Event context:

• Organized by the Khoury UG Advisory Committee
• Panel with Prof. Bell and Associate Dean Christo Wilson
• Topic: AI in the CS curriculum and beyond
• This deck provides talking points for the 14 questions submitted

Framing:

• These are genuinely hard questions with no single right answer
• The goal is conversation, not lecture
• Students should feel empowered to push back and share experiences

→ Transition: Let's jump into the questions...

AI at Khoury College

An Undergraduate Student & Faculty Conversation

Prof. Jonathan Bell & Associate Dean Christo Wilson

Organized by the Khoury UG Advisory Committee

©2026 Jonathan Bell, CC-BY-SA

Format:

• 14 questions submitted by students and advisory committee
• Each question gets a brief framing, then open discussion
• No question is off-limits — push back, disagree, share your experience

→ Transition: Question 1...

Q1: What Is Khoury's Role in AI?

Are we primarily builders of AI systems, critics and auditors of them, or both?

Key points:

• Both — and the "both" is what makes a CS education distinctly valuable
• Building gives you technical depth to understand what systems actually do (pattern predictors, not oracles)
• Critical evaluation gives judgment to know when AI is appropriate, dangerous, or oversold
• Undergraduate education should develop both: technical foundations AND design/evaluation skills
• CS 3100 models this: learn to use AI effectively, but only after building foundational skills to evaluate what it produces

Discussion prompt:

• Do students feel they're getting both sides?
• Where are the gaps?

→ Transition: What makes Khoury different?

Q2: What Differentiates Khoury's AI Education?

What do we uniquely offer students that they couldn't get elsewhere?

Three differentiators:

1. Co-op program:

   • Students use AI tools in real professional settings earlier than almost anyone
   • That feedback loop between industry practice and academic learning is genuinely rare

2. Software engineering emphasis:

   • When code generation gets cheaper, bottleneck moves to integration, validation, maintenance
   • Exactly what CS 3100 and CS 4530 emphasize

3. Active curriculum research:

   • Gathering data on student experiences, iterating, publishing
   • Students aren't consumers — they're helping shape the curriculum

Discussion prompt:

• What have you seen on co-op that we should know about?
• What's working? What's missing?

→ Transition: Should AI be required for everyone?

Q3: Should AI/ML Be Required for All CS Majors?

If AI is becoming foundational, does it belong alongside systems and algorithms as core knowledge?

The case:

• Foundational AI/ML literacy — yes, increasingly required
• Understanding LLMs are pattern predictors, they hallucinate, have context windows, are non-deterministic — this is becoming baseline literacy

The question is WHAT to require:

• Full ML course with gradient descent? Probably not for everyone
• Conceptual foundations without full mathematical depth? Yes

Two approaches:

• Weave AI literacy into existing courses (as CS 3100 does)
• Create a lightweight required module covering conceptual foundations
• Challenge: already-packed curriculum

Discussion prompt:

• How much theory do you actually need to use these tools well?
• Would a required AI course help or feel like one more box to check?

→ Transition: What about students outside CS?

Q4: Should All University Majors Get AI Training?

Should there be required foundational AI courses beyond Khoury?

Key points:

• Yes, but "training course" might be the wrong framing
• AI literacy looks different in different fields
• What CS faculty uniquely offer: the TECHNICAL understanding of what these systems do, why they fail, how to evaluate outputs

Cross-disciplinary engagement:

• Computing programs should be AT THE TABLE helping other programs
• Not just building courses for our own majors
• Worst outcome: every department independently inventing AI curriculum without technical grounding

Discussion prompt:

• Have you seen other departments grappling with AI?
• What would you want a non-CS major to understand about AI?

→ Transition: How do we balance tools vs. theory?

Q5: Tools vs. Theory — What Comes First?

Should students first learn how transformers work, or how to effectively use them?

