California Campus Level Artificial Intelligence and Human Formation Audit

CALIFORNIA CAMPUS-LEVEL EDUCATION AUDIT

Human formation and artificial intelligence across selected schools and universities

Investigation brief | September 2026 | Case set: UC Berkeley, UCLA, Stanford, USC, Caltech, Loyola Marymount University, Thomas Aquinas College, Mother of Divine Grace School, Harvard-Westlake, and High Tech High

What this audit does

• Compares public and private institutions at secondary and university levels through evidence of AI policy, curriculum, assessment, verification, ethics, privacy, and independent performance.
• Adds the post-COVID context: attention, belonging, family strain, interrupted learning, and mental fitness are educational conditions—not side issues.
• Separates institutional statements from observable practices, outcomes, and unresolved evidence gaps.

Interpretation rule

This is an evidence map, not a prestige ranking. A published policy is not proof of implementation; a promising course is not proof of schoolwide reach; and an innovative tool is not proof of better learning. Findings are provisional until institutions supply comparable practice and outcome data.

Case selection

1. Evaluation framework

A well-formed mind is more than employability, information recall, or tool fluency. It combines intellectual power with moral direction, self-command, social responsibility, and the capacity to act without surrendering judgment to a machine or crowd.

Evidence scale used in this report

Post-COVID lens

COVID-19 belongs in the investigation explicitly. School closures, emergency remote instruction, uneven parent capacity, social isolation, and disrupted routines altered the baseline from which current students learn. The campus question is therefore not merely “How is AI used?” It is “Does AI help rebuild—or further weaken—attention, relationships, confidence, foundational knowledge, and independent agency?” This audit records the issue as a cross-case requirement; it does not attribute a campus-specific mental-health effect without direct evidence.

2. Comparison at a glance

Preliminary pattern—not a ranking

• LMU presents the clearest published attempt to connect AI literacy with mission, ethics, privacy, access, and the humanities.
• Berkeley and Caltech provide especially useful bridges to verification or live independent performance; USC’s Keck policy offers a clear protected boundary for tasks meant to test a student’s own knowledge.
• Stanford combines student guidance with a significant experimentation ecosystem; its own education research also cautions that access to an AI tutor is not the same as effective human-supported learning.
• UCLA contributes something many institutions lack: evidence about student use and views, alongside citation and assignment-design guidance.
• Thomas Aquinas College and Mother of Divine Grace make the purpose of education and human formation unusually explicit, but their public evidence does not yet show a complete student-facing AI framework.
• Harvard-Westlake makes ethical reasoning and dialogue visible in named offerings; High Tech High makes authentic, public, project-based work visible. Neither public record yet closes the authorship-and-outcomes question.

3. Public universities

UC Berkeley

Published evidence: Teaching guidance asks instructors to state what generative AI use is allowed and why. Academic-integrity materials warn that substantive work produced by another entity can violate expectations. Berkeley engineering commentary also emphasizes critical evaluation, verification, and disclosure.

Formation value: The “allowed and why” formulation can connect rules to learning rather than punishment. Verification and disclosure develop epistemic responsibility: students must test claims and own their method.

Risk: Course-level discretion can produce inconsistent expectations. Students may learn to navigate rules without developing a stable account of which capacities must remain their own.

Evidence status: Policy/guidance: published. Student-use research: measured at broad disciplinary level. Common formation baseline and independent-performance outcomes: unresolved.

UCLA

Published evidence: UCLA provides generative-AI teaching guidance, citation expectations, and integrity language that rejects presenting AI output as one’s own. A Class of 2025 senior survey records student use and views.

Formation value: Acknowledgment makes assistance visible, while assignment redesign recognizes that the educational object—not simply the policing method—must change. Student survey evidence can reveal where actual use differs from faculty assumptions.

Risk: Citation alone does not show comprehension, judgment, or authorship. Surveying use is informative but does not establish gains or losses in learning, attention, or integrity.

