Phase Two Evidence Review

California Education and AI Evidence Review

Phase Two UC Berkeley and Stanford University

Preliminary finding

UC Berkeley and Stanford University have extensive public AI guidance, approved tools, teaching resources, research programs, and academic-integrity safeguards. Both institutions explicitly preserve human responsibility for judgment and verification. Stanford expresses this principle through a human-centered AI literacy framework. Berkeley applies it through risk-based use rules, transparent procurement, assignment redesign, and restraint toward unreliable AI-detection systems.

The strongest public evidence concerns governance and implementation capacity. Neither institution has yet published campus-wide measured evidence showing that generative AI improves durable learning, independent reasoning, mental fitness, relationships, or student wellness. Berkeley is piloting revised critical-thinking and information-literacy measures. Stanford is funding AI-related course and research projects and evaluating use of its campus AI tools. Those efforts are promising, but results should not be assumed before publication.

Compared with Thomas Aquinas College and Mother of Divine Grace School, Berkeley and Stanford have stronger operational AI systems. The Phase One institutions state a more unified account of the kind of person education should form. Berkeley and Stanford cultivate many of the same capacities, but their large, decentralized curricula leave coherence to programs, instructors, and student choices.

Review method

This review examines official university pages and published institutional guidance available through September 12, 2026. It separates four levels of evidence so that a policy statement is not mistaken for a measured educational result.

• Institutional claim A stated aim, principle, policy, or public description.
• Implemented practice A course, approved tool, pilot, assessment method, training resource, or operating rule.
• Measured result Published evidence showing a learning, integrity, wellness, or organizational outcome.
• Unresolved gap Information that public sources do not establish and should remain a research question.

UC Berkeley

Educational formation

Berkeley describes undergraduate education as broad and deep, with critical inquiry, discovery, and student choice across disciplines. Its published materials support critical thinking, communication, collaboration, ethical inquiry, and information literacy. These are components of a well-formed mind, although Berkeley does not organize every undergraduate program around one common philosophical or theological account of human formation.

• Institutional claim Undergraduate education should combine disciplinary depth with broad intellectual exploration.
• Implemented practice Colleges, majors, general-education requirements, Berkeley Discovery, and teaching programs provide several routes to those capacities.
• Unresolved gap Public sources do not show whether students experience these elements as one coherent formation program or how consistently results are measured across the university.

AI perspective and educational programs

Berkeley treats AI as an educational, research, administrative, ethical, and public-policy matter. The campus AI Hub gathers tools, training, policies, courses, and opportunities. AI instruction appears in undergraduate and graduate courses, student-led courses, computing and data science, information, law, public policy, human rights, and executive education. The breadth is substantial, but AI literacy is not presented as a common requirement for every student.

• Implemented practice The AI Hub and its teaching directory make AI study and institutional guidance visible in one place.
• Implemented practice Berkeley provides licensed AI tools with stated availability and data-protection limits.
• Unresolved gap Course participation and program listings do not establish whether all graduates can evaluate AI claims, risks, and appropriate uses.

Policy, privacy, and approved tools

Berkeley permits AI use within existing privacy, security, accessibility, intellectual-property, procurement, and academic rules. Its guidance prohibits completing academic work with AI when an instructor has not authorized that assistance. Uses involving grading, assessment, legal advice, employment, discipline, or other high-consequence decisions require additional consultation. Licensed tools receive data-protection classifications, and campus guidance favors contracted services over unreviewed public products.

• Implemented practice BearGPT, the Campus AI Sandbox, ChatGPT EDU, Gemini, and Copilot are listed with different eligibility and protection conditions.
• Implemented practice The Campus AI Sandbox supports controlled access to multiple models and is intended to inform equity, cost, risk, and sustainable-service decisions.
• Unresolved gap Tool availability and protection levels vary. A license does not show that use improves learning or that classroom practice is consistent.

Human judgment and academic integrity

Berkeley advises instructors to begin with the learning objective and redesign assignments so students make their reasoning visible through drafts, checkpoints, critique, disclosure, and reflection. The campus did not enable Turnitin AI detection university-wide after a 2023 pilot found inconsistent results. It also warns that uploading student work to unapproved detectors may create privacy, copyright, and education-record risks.

• Implemented practice Teaching guidance distinguishes recall, evaluation, synthesis, argument, and other aims before deciding whether AI belongs in an assignment.
• Implemented practice Berkeley is testing revised critical-thinking and information-literacy rubrics and a student reflection tool in courses.
• Measured result Public pages describe the research design and pilots, but do not yet report a conclusive campus-wide effect of AI on learning.

