Drexel UniversityDornsife School of Public Health

Department of Environmental
& Occupational Health

AI WORKSHOPS FOR DSPH

Foundations & Capabilities

What generative AI is, which Drexel tools to use, and how to ask for what you need. No experience needed.

Alex Quistberg, PhD, MPHAssociate Research Professor · Department of Environmental & Occupational Health

Core 1 · Fall 2026 · on Zoom

WHAT YOU'LL LEAVE WITH

Today's Goals

1Say in one sentence what tools like ChatGPT do, and what they get wrong.
2Know which Drexel tools to use and what never goes into them.
3Sign in to ChatGPT Edu, type a request, and ask a follow-up.
4Use four simple questions (PTCF) to get a better answer.

SECTION 01

What AI Is

What these tools do, which ones to use, and the one data rule.

THE MENTAL MODEL

AI Drafts, You Decide

ChatGPT and Copilot write text by predicting what words should come next, based on patterns in a huge amount of writing.

Good at drafting, summarizing, rewriting, and explaining.
Can sound confident and still be wrong. They are not search engines or databases.
Think of a fast, eager assistant who has read a lot but is sometimes wrong. You check the draft before you use it.

A FEW WORDS YOU'LL HEAR

Four Terms

Prompt

What you type into the box: a question, a request, or instructions.

Chat

One conversation. It remembers earlier messages in the same chat, not other chats.

Model

The AI system underneath the tool. You can use the default.

Hallucination

When the AI states something false as if it were true.

TOOLS & DATA

Which Tool, and the One Data Rule

Which tool to use

ChatGPT Edu, signed in with your Drexel account
Copilot Chat, signed in with your Drexel account
Personal ChatGPT, Claude, Gemini: public or low-risk information only

Never paste in

Health information about a person
Student records or Social Security numbers
Identifiable research data
When in doubt, leave it out or make the details up

SECTION 02

What It Can Do

Six real tasks from public health work: analyze data, review a draft, build slides, sort an inbox, find sources, form student groups, and one look past the chat box.

USE CASE 1 · DATA ANALYSIS + FIGURE

Ask a Question of a Spreadsheet

LIVE DEMO

You upload a file and ask

“This file has pedestrian crashes in our city. Which speed limits have the highest share of crashes where someone was killed or seriously injured? Make a bar chart, give me 3 plain-language findings, and tell me about any data problems.”

Workshop1_Demo_RivertonPedCrashes.csv · 1,045 rows

Fabricated data for a fictional city. ChatGPT Edu writes and runs the analysis code itself; you can click to see the code.

What comes back

Severe outcomes rise from 10% of crashes on 25 mph roads to 47% on 45 mph roads.
Crashes after dark were more often severe (26% vs. 16% in daylight).

Flagged:6 rows missing a speed limit; 3 rows with an impossible hour (“99”).

Minutes instead of an afternoon. You still check the numbers and decide what they mean.

USE CASE 2 · REVIEW YOUR OWN DRAFT

Get a Tough Reviewer Before the Real One

Your draft (you wrote it)

The Safer Corners Study will evaluate curb extensions installed at 60 intersections between 2021 and 2024.

Aim 1: Estimate the effect of curb extensions on pedestrian injuries, comparing police-reported crashes before and after installation.

Aim 2: Explore community perceptions of safety.

Aim 3: Assess equity in the rollout across neighborhoods.

Our findings will eliminate pedestrian deaths at treated intersections.

“Act as an NIH study section reviewer. List the 5 biggest weaknesses, most serious first. Do not rewrite it.”

Reviewer critique (actual AI output, trimmed)

1Aim 1’s before-after comparison has no control group, so it cannot separate the effect from citywide trends or regression to the mean.
2Police-reported crashes undercount pedestrian injuries and are too sparse at 60 sites for adequate power.
3“Eliminate pedestrian deaths” is implausible and costs credibility with reviewers.
4Aim 2 names no population, sample, measures, or analysis; “explore” signals an unformed aim.
5Aim 3 defines no equity measure and does not connect to Aim 1.

AI critiques your draft; it does not write it. NIH will not review applications substantially developed by AI.

USE CASE 3 · MAKE OR UPDATE A PRESENTATION

From a Dense Abstract to a Talk Outline

You paste a published abstract and ask

“Turn this abstract into a 5-slide outline for a community meeting. One idea per slide, plain language, no jargon, and speaker notes for each slide. Then give me a PowerPoint file.”

Also works for updates: “Here is last year’s deck and this year’s numbers. Which slides need to change, and what should they say?”

What comes back (example): slides + speaker notes

Why Your Corner Matters

More people are being hit while walking. Most of it happens on a few kinds of streets.

1

Faster Roads, Worse Injuries

At 45 mph, about half of crashes kill or seriously injure the person walking.

2

What Works

Street design (shorter crossings, slower turns) beats awareness campaigns alone.

3

Speaker notes, slide 2: “Point to the chart. Ask: which of these roads do you cross every day? Pause for answers before moving on.”

In practice: today’s slides were built with AI help, then checked and edited by me.

It gets you a first structure fast. Your expertise decides what stays.

USE CASE 4 · INBOX + LITERATURE

Sort the Thread, Then Check the Sources

Example: a long planning thread → action list

“Pull every task out of this thread: who, what, and when. List anything nobody agreed to do.”

