generative AI tools responsibly. Content covers approved platforms, data privacy obligations, and strategies for getting reliable outputs.
01. UBC Copilot Chat
UBC’s Copilot Chat is the recommended tool for exploring Generative AI. Approved for use at UBC, it keeps your information private and is currently cleared for Low and Medium Risk data.
Verify your session: Log in with your CWL and look for the green shield icon labeled “enterprise data protection” in the upper right.
DATA PRIVACY & RISK
UBC Copilot Chat is a secure, enterprise level platform, and does not use your inputs for training. It is approved for use with Low Risk and Medium Risk data. This covers most use cases, but please reference the Office of the CIO’s Data Security Classification document before getting started.
CONSENT & DISCLOSURE
If you are obtaining data from or about learners, faculty or staff, it is best practice to obtain informed consent before processing that data with AI. Additionally, if you are sharing a document or report generated with AI it is best practice to disclose where and how it was used.
DATA RETENTION
If AI is used in any way to make decisions about an individual’s performance you will be legally required to retain all the information, prompts, inputs, outputs and reports for 1 year. This could be as simple as using AI to summarize student feedback in a performance review. If you are unsure whether your use case applies, you can also consult with the Records Management Office.
Working with High Risk or Very High Risk data? Contact edmedia.med@ubc.ca for guidance.
02. Prompt Engineering & Avoiding Hallucinations
Prompt engineering is the practice of writing clear, well-structured instructions to get better, more accurate results from an AI model. The better the instructions, the better the output, without changing the AI itself.
Hallucinations occur when an AI model confidently produces false or fabricated information presented as fact. Each scenario below shows a common pitfall, what goes wrong, and how to avoid it.
Common Pitfall
What Went Wrong
How To Avoid It
Vague, open-ended prompt
“Tell me about medication errors in hospitals.”
Copilot fills the gaps with plausible-sounding but invented statistics or outdated studies and it will present it confidently as fact.
Be specific and constrained
“Using only the attached incident report from Q1 2025, summarize the top 3 medication error categories.”
No source material provided
“Summarize the key findings of our department’s last accreditation review.”
With nothing to anchor to, Copilot invents plausible findings your department never actually recorded.
Ground the response in source material
“Only answer using the attached PDF. Cite the section number for each finding you include.”
Asking a leading question
“What evidence exists that this training module improved learner outcomes?”
The premise is assumed true. Copilot fabricates supporting evidence rather than evaluating whether it exists.
Ask neutral, open questions
“Based on the attached evaluation data, what does the evidence show about learner outcomes from this module?”
One large, complex request
“Analyze this dataset, identify trends, and write a report with recommendations.”
Broad scope causes Copilot to speculate at each stage, compounding errors across the output.
Break complex tasks into steps
Step 1: “Summarize this dataset in 5 bullet points.” → Step 2: “What patterns do you notice?” → Step 3: “Draft 3 recommendations based on those patterns.”
Accepting the first output without checking
You publish a Copilot-generated summary without reviewing the cited sources.
Errors, fabricated references, or outdated information reach your audience unchallenged.
Verify and iterate
“Double-check this summary against the attached document. Flag any statement you cannot directly support.” Always have a human cross-reference outputs with trusted sources.
⚠️ Never assume an output is correct. All Copilot responses must be verified by a human and cross-referenced with trusted sources. The more grounded, specific, and verifiable your prompt is, the less likely hallucination becomes.
03. Getting Help & Training
The Digital Solutions Educational Technology (EdTech) team offers support, consultation, and training sessions for UBC Copilot Chat.
UBC Copilot Chat is approved for use with Low Risk and Medium Risk data. Data classification information is available on the Office of the CIO’s website. If you are unsure whether the data you are processing needs to be retained you can consult with the Records Management Office.