Evidence status: Guidance and survey: published/measured. Course implementation and learning effects: unresolved.

Public-university judgment

Both systems face a scale problem: decentralized policies can be pedagogically appropriate, but students still need a recognizable institution-wide floor. That floor should define protected capabilities, disclosure, source verification, privacy, and the right to a meaningful AI-free path when a tool is not essential.

4. Private research universities

Stanford University

Published evidence: Stanford’s student guide compares generative-AI assistance to help from another person and directs students to the policy governing each learning context. Teaching resources address integrity and AI detection; the university funds educational AI experiments through AIMES and related initiatives.

Formation value: The outside-help analogy makes authorship intelligible. Research on AI-supported learning can move the discussion from access and novelty toward tested design. Stanford SCALE research reporting that human support improves engagement with AI tutoring reinforces a core formation principle: tools do not replace relationships.

Risk: A rich pilot ecosystem can fragment into isolated experiments. The institution still must distinguish demonstrations of technical efficacy from durable learning, judgment, belonging, and agency.

Evidence status: Student guidance and pilots: published/practiced. Research outcomes: emerging. Campuswide formation outcomes: unresolved.

University of Southern California

Published evidence: USC’s integrity office places the extent of permitted AI use with instructors, while an institutional resource hub supports teaching and ChatGPT Edu. Keck School of Medicine prohibits generative AI on assignments designed to evaluate a student’s own knowledge.

Formation value: Keck’s wording identifies the reason for an AI-free boundary: the assignment is evidence of the learner’s own competence. Institutionally supported access may also reduce shadow use and enable guided experimentation.

Risk: Broad tool availability can normalize use faster than shared pedagogy develops. School-by-school variation may leave students without a coherent progression from protected foundations to accountable augmentation.

Evidence status: Resources and policies: published/practiced. Cross-school consistency and learning outcomes: unresolved.

Research-university judgment

Stanford is strongest here in visible experimentation; USC is strongest in broad institutional enablement combined with at least one sharply stated knowledge boundary. Neither public record yet establishes a university-wide developmental sequence showing when students must work independently, when AI adds value, and how mastery is verified afterward.

5. Caltech and Loyola Marymount University

California Institute of Technology

Published evidence: Caltech’s writing center states that there is no campuswide student generative-AI policy and frames course decisions around whether use promotes or impedes learning. Teaching guidance identifies misinformation risks and suggests oral examinations, in-person examinations, alternative assessments, and active classrooms. Institutional guidance calls for disclosure.

Formation value: Oral defense and live problem-solving make reasoning observable. In a demanding technical environment, these approaches can preserve competence that must exist before augmentation. The explicit information/misinformation pairing supports verification.

Risk: Decentralization leaves the common baseline unclear. Technical rigor alone does not demonstrate breadth in ethical, civic, aesthetic, relational, or spiritual formation.

Evidence status: Guidance and assessment options: published. Actual uptake and broader formation outcomes: unresolved.

Loyola Marymount University

Published evidence: LMU’s resources explicitly address intentional course policy, student learning, privacy, accessibility, cost, and alternatives to required AI use. A provost task force links AI literacy to institutional mission; humanities and arts initiatives bring philosophy, bioethics, theology, and creative practice into the conversation. A 2026 certificate joins technical work with fairness, accountability, transparency, and ethics. Ai-DEAL extends education to teachers, administrators, and parents.

Formation value: LMU offers the most explicit synthesis in this set: tool literacy is situated inside a moral and humanistic frame, and access/privacy concerns are treated as educational design issues. Parent-facing work supports a wider learning ecology.

Risk: Much of the evidence describes initiatives in development or selected courses. The decisive questions are reach, requirements, instructional quality, independent assessment, and measured student growth.

Evidence status: Frameworks and programs: published/practiced. Scale, continuity, and outcomes: emerging or unresolved.

