Claude Research #2: CS Curricula Analysis

Claude Research #2: CS Curricula Analysis

Category: research
Date: 15 February 2026, 17:36 UTC
Original File: CLAUDE_RESEARCH_2_CS_CURRICULA.md

26 universities compared - GenAI in CS education

Generative AI in CS Curricula: 26 Universities Compared

Claude Deep Research Output #2 - Analysis

Source: Claude.ai Artifact (0ca9705a-cc1b-4cdc-b4f3-6086418a68a1)
Date: 2026-02-15
Researcher: Thomas Lancaster (via Claude Deep Research)
Topic: Prompt B - GenAI in CS Education
Universities Analyzed: 22 Russell Group UK + MIT, Stanford, CMU, ETH Zurich

EXECUTIVE SUMMARY

Key Finding: No university has banned GenAI outright, but a “performance-learning paradox” has emerged:

• Students using AI complete tasks faster and score higher on homework
• But perform measurably worse on exams (2 letter grades lower at CMU)
• Stanford reported: CS problem set scores increased, exam scores declined

Central Driver: This paradox is forcing curriculum redesign across the sector.

CRITICAL FINDING: The Performance-Learning Paradox

Stanford (November 2025 Faculty Senate)

• CS problem sets: Scores increased with AI use
• Midterm/final exams: Scores declined
• VP for Undergraduate Education reported this pattern

Carnegie Mellon (Course 15-122)

• Students using AI received 2 letter grades lower on average
• Spring 2025: Highest drop rate in years
• Instructor Michael Taylor: “Every change we made somehow all tied back to AI”

Implications:

• AI assists immediate task completion
• But undermines deep learning
• Students fail when AI unavailable (exams)

UK RUSSELL GROUP ANALYSIS (22 Universities)

Framework: All Follow 5 Shared Principles

Russell Group Principles (July 2023, updated Feb 2025):

• Promote responsible use
• Not prohibition
• Tiered/categorical systems
• Module-level decision making

DETAILED INSTITUTION PROFILES

University of Cambridge

Policy Date: 15 October 2024 (CS Department-specific)

Approach: Most detailed publicly available CS policy

Key Provisions:

• ✅ AI prohibited in ALL examinations (regardless of location/format)
• ✅ Local editing tools permitted (grammar, spell-check)
• ✅ Viva voce mechanism: Suspected students must demonstrate work is their own
• ❌ AI-generated output never presented as student’s own work unless explicitly permitted

Unique Feature: Oral examination option for verification

University of Edinburgh

School of Informatics

Approach: Most restrictive default among UK institutions

Key Provisions:

• ❌ Default: GenAI systems MUST NOT be used for assessed work
• ✅ Unless explicitly allowed in writing by course organiser
• Infrastructure: Built ELM (Edinburgh Language Models)
  • Zero-data-retention agreement with OpenAI
  • Integrated with JupyterAI for coding instruction

Unique Feature: Institutional AI platform with privacy protection

University of Oxford

First UK Uni to Provide ChatGPT Edu

Timeline: September 2025

• Member of OpenAI’s NextGenAI consortium
• GPT-5 model access for all staff/students

Policy:

• Every assessment must explain AI permissions
• Unauthorized use = academic misconduct
• No CS department-specific policy - follows university-wide guidance

Gap: No publicly available CS-specific guidance

Imperial College London

Most Sophisticated AI Infrastructure

Platform: dAIsy (custom-built institutional platform)

• Multiple models: GPT-5, Claude, DeepSeek
• Single secure interface

Courses:

• “Introduction to AI-Assisted Programming” - doctoral students
• GitHub Copilot training
• Mandatory: “Introduction to Generative AI” online course for all students

Tool: Assessment “stress test” for staff

• Evaluate how GenAI affects specific assessment tasks

Unique Feature: Comprehensive institutional platform + mandatory training

UCL

Three-Category Framework

Critical Stance:

• Does NOT use GenAI detectors when marking
• Shared by most Russell Group institutions

University of Bristol

Four-Category System

Research: Funding AI-generated feedback for first-year programming

University of Birmingham

Most Restrictive University-Wide Default

• ❌ AI use NOT permitted unless explicitly stated
• ✅ GenAI proofreading restricted to spelling/grammar only

King’s College London

Four Levels of AI Use

From routine tools → AI-integral assessment

Advice: “Against integrating generative AI into most summative assessments”

University of Manchester

World’s First: Microsoft 365 Copilot to 65,000

• Entire community (students + staff)
• Free GitHub Copilot for all

University of Leeds

Traffic-Light System

• 🟢 Green: AI permitted
• 🟡 Amber: Limited use
• 🔴 Red: Prohibited

Investigation: Student newspaper found 17% admitted using AI in “Red” assignments

University of Glasgow

Explicit Code Prohibition

“Generating code or generating code comments” prohibited in restricted assessments

University of Exeter

Four-Tier System

• AI-integrated
• AI-assisted
• AI-limited
• AI-prohibited

Feature: Mandatory compliance checkbox at submission

Queen’s University Belfast

RAISE Framework (bespoke)

• Responsible use
• AI best practice
• Integrity
• Support
• Equitable access

CRITICAL GAP ACROSS UK SECTOR

Finding: None of 22 Russell Group CS departments have publicly available, department-specific GenAI policies

• All operate under university-wide frameworks
• Devolved decision-making to module level
• Internal VLEs/module handbooks contain guidance (not public)

INTERNATIONAL INSTITUTIONS

MIT

Course-Level Experimentation (No Department-Wide Policy)

