Phase Three Course Evidence Review

California Education and AI Course Evidence Review

Phase Three UC Berkeley and Stanford University

Main finding

The strongest course designs match AI rules to the mental work students are expected to develop. Writing courses protect original reading and composition. Quantitative courses permit explanation or feedback while requiring students to solve and verify. Technical courses introduce AI after students build foundational mental models, or permit it for selected coding tasks while retaining project, presentation, and examination requirements. Courses that study AI directly use it as an object of ethical, civic, and practical judgment.

A prohibition can protect productive struggle, but it does not prove learning. Permission can support access and practice, but it does not prove that students performed the reasoning. The most credible designs combine a clear purpose, task-specific boundaries, disclosure, source verification, visible drafts or work, and an independent check of understanding.

The course pages provide considerable evidence of policy and practice. They provide little comparative outcome evidence. No reviewed course publishes a controlled comparison showing that its AI approach improves durable learning, independent judgment, mental fitness, or wellness.

Scope and evidence standard

This phase examines eight publicly documented course examples from 2025 and 2026. The sample is intentionally varied and is not statistically representative of either university. Each case is evaluated as a stated rule, an implemented practice, a method of checking understanding, a human-development indicator, or a measured result.

• AI prohibited The student performs the relevant reading, reasoning, writing, or assessment without generative AI.
• AI bounded AI is allowed for named purposes but barred from the intellectual act being assessed.
• AI staged Students first build foundational knowledge or mental models, then learn to use AI.
• AI integrated AI is part of the course subject, production process, or required learning experience.

Course sample at a glance

What the sample establishes

• Course-level policies differ materially within each university because the intellectual work differs by discipline and assignment.
• Several instructors explicitly identify writing, recall, problem solving, discussion, and productive struggle as capacities AI can weaken when it substitutes for the learner.
• In-person quizzes, oral explanations, supervised problem solving, presentations, and process documentation provide stronger evidence of individual understanding than an unsupervised final product alone.
• The reviewed pages usually describe intentions and assessment structures, not validated learning or wellness effects.

What the sample does not establish

These examples cannot determine the prevalence of each approach, the degree of compliance, or whether one policy causes better results. Public pages also tend to feature unusually explicit instructors and innovative courses. Their practices should not be treated as the ordinary experience of every Berkeley or Stanford student.

UC Berkeley course evidence

SLAVIC 134E Anton Chekhov

The spring 2026 writing-intensive literature course states that AI use is prohibited. The subject requires students to interpret literary form, historical context, and questions about human experience. The prohibition therefore protects the student’s encounter with the text and the formation of an individual interpretation. The public listing does not explain enforcement, assessment design, or measured results.

• Strength The rule is unambiguous and aligned with original reading and interpretation.
• Limit A ban without published process evidence does not show whether students comply or learn more.

Stat 20 Introduction to Probability and Statistics

Stat 20 permits AI for conceptual help and coding errors but bars students from feeding complete assignments to a chatbot or copying generated output. The policy explains the cognitive reason: recognizing an answer is easier than producing it, so AI-assisted homework can create an illusion of understanding that fails under independent testing. Students are told to attempt problems first, use AI for explanations rather than solutions, verify outputs against course materials or people, and develop the final solution themselves.

The course uses in-person work, individual submissions, quizzes, quiz retakes, and a final examination. It also emphasizes small-group work, tutoring, friendship, and a manageable workload in which much work is completed during class. These practices support learning and social connection. The syllabus reports instructor observations about students who copied AI-supported solutions and later failed similar quiz questions, but it does not publish systematic course-level results.

• Strength The policy teaches students why cognitive offloading can impair learning and gives specific alternatives.
• Strength Independent assessments test whether students can produce answers without AI assistance.
• Limit The reported pattern is an instructional observation rather than a controlled outcome study.

Stat 133 Concepts in Computing with Data

Stat 133 asks students to refrain from AI until the course introduces it. The stated reason is that students need firm mental models of code before they can use AI well. AI may provide feedback on writing, and the course later teaches AI coding tools. The assessment structure places 49 percent of the grade on quizzes, 30 percent on a paper final, and 20 percent on project studios where students discuss their reasoning. Problem sets and projects function mainly as practice.

• Strength AI use follows foundational knowledge instead of replacing its formation.
• Strength Oral project studios and individual examinations make student understanding visible.
• Limit The public syllabus does not compare performance before and after AI instruction.

CS 184 284A Computer Graphics

The spring 2026 course permits AI for programming assignments and final projects and encourages students to write code before using AI as a coach. It acknowledges that reliable policing is difficult and encourages experimentation with specialized coding tools. Students also complete programming assignments, team projects, presentations, participation, and examinations. This is the sample’s most permissive production policy.

• Strength The policy recognizes professional tool use and asks students to retain responsibility for the work.
• Risk Permission for both code and final-project writing can make individual contribution difficult to distinguish without process records or oral validation.
• Limit No public comparison shows whether permissive use changes mastery, originality, workload, or integrity.

Berkeley assessment

Stanford course evidence

PWR 1 Writing and Rhetoric

A published PWR 1 example requires students to read sources and write their own language without generative AI. Limited use may be allowed while exploring a research topic. The policy connects human reading and writing to self-understanding, expanded thought, perspective taking, empathy, and participation in democracy. It also commits the instructor to create course materials and provide feedback without AI assistance.

• Strength The policy protects writing as a process of thought and preserves reciprocity between student and instructor.
• Strength The rule explains the human capacities at stake instead of relying only on an integrity prohibition.
• Limit The published example does not include comparative learning, voice, empathy, or wellness results.