Distinctive contrast

Caltech shows how assessment architecture can protect live reasoning; LMU shows how mission and the humanities can discipline the ends for which AI is used. A stronger model would combine both: moral direction plus independently demonstrated competence.

6. Formation-centered institutions

Mother of Divine Grace School

Published evidence: Mother of Divine Grace School offers a Catholic classical distance-learning model in which parents play an integral educational role, supported by individualized plans, scripted lessons, consultants, online classes, tutoring, and grading. Its curriculum follows developmental stages and explicitly aims at habits of thought, good judgment, truth, goodness, and beauty.

Formation value: This case directly joins intellectual formation, moral purpose, developmental readiness, and family life. Its account of grammar, dialectic, and rhetoric provides a strong sequence for deciding what a student should first internalize, then analyze, and finally express and defend.

AI perspective: No sufficiently specific public schoolwide generative-AI policy or AI curriculum was located in this review. That absence should be recorded as unresolved—not interpreted as prohibition, approval, or lack of private guidance.

Risk: Because much work occurs at home, the model depends on parent capacity and consistent expectations. AI can quietly replace practice or authorship unless families receive age-specific rules, examples, and verification routines. Post-COVID differences in family time, confidence, connectivity, and stress make this especially important.

Evidence status: Educational philosophy, curriculum sequence, and family support: published/practiced. AI rules, tool guidance, implementation, and outcomes: unresolved.

Thomas Aquinas College

Published evidence: Thomas Aquinas College uses one Great Books program across mathematics, science, language, literature, philosophy, and theology, taught through small, vigorous discussions. In 2026 it hosted an AI and Human Flourishing symposium that asked whether AI can attain wisdom, how it affects thought and moral agency, and how technology can remain subordinate to the human person.

Formation value: The discussion method, original texts, first-principles questions, and integrated liberal curriculum directly cultivate reasoning, listening, argument, and intellectual humility. Its AI inquiry begins with the nature of intelligence, wisdom, work, relationship, and the common good rather than with tool adoption.

AI perspective: The symposium is meaningful evidence of institutional intellectual engagement, but it is not yet evidence of a student-facing AI curriculum, assignment policy, or assessment sequence. The public record reviewed here does not establish how students may use AI in required academic work.

Risk: A strong philosophy of the human person can remain external to everyday tool decisions unless translated into concrete student practice: disclosure, protected independent work, source verification, privacy, and disciplined non-use.

Evidence status: Educational model and AI symposium: published/practiced. Student-use policy, curricular reach, and outcomes: unresolved.

Formation-centered judgment

These two cases contribute something the wider set often leaves implicit: an articulated end of education and a developmental account of intellectual formation. Their next challenge is translation. A credible AI framework should specify how classical habits—memory, analysis, dialogue, rhetoric, moral judgment, and family responsibility—govern actual tool use and independent assessment.

7. Secondary schools

Harvard-Westlake School

Published evidence: The 2026–27 curriculum includes Ethics, Artificial Intelligence, and Society. Other offerings develop dialogue across difference, active listening, empathy, inquiry, respectful disagreement, primary research, and fact-checking. The Kutler Center describes original research, public presentation, and ethical reasoning.

Formation value: This is unusually explicit secondary-level evidence that AI belongs inside ethics and society, not only computer science. Dialogue and fact-checking connect intellectual formation with character, relationship, and public responsibility.

Risk: The public evidence is course- and program-centered. It does not yet show whether every student encounters these capacities, which AI uses are permitted across assignments, or how unaided authorship and growth are verified.

Evidence status: Curriculum and program descriptions: published/practiced. Schoolwide reach, AI policy, and outcomes: unresolved.

High Tech High

Published evidence: High Tech High publicly emphasizes student-centered project-based learning, internships, publications, and student work. Its 2026–27 student/parent handbook states that technology, including generative AI, must be used according to the High Tech High AI Use Policy, but the accessible public evidence reviewed here does not disclose enough policy detail for a full audit.

