Instructor Notes
Instructor Notes: Responsible Use of AI Tools on HPC
Maintenance: This Lesson Has a Shelf Life
More than any other workshop in the series, this one goes stale. Product names change (Bard became Gemini in February 2024), models are withdrawn (GPT-5.1 was removed from ChatGPT in March 2026), and prices move constantly.
The lesson is deliberately built so that only a small, marked surface needs maintenance. Everything else is framed around data classification, which does not change. Before each delivery, check these and nothing else:
| Location | What to check | Last verified |
|---|---|---|
episodes/01-ai-landscape.md |
Model size table, marked “as of August 2026” | 2026-08-25 |
episodes/02-types-of-models.md |
Open-weight model list and licences | 2026-08-25 |
learners/reference.md |
Cloud tool table; local model table | 2026-08-25 |
episodes/03-pomona-ai-policy.md |
Link to Data Governance page still resolves | 2026-08-25 |
Deliberate choices worth preserving if you edit:
- No quality rankings or leaderboard claims. They rot within weeks and are contested. The lesson ranks nothing.
- No prices. Tiers and figures change; the lesson says only that free and paid tiers exist.
-
Ungated, permissively licensed models in all
examples (
gpt-oss-20b), so no learner is blocked waiting on a Hugging Face access request mid-workshop. - The teaching point is never the model. If you swap in a newer model, the exercises should still work unchanged.
Workshop Overview
Duration: 1 day (3-4 hours contact time) Format: Lecture with discussion and interactive exercises Target Audience: All Pomona researchers (no technical background required)
Learning Outcomes: By the end, learners can: - Classify their research data into Pomona’s framework - Select appropriate AI tools based on data classification - Understand federal regulations affecting AI use (FERPA, HIPAA, export control) - Implement responsible AI practices (verification, disclosure, bias auditing) - Know how to get help from institutional compliance teams
Key Message: “Using AI responsibly is everyone’s responsibility. When in doubt, ask.”
Episode-by-Episode Teaching Notes
Episode 1: The AI Tools Landscape (15 min teaching + 5 min discussion)
Goal: Demystify AI tools and set context for responsibility discussions
Key Points to Emphasize: - AI tools are tools (not magical, not conscious) - Cloud AI is easy but sends data externally - Local AI is harder to set up but keeps data private - Choice depends on your data, not just convenience
Live Demo: - Show ChatGPT/Claude interface (2-3 minute demo) - Ask it a simple question live - Show result quality and speed
Common Misconceptions to Address: - “Cloud AI is always better” (False: local AI is safer for sensitive data) - “AI is perfect” (False: AI makes mistakes, hallucinates, has biases) - “Nobody can find out if I share data” (False: AI companies log everything)
Discussion Questions: - “Who here has used ChatGPT? What for?” - “What hesitations do you have about using external AI?” - “Have you thought about data privacy before?”
Troubleshooting: - If WiFi is slow, use pre-recorded demo - Have screenshots ready if ChatGPT is down
Episode 2: Pomona College AI Policy (20 min teaching + 10 min exercises)
Goal: Help learners understand institutional policy and data classification
Key Points: - Pomona has an official AI policy (not just guidelines) - Policy is based on data classification - Different data → different rules - FERPA is non-negotiable federal law
Structure of Teaching: 1. Show the four data classifications (2 min) 2. Explain each with examples (8 min) 3. Show decision tree (3 min) 4. Practice with examples (7 min)
Critical Concepts to Drive Home: - PUBLIC data: “Safe with ChatGPT” - PROPRIETARY data: “Local AI only, or you break NDAs” - RESTRICTED data: “NEVER external AI - federal law”
Use Real Examples: - Show Pomona’s actual policy document - Reference current news stories about AI/privacy breaches - Mention FERPA violations from other institutions
Interactive Exercise: Present 5 scenarios, have learners vote on classification: 1. “Published research paper abstracts” (PUBLIC) 2. “Student course feedback forms with names” (RESTRICTED - FERPA) 3. “Our lab’s unpublished methods (no data)” (PROPRIETARY) 4. “Patient survey data from hospital partner” (RESTRICTED - medical) 5. “Anonymized student survey responses” (PROPRIETARY/PUBLIC - explain anonymization nuance)
Discussion: - “What happens if you violate Pomona policy?” (Disciplinary action, retraction, funding loss) - “Has anyone gotten unclear guidance before?” (Normalize the question!)