Who	Task	Due
Maria	Confirm speaker and title	Oct 20
Dev	Draft flyer for review	Oct 24
Priya	Approve catering budget	Oct 27
No one	Book the room (raised twice, never assigned)	—

Asked for 5 studies on leading pedestrian intervals

The model gave 2 and said it could not name 3 more without inventing them. Then we checked both DOIs:

REALFayish & Gross (2010). Transp Res Rec 2198:15–22. DOI correct.

WRONG DOIVan Houten et al. (2000). Real paper, but the DOI it gave (1734-11) opens a study of motorcycle helmets. Correct: 10.3141/1734-13.

Even a careful answer had a wrong citation. Open every source before you cite it.

USE CASE 5 · TEACHING ADMIN

Balanced Student Groups, No Names Shared

Three steps

In Excel, replace each name with an ID (S01, S02…). Keep the name key on your computer.
Upload only the IDs and what matters for grouping: program, concentration, coding experience.
Paste the groups back into Excel and match IDs to names yourself.

“Make 6 groups of 4. Each group needs a strong coder, at least two programs, and no more than 2 students from one concentration. Show a table and check each rule.”

What comes back (example, 24 fabricated students)

Group	Students	Strong coder	Programs
1	S04 S09 S20 S21	S21	MPH, MS, PhD
2	S05 S06 S07 S16	S06	MPH, PhD
3	S10 S12 S13 S23	S23	MPH, MS
4	S03 S17 S22 S24	S03	MPH, MS, PhD
5	S08 S11 S15 S19	S08	MPH, MS
6	S01 S02 S14 S18	S14	MPH, MS, PhD

Check it: count each rule yourself. Models sometimes report a rule as met when it is not.

Even in ChatGPT Edu, send the least information needed. Names stay with you.

BEYOND THE CHAT BOX

AI That Works Across Your Files

NOT DREXEL-SUPPORTED · LOW-RISK DATA ONLY

How this section of today’s slides was made

One request: “Add 5–10 minutes of real use cases to today’s slides.” An AI agent (Claude, a paid tool) then:

Opened this deck, the workshop page, and my instructor notes in my OneDrive folder.
Asked me 4 questions about format, examples, tools, and timing.
Built the fake crash dataset, ran the analysis, and drew the chart.
Asked a second AI model for the reviewer critique and the citations, then checked the DOIs itself.
Wrote these slides into the deck. I reviewed every one.

What is different from a chat window

It reads and writes real files instead of copy-paste.
It runs code and checks its own work.
It can hand pieces of a job to other AI models.

Why it is not where you start

Personal accounts: no student, health, or research data.
More power means more to check.
Examples: Claude, OpenAI Codex, Copilot agents.

Start with ChatGPT Edu today. This is where the tools are heading.

SECTION 03

How AI Gets Things Wrong

And who is responsible when it does.

THREE THINGS TO REMEMBER

Wrong, Biased, and Yours

It Makes Things Up

It predicts likely-sounding text, so it can invent facts, numbers, and citations. Check anything that matters.

It Can Be Biased

It learned from writing that reflects society's inequities, so answers can repeat stereotypes. Our work affects real communities.

You Are Responsible

If your name is on it, you own it: accuracy, fairness, and tone. “The AI wrote it” is never the answer.

SECTION 04

Hands-On

A guided tour, three first requests, and a simple recipe for better answers.

GUIDED TOUR · ~10 MIN · FOLLOW ALONG

ChatGPT Edu, Step by Step

1Sign in with your Drexel account. Check you are in the Drexel workspace.
2Find the message box at the bottom of the screen.
3Type: “In two sentences, what can you help a public health staff member with?”
4Follow up in the same chat: “Give me three examples.”
5Copy the answer with the copy icon.
6Click “New chat.” It does not remember the old one.
7Find past chats in the list on the left.

React with 👍 in Zoom after each step. Stuck? Say so in the chat.

ACTIVITY 1 · ~11 MIN

Your First Three Requests

Copy each prompt from the workshop page into ChatGPT Edu, one at a time. There are no wrong answers.

Summarize: a short paragraph on pedestrian deaths, in 3 bullet points.
Draft: a friendly reminder email to a committee, under 100 words.
Explain: what a “data use agreement” is, in plain language.
Then follow up: “Make it shorter,” “More formal,” or “Add a sentence about ___.”

ACTIVITY 2 · ~11 MIN

Four Questions Before You Type

PTCF

Persona: who should it act as?
Task: what do you want it to do?
Context: what does it need to know?
Format: what should the answer look like?

Try both

Short: “Write an email about a meeting.”
Full: “You are an experienced department administrator. Write an email announcing a staff meeting Tuesday at 10 in Room 204 about the new travel policy. Under 100 words, warm tone, date and time first.”

WRAP-UP

What to Take Away

AI writes drafts. You check them and decide.

For Drexel work, use ChatGPT Edu or Copilot and follow the data rule.
Before you type, ask the four questions: Persona, Task, Context, Format.
Before Core 2: do one weekly task (an email, a summary, notes) with ChatGPT Edu, and bring it.

Drexel UniversityDornsife School of Public Health

THANK YOU

Questions & Discussion

Next: Core 2 — Workflow Enhancement. Start from your own plan, then let AI draft or refine.

Alex Quistberg, PhD, MPHdaq26@drexel.edu

Department of Environmental & Occupational Health