The "forklift in the gym" answer:

• If building foundational skills — do the work yourself first (forklift in the gym)
• Once you have foundations — AI tools are the forklift in the warehouse

Three distinct competencies:

Tier 1 — Using AI assistants (every graduate):

• Don't need transformer internals
• Need judgment to evaluate outputs, spot hallucinations, debug AI-written code
• This is what CS 3100 teaches

Tier 2 — Building AI-enabled software (growing share):

• Integrating AI into real products (summarization, classification, generation)
• Need to understand context windows, non-determinism, failure modes
• Like knowing databases well enough to design a schema without building a storage engine
• THIS IS THE BIGGEST GAP in current curricula

Tier 3 — Advancing AI itself (researchers):

• Requires deep theory: linear algebra, probability, optimization
• Graduate-school territory for most

Discussion prompt:

• Which tier are you most interested in?
• Are you getting what you need for Tier 2?

→ Transition: What skills matter most?

Q6: What Skills Matter Most in an AI-Driven Workforce?

Math depth? Systems thinking? Product intuition? Ethics? Something else?

The skills that matter most are the ones AI doesn't provide:

1. Evaluation and judgment — Can you tell if AI-generated code is correct, secure, performant?
2. Requirements and design thinking — What should we build? For whom? With what tradeoffs?
3. Communication and collaboration — As code generation accelerates, coordination becomes the bottleneck
4. Systems thinking — How components interact, where failures propagate, what happens at scale
5. Adaptability — The meta-skill: ability to learn new tools, evaluate them, integrate them

The common thread is JUDGMENT:

• The ability to evaluate, decide, and take responsibility
• Math depth matters if building AI systems
• Ethics awareness matters for everyone

Discussion prompt:

• Which of these skills do you feel strongest/weakest in?
• Are there skills you think are missing from this list?

→ Transition: How do we prevent over-reliance?

Q7: How Do We Prevent Over-Reliance on AI Tools?

Where should the line be drawn in coursework?

CS 3100's three-part approach:

1. Sequence matters:

   • Restrict AI on early assignments — not because AI is bad
   • Students haven't yet built skills to review AI-generated code
   • "Task familiarity quadrant" — if you can't evaluate output, don't use AI yet

2. Be explicit about learning goals:

   • Each assignment specifies AI usage level (Not Allowed → Permitted → Required)
   • When students understand WHY, they engage authentically
   • Prohibition with rationale = coaching; without = arbitrary

3. Assess the process, not just the product:

   • Proctored exams, oral walkthroughs, self-reflections
   • If AI can produce the artifact, assess something AI can't fake: understanding

The line isn't fixed — it moves as students build competence.

Discussion prompt:

• Do AI restrictions feel arbitrary or well-explained in your courses?
• What assessment methods actually measure YOUR understanding?

→ Transition: How is AI changing software itself?

Q8: How Is AI Changing Software Engineering?

Are we training students for a future of less writing and more reviewing?

Yes — and this shift was already happening:

• Code review became first-class with GitHub PRs (2008)
• Developers already spent more time READING than WRITING
• AI accelerates this trend

What's changing:

• More code, faster — Jevons' paradox: cheaper production → more production
• Bottleneck shifts — Is it correct? Does it integrate? Will it be maintainable?
• Code review becomes more critical — Reviewing code from a model that's seen millions of codebases but doesn't understand YOUR project

This is why CS 3100 emphasizes:

• Design for change, specifications, testing, stakeholder analysis
• These skills remain hard regardless of how code is produced

Discussion prompt:

• On co-op, have you seen this shift in how teams work?
• How much time do you spend reading vs. writing code?

→ Transition: What new challenges are emerging?

Q9: What New Engineering Challenges Does AI Create?

Reliability, testing, reproducibility, long-term maintenance?

Six emerging challenges:

1. Non-determinism as a feature — Same prompt, different results. Breaks traditional testing.
2. Testing and validation — How do you test "generate a good code review"? Easy tasks have clear yes/no; hard tasks need user testing, expert review, time.
3. Maintenance of AI-assisted codebases — If developers accepted code they don't understand, future maintainers inherit that gap.
4. Supply chain and provenance — AI-generated code may reproduce training data patterns. Licensing? Attribution?
5. Reproducibility — Model updates change behavior. Code that "worked" may break with next model version.
6. The "vibe coding" failure mode — Only evaluating execution, not code. Brittle systems that collapse when requirements change.