Key Takeaway: “If you’re unsure about classification, email its-hpc@pomona.edu. That’s what they’re there for.”
Episode 3: Data Privacy and Classification for AI Usage (25 min teaching + 15 min exercises)
Goal: Deep dive into data protection; connect policy to actual research decisions
Teaching Structure: 1. Decision tree walkthrough (5 min) 2. Real-world classification examples (7 min) 3. Compliance frameworks (NIST, FERPA, HIPAA) (5 min) 4. Documentation templates (3 min) 5. Interactive exercises (15 min)
Emphasize FERPA: - Show what counts as educational record - Explain consequences (federal law, not just institutional) - Mention actual FERPA violation cases - Give clear examples of what to do/not do
Local AI Setup: - Show gocryptfs for encryption (safe storage) - Explain why encryption matters - Mention that Sagehen HPC is US infrastructure (export control safe)
Interactive Activities: 1. Scenario Analysis: Present 3 research projects, classify all data 2. Policy Application: Given classification, what AI tools are allowed? 3. Risk Assessment: What could go wrong if classification is wrong?
Discussion: - “What data do you work with in YOUR research?” - “How would you classify it?” - “Are you confident in that classification?”
Important Nuance: - Anonymization is tricky (difficult to truly anonymize) - Recommend: “If you have to anonymize, probably shouldn’t share” - Better: Use a local model on Sagehen
Closing: “The principle is: When in doubt, assume stricter interpretation.”
Episode 4: Running AI Models Locally on Sagehen HPC (25 min teaching + 20 min optional hands-on)
Goal: Show practical path for handling sensitive data safely
Two Options: 1. Hands-On: Actually run a local model on Sagehen (requires Sagehen access) 2. Demonstration: Show pre-recorded demo (safer if no WiFi/access)
If Hands-On (assumes SSH access available):
BASH
module load miniconda3 # PyTorch comes from conda: conda activate pytorch_env
# Create environment, load gpt-oss-20b, run query
# Time: 10-15 minutes for first load, 5 seconds for subsequent queriesKey Points: - Data stays on Pomona infrastructure - No external dependency on cloud service - Trade-off: Takes longer than ChatGPT (acceptable) - Quantization can speed up (explain 4-bit reduction)
Demonstrate: - Loading model (slow first time, normal) - Asking a question (show inference speed) - Comparing to ChatGPT (much slower but private)
Discussion: - “Would you use a local model for sensitive data?” (Yes, despite slowness) - “What if you need real-time speed?” (Use ChatGPT for PUBLIC data only) - “Can you combine both?” (Yes: local for sensitive, cloud for public)
Practical Workflow to Show: 1. Get research data 2. Classify it 3. If RESTRICTED: Use a local model on Sagehen 4. If PUBLIC: OK to use ChatGPT 5. Always verify outputs
Troubleshooting: - Model download fails? Pre-download and cache - CUDA not available? Check module loads - GPU memory error? Use quantized model (4-bit)
Episode 5: Ethical Considerations and Bias in AI Systems (25 min teaching + 15 min exercises)
Goal: Critical thinking about AI outputs; not all AI advice is good advice
Key Concepts: 1. AI can encode bias (from training data) 2. AI can hallucinate (make up facts) 3. AI can be confidently wrong (sounds authoritative when false) 4. This matters for research credibility
Teaching Structure: 1. Sources of bias (2 min) 2. Real-world bias examples (5 min) 3. Audit checklist (3 min) 4. Disclosure in publications (5 min) 5. Interactive exercises (15 min)
Real Examples of AI Bias: - Facial recognition worse on darker skin tones - Hiring algorithms that discriminate based on gender/race - Medical algorithms that undertreat certain populations - Language models that reflect stereotypes
Emphasize: - You (the researcher) are responsible for verifying - If you use AI output, you vouch for accuracy - Disclosure doesn’t excuse poor verification
Audit Checklist Activity: Present AI response, have learners identify problems:
Example 1:
AI: "The causes of poverty are primarily individual: laziness and lack
of work ethic account for 80% of poverty."Problems: - Ignores structural factors - Reflects harmful stereotype - Contradicts evidence
Example 2:
AI: "According to the latest research in 2025, quantum computers have
achieved general artificial intelligence."