Course Examples:

Math Department Request: All courses include in-person assessments (AI “big factor”)

Infrastructure:

• ChatGPT Enterprise for faculty
• OpenAI NextGenAI consortium member

Stanford

Most Structured Institutional Response

Academic Integrity Working Group (AIWG):

• Proctoring pilot: 3rd year, 50+ courses
• Found AI detectors unsuitable for misconduct evidence (high false positives)

Platform: AI Playground

• ChatGPT (GPT-5)
• Claude (4 Sonnet, 4.5 Sonnet)
• Gemini, DeepSeek, Llama

CS106A (Chris Piech):

• ❌ AI code completion prohibited (must disable in PyCharm)
• ✅ Students build AI-powered apps in “Infinite Story” assignment
• Teaching to program WITH GenAI as component, not crutch

Graduate School of Business:

• Instructors may NOT ban AI for take-home coursework

Data (Nov 2025 Faculty Senate):

• Problem set scores increased
• Exam scores declined

Carnegie Mellon (CMU)

Ranked #1 Globally in AI

Default: GenAI = “unauthorized assistance” unless explicitly permitted

Eberly Center: 6 template syllabus policies

• From full prohibition → full encouragement

Course 15-112 (Intro):

• Highest drop rate in years (Spring 2025)
• Students using AI heavily failed quizzes/exams

Instructor Response (Michael Taylor):

• Increased quiz weighting
• In-class homework rewrites (live coding)
• Flipped classroom formats
• Quote: “Every change we made somehow all tied back to AI”

New Course: 15-113 “Effective Coding with AI”

• Build projects using AI tools strategically
• Evaluate code for correctness, security, maintainability

Provocative Course: “Students will not write actual code”

• Manage AI-assisted development
• Tools: Cursor, Windsurf
• Method: “Mob programming” sessions

ETH Zurich

“Proactive Approach” Emphasizing Responsible Use

• Developing institutional frameworks
• Emphasis on responsible integration
• (Content truncated in source)

COMPARATIVE ANALYSIS

Assessment Restructuring Trends

Infrastructure Investment

Policy Spectrum

IMPLICATIONS FOR RESEARCH

High-Value Research Questions:

1. Performance-Learning Paradox Replication

   • Test at other institutions
   • Longitudinal tracking
   • Causal mechanisms

2. Assessment Redesign Efficacy

   • Which changes work?
   • Cost-benefit analysis
   • Student learning outcomes

3. Infrastructure Comparison

   • Which platforms most effective?
   • Privacy vs functionality
   • Cost analysis

4. Policy Compliance

   • 17% at Leeds used AI in “Red” assignments
   • Actual vs stated policy adherence
   • Enforcement mechanisms

5. Department vs University Policies

   • Gap in publicly available CS-specific guidance
   • Module-level variation
   • Best practice identification

Methodological Opportunities:

1. Scrape and Compare - Policy document analysis (your method)
2. Survey Faculty - Implementation experiences
3. Student Interviews - Understanding of policies
4. Longitudinal Tracking - Grade patterns pre/post AI

CONTENT CREATION OPPORTUNITIES

Twitter/X Threads:

1. “The AI Paradox: Higher homework scores, lower exam scores”
2. “26 universities, 26 approaches: How CS is handling GenAI”
3. “Why Oxford gives all students GPT-5 (and what happened)”
4. “CMU’s radical experiment: A course where you don’t write code”

Blog Posts:

1. “The Performance-Learning Paradox: Data from 26 Universities”
2. “Comparing CS GenAI Policies: From Prohibition to Integration”
3. “Infrastructure Wars: Which University Built the Best AI Platform?”

Academic Paper:

1. “The Performance-Learning Paradox in CS Education: A Multi-Institutional Analysis”
2. “Policy Responses to GenAI: A Comparative Study of 26 Leading Universities”

TEAM ACTION ITEMS

Immediate (This Week):

• Riley: Verify key statistics with original sources
• Ava: Create comparison matrix (institution × policy features)
• Quinn: Draft “Performance-Learning Paradox” Twitter thread
• Kai: Design infographic showing policy spectrum

Short-term (Next 2 Weeks):

• Riley: Find full text of Cambridge, Imperial, CMU policies
• Ava: Quantitative analysis of policy patterns
• Quinn: Blog post “26 Universities, One Paradox”
• Zak: Propose conference paper structure

Medium-term (Conference):

• Replicate performance-learning analysis
• Survey UK CS departments on implementation
• Design assessment redesign efficacy study
• Prepare live demo dataset

CONNECTION TO YOUR PRIOR WORK

Building on:

• Lancaster (2016) - Contract cheating handbook (assessment redesign)
• Lancaster & Clarke (2007) - Detection process (policy analysis)
• Lancaster (2019) - Student reasons (understanding why students use AI)

New Contribution:

• First multi-institutional CS-specific analysis
• Performance-learning paradox documentation
• Infrastructure/platform comparison
• Real-time policy tracking (2024-2026)

SUMMARY

This research reveals a sector in rapid transformation:

• No outright bans anywhere
• Universal concern about the performance-learning paradox
• Massive infrastructure investment (custom platforms)
• Assessment restructuring as primary response
• Gap between policy and practice (17% non-compliance at Leeds)

Central Challenge: Students appear to be learning less while performing better on AI-assisted tasks—a fundamental threat to educational integrity.

Saved: 2026-02-15
Status: Ready for analysis, content creation, and research development
Next: Team begins content creation and comparison analysis

Original file: CLAUDE_RESEARCH_2_CS_CURRICULA.md