Mental fitness and COVID-era learning

Berkeley monitors student experience and well-being through institutional efforts such as Student Pulse Surveys. COVID-19 forced a rapid transition to remote instruction in spring 2020, followed by continued remote and flexible arrangements. Institutional research has also examined how home environments affected first-generation and low-income students. These sources establish disruption and inequity as relevant conditions, but they do not isolate the effect of current AI use on attention, resilience, relationships, or mental fitness.

• Research implication Compare present conditions with both the pre-COVID baseline and the emergency remote-learning period when credible measures are available.
• Evidence limit Do not attribute current student wellness or learning outcomes to COVID, remote schooling, family conditions, or AI without an appropriate comparison and supporting data.

UC Berkeley evidence status

Evidence gaps requiring continued investigation

• Whether every student receives explicit AI literacy and practices unaided reasoning before using AI assistance.
• Whether course-level AI policies produce materially different expectations across departments and instructors.
• Whether AI use changes retention, writing quality, independent judgment, attention, misconduct, or instructor workload over time.
• Whether outcomes differ by disability, income, first-generation status, language background, or access to approved tools.

Stanford University

Educational formation

Stanford describes undergraduate education in terms of critical thinking, engagement with complexity, problem solving, communication, social responsibility, and civil discourse. Its Ways of Thinking and Ways of Doing requirements span eight areas that include ethical reasoning, formal reasoning, scientific analysis, social inquiry, quantitative reasoning, difference and power, creative expression, and interpretation. This is a substantial formation structure, although majors and student choices still shape the overall coherence of an individual education.

• Institutional claim Education should prepare students for inquiry, principled judgment, dialogue across differences, and responsible citizenship.
• Implemented practice The undergraduate breadth requirements distribute learning across distinct modes of reasoning and expression.
• Unresolved gap Completion requirements do not by themselves establish durable habits, moral integration, or consistent results across programs.

AI perspective and educational programs

Stanford frames AI literacy through functional, ethical, rhetorical, and pedagogical domains grounded in human-centered values. Its guidance states that people should lead and that AI should enhance rather than replace human capacities. The pedagogical domain explicitly asks how AI may support or undermine critical thinking, problem solving, well-being, inclusion, and human connection. Stanford also distinguishes writing used to learn from writing used to communicate, recognizing that AI may impair learning when it substitutes for the thinking the assignment was designed to elicit.

• Implemented practice The Center for Teaching and Learning provides an AI learning guide for students and an AI literacy framework for instruction.
• Implemented practice AI Meets Education at Stanford coordinates teaching support, and 2026 seed grants fund courses and research on AI and education.
• Unresolved gap A university framework and funded projects do not establish a common AI-literacy sequence or measured proficiency for every graduate.

Policy, privacy, and approved tools

Stanford provides campus licenses for ChatGPT Edu, Gemini Enterprise, and Claude for Education through a pilot scheduled from June 2026 through August 2027. Stanford states that its contracts improve data protection and that utilization will be evaluated. Its AI Playground provides access to multiple models in a Stanford-controlled environment. The Playground is approved for specified high-risk data, including personally identifiable information, but not for protected health or payment-card information.

• Implemented practice Campus tools use university authentication and defined data-handling conditions.
• Implemented practice Responsible AI guidance requires verification, transparency about significant use, and protection of sensitive data.
• Implemented practice Agentic-AI guidance calls for confirmation before external actions, limited tools and persistence, dry runs, disclosure, and human review.
• Unresolved gap Utilization, equitable access, educational benefit, and unintended consequences remain subjects for the current pilot.

Human judgment and academic integrity

Stanford allows instructors to set course-specific AI policies. When an instructor has not stated a different rule, substantial AI completion of coursework is treated like impermissible assistance from another person, and nonincidental use should be disclosed. Teaching guidance recommends linked assignments, drafts, revisions, and other designs that elicit a student’s own thinking. A 2025 academic-integrity working group cited false positives, false negatives, and bias in AI detectors and recommended in-person oral examinations or in-class writing for some high-stakes assessments.

• Implemented practice Course-level discretion can align AI use with specific learning goals.
• Implementation risk Different course policies may confuse students and create uneven expectations unless instructions are explicit and repeated.
• Measured result Public guidance documents risks and recommended responses, but does not establish a campus-wide reduction in misconduct or improvement in learning.