(This is false - no such breakthrough exists)Problems: - Confident hallucination - Cites fake sources - Spreading misinformation
Interactive Exercise: Ask learners: “If you included this in your paper without verification, what would happen?” (Answer: Retraction, damage to credibility)
Disclosure Language: Show how to write in methods: - “ChatGPT was used for hypothesis generation [specific way]” - “All reported results were independently verified” - “No AI output was used in statistical inference”
Discussion: - “How do you currently verify sources?” (Apply same rigor to AI) - “What if AI gets it wrong?” (Your responsibility to check) - “Should you disclose when you DON’T use AI?” (No, only when you do)
Key Takeaway: “AI is a research tool, like Google Scholar. You wouldn’t copy Google Scholar results into your paper without reading them. Same principle for AI.”
Episode 6: Institutional Policies, Compliance, and AI Governance (25 min teaching + 15 min exercises)
Goal: Connect policy to action; know how to get help
Teaching Structure: 1. Regulatory overview (FERPA, HIPAA, export control) (8 min) 2. Compliance workflow (pre-research checklist) (5 min) 3. Consequences of violations (3 min) 4. How to get help (3 min) 5. Case studies (6 min)
Regulatory Focus:
FERPA (Most common at college): - What it protects (student records) - Why it matters (federal law) - What you can’t do (share with external AI) - What you can do (local model or fully anonymize) - Real consequences (institution loses federal funding)
HIPAA (If medical research): - What it protects (patient health info) - Why it matters (privacy + security) - What you can’t do (share with unapproved external AI) - What you can do (HIPAA-compliant service or local)
Export Control (If sensitive research): - What it protects (US technology) - Why it matters (national security) - What you can’t do (share with non-US entities) - What you can do (computation on US infrastructure)
Pre-Research Checklist: Walk through the provided checklist with example project:
Project: Analyzing student writing samples
Data: Student essays with names and student IDs
Classification: RESTRICTED (FERPA)
Regulations: FERPA applies
Approved Tools: Local models on Sagehen ONLY
Verification Plan: Manual checking of outputs
Disclosure: Methods section note about AI usage
This checklist prevents problems!Consequences Matter: - Violation = disciplinary action - Publications can be retracted - Funding can be lost - If FERPA: Federal law violated
Who to Contact: Create visual with flowchart:
Unsure about classification? → ITS (its-hpc@pomona.edu)
FERPA question? → Registrar's office
HIPAA question? → IRB
Export control? → Provost's office
General policy? → Your advisorCase Studies: Present 3 scenarios, discuss what went right/wrong:
Case 1: Good Practice - Researcher had student data - Classified as RESTRICTED - Used a local model on Sagehen - Documented decision - → COMPLIANT ✓
Case 2: Problem but Fixed - Researcher started to use ChatGPT with sensitive data - Realized it was FERPA-protected - Pivoted to local model - Still OK → COMPLIANT (after correction)
Case 3: Violation - Researcher sent student data to ChatGPT without checking - Student found their data online - Reported to administration - Paper retracted, disciplinary action - → VIOLATION ✗
Discussion: - “How would you handle Case 2 if it happened?” (Immediate action to switch to local AI) - “What early warning signs indicate violation risk?” (Unclear classification, unsure of policy, shortcuts)
Key Takeaway: “Institutional compliance exists to protect you and your research. Ask questions; that’s what support staff are for.”