These are genuinely new challenges without established best practices — exciting time to study SE.

Discussion prompt:

• Have you encountered any of these on co-op or in projects?
• Which feels most concerning to you?

→ Transition: Who is responsible when AI causes harm?

Q10: Who Is Responsible When AI Systems Cause Harm?

Developer? Company? Researcher? Institution?

The developer who ships the code is responsible — full stop.

• "The AI wrote it" is not a defense
• The Documentation Test: can you justify the decisions? understand what it does? take responsibility if wrong?

But responsibility is layered:

• Developer: Evaluates and validates before shipping
• Company: Establishes review and validation processes
• Model provider: Transparency about capabilities and limitations
• Institution: Teaches evaluation skills for responsible use

Historical parallel (from L1):

• Grace Hopper faced skeptics about compilers: "Machines can't write programs"
• Compilers didn't eliminate programmer responsibility
• Same pattern: AI generates, humans validate and take responsibility

Course connection:

• "You are responsible for any material you turn in"
• "You must understand all submitted work"
• The principle doesn't change because the tool changed

Discussion prompt:

• Where should the line be drawn when AI gets something wrong?
• How does this change if AI is making medical or legal decisions?

→ Transition: How do you get into AI research?

Q11: How Do Undergraduates Get Into AI Research?

What's the first concrete step?

First concrete step:

• Talk to a professor whose work interests you
• NOT a cold email with "I'm interested in AI"
• Read one of their recent papers, come with a specific question
• Most faculty have more research ideas than hands

Beyond that:

1. Build strong foundations — Linear algebra, probability, algorithms, systems. Best researchers have strong fundamentals.
2. Don't just follow the hype — Some of the most impactful work is in evaluation, testing, safety, understanding failures — not just making models bigger.
3. Build things and break things — Implement papers from scratch. Reproduce benchmarks. The gap between reading about AI and building AI is where learning happens.
4. Take CS 3100 seriously — Design, testing, evaluation skills are directly applicable to AI research.

Discussion prompt:

• What research areas interest you?
• What's stopping you from reaching out to faculty?

→ Transition: How do you stand out in a crowded field?

Q12: How Do You Stand Out When Everyone Lists "AI"?

What signals depth over trend-chasing?

What signals depth:

1. Show what you built and WHY — Not "fine-tuned GPT" but "identified this problem, evaluated three approaches, chose this because of these tradeoffs, learned from failure modes"

2. Demonstrate evaluation skills — Can you critically assess an AI system? Identify when it's wrong? Rarer and more valuable than ability to prompt.

3. Go deep on fundamentals — Strong systems knowledge, solid math, debugging complex systems. Fundamentals age well; framework-of-the-month doesn't.

4. Co-op experience — Khoury's superpower. Real experience with production AI systems is worth more than tutorials.

5. Contribute to open source — Shows you can work in real codebases, collaborate, navigate constraints.

6. Be honest about AI usage — "I used Copilot for this part but verified by doing X and Y" shows professional judgment.

Discussion prompt:

• What's on your resume right now? Does it show depth or surface?
• What's one thing you could do this semester to build a depth signal?

→ Transition: Is grad school necessary?

Q13: Is Graduate School Becoming Necessary for AI?

Can undergraduates still break into meaningful AI roles directly?

Undergraduates can absolutely break into meaningful AI roles.

• Demand for people who can USE AI effectively far exceeds demand for people who can BUILD new models
• The latter usually requires grad school; the former requires strong engineering + judgment

The key question:

• Do you want to BUILD WITH AI or ADVANCE AI?
• Both valuable, different preparation

Economic reality:

• Jevons' paradox: AI makes tasks cheaper → organizations pursue MORE ambitious projects → MORE demand for engineers
• Integration, evaluation, design, collaboration — strong BS with co-op positions you well

When grad school makes sense:

• Fundamental AI research — new architectures, training methods, safety
• Provides depth and time that's hard to get otherwise
• Strong undergrads can contribute, especially on evaluation and testing

Discussion prompt:

• Are you considering grad school? Why or why not?
• What roles are you targeting after graduation?