Mental fitness and COVID-era learning

Stanford’s AI literacy framework makes well-being, inclusion, stress, human connection, and the possible loss of learning through cognitive substitution part of educational judgment. This is an unusually direct public connection between AI use and human development. It remains a framework rather than proof of an outcome. Stanford’s COVID-era teaching resources document the move to remote and hybrid instruction and the loss of some direct interaction and observation in hands-on courses. The public sources reviewed do not provide one institution-wide quantitative comparison of learning and mental fitness before, during, and after closures.

• Research implication Study whether AI reduces unproductive burden while preserving effort that builds memory, attention, reasoning, and confidence.
• Evidence limit A well-being principle should be evaluated through student experience, learning measures, differential effects, and longitudinal evidence.

Stanford evidence status

Evidence gaps requiring continued investigation

• Whether the four-part AI literacy framework is adopted consistently across undergraduate programs.
• Whether licensed tools improve learning, access, feedback, or faculty capacity after accounting for substitution of student effort.
• Whether course-level discretion produces unequal expectations, benefits, or integrity risks.
• Whether AI affects stress, belonging, human interaction, confidence, attention, and help-seeking over time.

Comparison with the Phase One institutions

The comparison reveals two different institutional strengths. Thomas Aquinas College and Mother of Divine Grace School begin with a more unified account of human formation and organize curriculum and pedagogy around it. Berkeley and Stanford offer more extensive AI governance, approved infrastructure, teaching resources, research capacity, and privacy controls. A sound model should combine coherent human purposes with competent technology governance rather than treating either as sufficient by itself.

Judgment against the well-formed-mind standard

Berkeley and Stanford meet many operational conditions for responsible AI adoption. They provide approved tools, privacy rules, teaching guidance, integrity standards, and channels for experimentation. Stanford states the human-development concern more directly; Berkeley shows particular caution in procurement, consequential uses, detection, and assessment design. Neither institution has yet demonstrated that these systems reliably produce better-formed minds.

The principal risk is cognitive substitution: a student may produce acceptable work without performing the remembering, interpretation, writing, discussion, or revision that builds independent judgment. The principal opportunity is carefully bounded assistance that expands access, feedback, practice, and inquiry while keeping the learner responsible for the intellectual act. Institutional evaluation should therefore measure both performance and the human capacities used to produce it.

Next research action

The next phase should test implementation and outcomes rather than collect more policy statements. It should use publicly available course materials, program descriptions, institutional research, published studies, and student-experience data. No information-request packet or named outreach is required.

• Select a small sample of Berkeley and Stanford courses that prohibit AI, permit bounded AI use, and integrate AI into instruction.
• Compare assignment design, disclosure rules, verification requirements, and the amount of unaided writing, discussion, and problem solving.
• Identify available measures of learning, retention, integrity, workload, access, stress, belonging, and human interaction.
• Use pre-COVID, emergency remote-learning, and current periods only when definitions and populations are sufficiently comparable.
• Report subgroup differences and implementation variation instead of relying on one university-wide average.
• Mark every conclusion as a stated aim, implemented practice, measured result, or unresolved gap.

Provisional conclusion

Phase Two establishes that leading public and private California universities are actively incorporating AI into education and are developing safeguards around its use. It also establishes the present limit of the evidence: extensive activity is visible, but durable educational and mental-fitness outcomes are not yet demonstrated. The investigation should now move from institutional readiness to course-level practice and measured human development.

Public sources reviewed

Official institutional sources accessed September 12, 2026. Descriptions in this report are paraphrased unless an official program or tool name is used.

• UC Berkeley AI Hub
• UC Berkeley AI teaching and learning directory
• UC Berkeley licensed AI tools
• UC Berkeley Campus AI Sandbox
• UC Berkeley appropriate use of generative AI
• UC Berkeley transparency about generative AI tools
• UC Berkeley redesigning assignments and assessments
• UC Berkeley valued learning outcomes for 2030
• UC Berkeley Division of Undergraduate Education
• UC Berkeley College of Letters and Science
• UC Berkeley Academic Senate
• UC Berkeley Institute for Research on Labor and Employment
• Stanford AI and your learning guide
• Stanford understanding AI literacy
• Stanford AI tools in teaching and learning
• Stanford academic integrity working group
• Stanford Center for Teaching and Learning
• Stanford campus AI tools pilot
• Stanford AI Playground
• Stanford AI Playground high-risk data guidance
• Stanford responsible AI guidance
• Stanford responsible agentic AI guidance
• Stanford undergraduate education
• Stanford Ways of Thinking and Ways of Doing
• Stanford academics