Workshop Logistics
Teaching Considerations
Tone
This workshop covers compliance and responsibility, which can feel heavy. Balance with: - Normalize confusion: “These are legitimate questions” - Empower, don’t scare: “You CAN use AI responsibly” - Practical focus: “Here’s how to actually do this”
Audience Diversity
Learners will have very different AI experience: - Complete beginners: Haven’t used ChatGPT - Regular users: Already using AI in research - Power users: Asking about advanced use cases
Teaching Strategy: - Beginners: Focus on decision tree and basic concepts - Experienced: Discuss edge cases, nuances - Everyone: Understand Pomona’s specific policy
Common Pushback
“I’m not doing anything sensitive, so it’s fine to use ChatGPT” - Response: “Great! But how do you know what’s sensitive? Let’s classify first.”
“Everyone uses external AI with their data” - Response: “They might be violating policy. Our institution has specific rules to protect you.”
“Local AI is too slow” - Response: “Correct! Trade-off: slower but private. For public data, ChatGPT is fine.”
“This policy limits research” - Response: “Actually enables research. Using local AI for sensitive data is liberating, not limiting.”
Assessment and Feedback
Assess Learning Through:
- Scenario exercises: Can they classify data correctly?
- Decision tree: Can they navigate to right tool?
- Contact knowledge: Do they know whom to email?
- Practical application: Could they apply this to their own research?
Feedback Survey Items
1. Did you understand Pomona's data classification system?
(Very unsure / Somewhat unsure / Confident / Very confident)
2. Could you classify your own research data after this workshop?
(Yes / No / Not applicable)
3. Do you know when to use ChatGPT vs. local AI?
(Yes / No / Somewhat)
4. What question still remains about AI policy?
(Open text - helps identify follow-up needs)
5. Would you recommend this workshop?
(1-5 scale)Common Learner Questions and Answers
Q: “Can I use ChatGPT to analyze student feedback?” A: Only if you remove all identifiers. Better: Use a local model on Sagehen.
Q: “What counts as identifying?” A: Names, student IDs, course numbers, dates, assignment names. If you can trace back to an individual, it’s identifiable.
Q: “Does anonymization make it OK?” A: Not really. Safer to just use local AI instead.
Q: “What if I accidentally send FERPA data to ChatGPT?” A: Contact ITS immediately. The faster you act, the better. Don’t hide it!
Q: “Is our lab data proprietary?” A: Check with your advisor or department chair. When in doubt, ask ITS.
Q: “Can I use ChatGPT with data from collaboration?” A: Only if partner agrees. Check your NDA!
Q: “How do I know if research is export-controlled?” A: Government websites, funding agency, or ask Provost’s office.
Q: “Is a local model actually secure?” A: Yes, it’s on US infrastructure and doesn’t send data anywhere.
Follow-Up Resources
Post-Workshop Actions
Immediate
- Distribute slides and reference materials
- Provide ITS contact info
- Share local model setup guide
Notes from Previous Runs
(Update after each workshop)
What Went Well
- Real-world case studies helped people understand consequences
- Interactive scenarios made policy concrete
- Decision tree demystified complex policy
- Showing the local model demo increased confidence
Critical Reminders for Instructors
- You are not a lawyer: Refer complex legal questions to compliance office
- Policy changes: Confirm Pomona policy is still current before workshop
- Emphasize support: This isn’t about punishing violations; it’s about enabling research safely
- No judgment: People ask because they care about compliance; that’s good
- Document everything: When people ask questions, note them for policy updates