→ Transition: Final question — if you were starting over...

Q14: If You Were an Undergraduate Again Today...

What would you focus on — and what would you ignore?

Double down on:

• Fundamentals — Data structures, algorithms, systems, networking. Don't expire. When tool landscape shifts, fundamentals transfer.
• Reading and understanding code you didn't write — Becoming THE core skill. Career = reading, reviewing, evaluating, modifying code.
• Communication — Writing clearly, presenting, feedback, teamwork. AI makes this MORE important as coordination bottleneck grows.
• Building complete systems — End-to-end, deploy, maintain, fix bugs in production. "Works on my laptop" → "works for users" is where engineering lives.
• Co-ops — Use Khoury's program aggressively. Nothing replaces real stakes with real stakeholders.

Ignore:

• Framework of the month — Tools change. Learn one well, understand principles, pick up the next quickly.
• The hype cycle — Focus on what systems actually do (pattern prediction), not what marketing says.
• The anxiety — "Will AI replace programmers?" has been asked about every productivity tool since compilers. Answer: no, but it changes what we do. Build durable skills and you'll adapt.

Discussion prompt:

• What are you doubling down on?
• What anxiety do you want to talk through?

→ Transition: Closing slide...

Bonus: "We'll Solve All Your Software Problems"

Dario Amodei says software engineering is obsolete in 12 months. Where have we heard this before?

The pattern repeats every 10-15 years:

1. 1960s-70s Mainframes: IBM promised centralized computing would handle everything. Created vendor lock-in, COBOL armies, 6-month lead times.

2. 1980s PCs: "Empower every worker!" Created shadow IT, incompatible spreadsheets, "my Access database runs the company."

3. 1990s-2000s ERP: SAP, Oracle, PeopleSoft — "one system to rule them all." Multi-year implementations, billions over budget. And of course, Northeastern's Banner → Workday migration — a system that worked (if creakily) replaced by a shiny new system that... well, students and staff have opinions.

4. 2000s-2010s SaaS: Salesforce's "No Software!" campaign. Now organizations drown in 50+ SaaS subscriptions that don't talk to each other.

5. 2010s-2020s No-Code: Monday.com, ClickUp, Airtable — "anyone can build software!" Until you need it to scale, integrate, or handle edge cases.

6. 2025 AI: "Software engineering obsolete in 12 months." Software stocks tumble. Same pattern, shinier beans.

The punchline:

• Each wave DOES make things better — AND more complex
• Mainframes were real. PCs were real. Cloud was real. AI is real.
• But none eliminated the need for people who understand requirements, evaluate systems, maintain them, and take responsibility
• The tools change. The hard problems don't.

Amodei's specific claim:

• Software stocks dropped on his statement (ServiceNow, Salesforce, Workday all fell)
• Worth noting: Anthropic sells AI. Of course the CEO says AI replaces everything.
• Same energy as Salesforce saying "No Software" while being... a software company.

Discussion prompt:

• Has anyone lived through the Workday migration? (knowing laughs)
• Does the pattern make you more or less anxious about AI?

→ Transition: Closing...

Thank You — Let's Keep This Conversation Going

The through-line of every answer tonight:

• Technology changes. Responsibility doesn't.
• The hardest parts of building software have never been typing code.
• Build the judgment that makes tools valuable — then use the tools aggressively.

Stay connected:

• These conversations continue in CS 3100 — AI is woven throughout the rest of the semester
• Your survey responses and feedback directly shape what we teach next
• The UG Advisory Committee is your voice — use it

Closing message:

• These are hard questions with no single right answer
• The fact that students are asking them is a sign of the seriousness with which this community takes AI
• Faculty don't have all the answers — but we're committed to working through them together
• The conversation doesn't end when you leave this room

Call to action:

• Fill out the surveys in Lab 6
• Come to office hours to continue the conversation
• Push back when something doesn't make sense
• Share what you're seeing on co-op — you're on the front lines