CS 389 Internet Censorship

CS 389 prohibits generative AI for paper reviews while encouraging it for coding assignments. This division protects direct reading, synthesis, discussion preparation, and written interpretation while permitting tools for technical production. Students also lead discussions, deliver presentations, complete an original research project through staged milestones, and include ethics and limitations in the final paper.

• Strength The policy draws its boundary around the kind of thinking each task is meant to develop.
• Strength Presentations, discussion, milestones, and a final paper expose the student’s reasoning process.
• Limit The course does not publish learning or integrity results by type of AI use.

COLLEGE 102 Citizenship in the 21st Century

More than 1,200 first-year students in over 70 sections considered how Stanford should promote learning, reduce cheating, and assess understanding in the AI era. Students drafted 109 policy memos, deliberated in small groups and sections, voted across sections, and presented the most-supported proposal to university leaders. The winning proposal recommended brief in-person understanding checks, oral explanations, supervised problem solving, and reduced weight for other written assignments so that verification would not simply add workload.

This course used AI policy as an exercise in citizenship, perspective taking, negotiation, and institutional judgment. The number of participants, memos, votes, and public presentation are implementation results. They do not measure whether critical thinking, civic judgment, or AI literacy improved.

• Strength Students practiced the civic and intellectual capacities the course aimed to teach.
• Strength The proposed assessments address comprehension while recognizing student workload and anxiety.
• Limit Selection of a winning memo is not a validated measure of individual learning.

AI for Human Flourishing

The 2026 Graduate School of Business course asked how AI might support lives of growth, purpose, health, impact, and depth. Forty MBA students used AI coding agents in a nearly five-hour buildathon, created applications addressing personal needs, wrote individual essays, met practitioners, and concluded with small-group discussion among 40 students and 40 guests. The reported final conversation lasted about an hour and was notable for sustained attention without phone use.

• Strength AI use is subordinated to explicit questions about purpose, human needs, judgment, creativity, and connection.
• Strength Building, reflection, and face-to-face discussion require more than passive consumption of AI output.
• Limit Project examples and student quotations are descriptive evidence, not measured flourishing or mental-fitness outcomes.

Stanford assessment

Comparison against the well formed mind standard

Mental fitness and wellness

The sample supports a practical definition of mental fitness in education: the capacity to sustain attention, retrieve knowledge, tolerate productive difficulty, explain reasoning, revise one’s judgment, take responsibility, and remain connected to other people. Several courses deliberately exercise these capacities. This is stronger evidence than a generic wellness statement, but it remains evidence of design rather than measured psychological improvement.

Course evaluation should distinguish healthy reduction of needless burden from removal of effort that builds capacity. It should also test whether AI changes confidence, anxiety, attention, isolation, help-seeking, belonging, and the quality of interaction with instructors and peers. These outcomes may differ sharply among students with stable support, students whose family or social relationships are strained, and students still experiencing effects associated with pandemic disruption.

COVID era comparison

The reviewed post-pandemic courses combine restored in-person teaching with continued online practices. Stat 20 and Stat 133 emphasize physical classrooms, tutoring, oral explanation, and independent examinations. CS 184 284A retains online lectures, discussions, and office hours. Stanford examples emphasize live discussion, presentations, and supervised understanding checks. This variation permits a useful comparison of educational conditions, but the public sources do not compare equivalent cohorts before, during, and after COVID-19.

Future analysis should test whether direct contact, regular routines, family and friendship support, and immediate human feedback help students preserve learning and wellness when education is disrupted. It should also examine whether students who entered the pandemic with fragile mental health, dysregulated family life, or limited friendships experienced greater and more persistent losses than students with stronger support. COVID history should remain an explicit comparison factor, not an assumed explanation for every present outcome.

Evidence based course standard

A course that uses AI while seeking to form independent judgment should do the following.

• Name the human capacity each assignment is designed to develop.
• State permitted and prohibited AI uses by assignment, with examples and a reason for each boundary.
• Require an unaided first attempt when retrieval, reasoning, interpretation, or writing is the learning objective.
• Require disclosure of tools, prompts, outputs, sources, and the student’s use of them when AI is permitted.
• Preserve direct reading, original writing, live discussion, and independent problem solving.
• Use drafts, oral explanation, in-person work, or supervised checks to make understanding visible.
• Measure learning, retention, integrity, workload, attention, confidence, belonging, and differential effects over time.

Conclusion and next investigation

Course practice confirms that AI policy is most educationally sound when it protects the precise human work a student must learn to perform. Berkeley provides especially clear examples of productive struggle, staged competence, and independent assessment. Stanford provides especially clear examples of human writing, task-specific rules, civic deliberation, and reflection on flourishing. Neither university yet supplies enough public outcome evidence to determine which approach best forms durable judgment or supports mental fitness.

The next investigation should examine published empirical research on cognitive offloading, memory, writing, AI tutoring, academic performance, integrity, and student wellness. It should then use that research to grade the course practices identified here by strength of evidence.

Public sources reviewed

Official institutional sources accessed September 12, 2026. Course policies can change by instructor and term.

• UC Berkeley SLAVIC 134E class listing
• UC Berkeley Stat 20 fall 2026 syllabus
• UC Berkeley Stat 133 fall 2026 syllabus
• UC Berkeley CS 184 284A spring 2026 policies
• UC Berkeley course policy guidance and instructor examples
• Stanford clear AI policies and course examples
• Stanford CS 389 Internet Censorship
• Stanford COLLEGE 102 course report
• Stanford COLLEGE 102 winning policy memo
• Stanford AI for Human Flourishing course report
• Stanford AI and your learning guide