Formation value: Authentic projects and public products can make learning purposeful, collaborative, and accountable to audiences beyond a test. Internships connect knowledge to adult responsibility and real constraints.

Risk: In collaborative and AI-assisted projects, an impressive product can conceal uneven individual understanding. The model needs explicit checkpoints—process records, oral defense, individual writing, source trails, and unaided demonstrations—to verify formation.

Evidence status: Learning model and handbook reference: published. AI policy detail, implementation, individual-mastery measures, and longitudinal outcomes: unresolved.

Secondary-school judgment

Harvard-Westlake supplies visible ethical and dialogical content; High Tech High supplies visible authentic work. The most promising synthesis would unite ethical inquiry, respectful disagreement, project responsibility, and reliable individual demonstrations. Families should receive clear guidance because secondary-school AI habits are formed partly outside school and because post-COVID learning conditions vary widely by household.

8. Cross-institution conclusions

Emerging minimum standard

• State the human purpose of each course and major assessment.
• Identify foundational knowledge and skills students must retain without AI.
• Use protected demonstrations: live reasoning, oral defense, supervised writing, performance, or practical application.
• Require traceable assistance: tool, prompt or process description, sources, verification, and substantive revisions.
• Teach refusal as a form of competence: students should know when AI degrades privacy, attention, learning, or responsibility.
• Pair AI experiments with human support and equitable alternatives.
• Measure learning, agency, integrity, belonging, and mental fitness—not usage alone.
• Provide parents and students a shared, plain-language developmental framework.

A developmental sequence

9. Mental fitness, family dynamics, and COVID

A historical comparison of mental fitness can be useful, but only if it avoids nostalgia and false equivalence. The defensible unit of comparison is not “people today versus people then” in the abstract. It is a set of measurable educational conditions across cohorts: sustained attention, loneliness and belonging, sleep, anxiety, help-seeking, family time and conflict, attendance, reading stamina, physical activity, and independently demonstrated learning.

Why COVID must remain explicit

• School closures and emergency remote education changed instruction abruptly and unevenly.
• Parents differed greatly in time, confidence, language access, devices, connectivity, and ability to supervise learning.
• Some students received structured online teaching; others received inconsistent instruction or disengaged.
• Isolation and disrupted routines affected social development, motivation, and belonging alongside academic progress.
• AI now enters classrooms before those effects are fully resolved, creating both remediation opportunities and risks of further dependency or avoidance.

Recommended comparison design

Campus evidence request

For each institution, request: AI policies by school and course; sample assignments; assessment redesigns; misconduct trends; student and faculty surveys; tutoring and advising models; privacy/access rules; post-COVID learning and wellness measures; parent communication for secondary schools; and disaggregated outcomes. Interviews should include students, faculty, counselors, learning-support staff, parents, and recent graduates—not administrators alone.

10. Investigation plan and decision gates

Questions that could change the current judgment

• Which capabilities does each institution require students to demonstrate without AI, and at what intervals?
• Can students accurately explain how AI changed their reasoning—not merely disclose that it was used?
• Which courses systematically teach verification, source provenance, model limitations, and privacy?
• Do AI-integrated assignments outperform well-designed non-AI alternatives on delayed learning?
• Are attention, belonging, confidence, and help-seeking improving or deteriorating for different student groups?
• How do institutions respond when families or students decline a tool on privacy, cost, accessibility, or educational grounds?

Next task

Build the AI Tool Suitability Matrix: evaluate categories of AI tools—tutors, writing assistants, search/research systems, coding copilots, creative generators, adaptive platforms, and monitoring systems—against the developmental sequence and minimum standard above. The matrix should recommend appropriate, conditional, and inappropriate uses by age and learning purpose.

11. Selected official sources

Sources below are official institutional materials or institution-hosted research reviewed for this audit. Dates and policies should be reconfirmed during interviews and before publication.

UC Berkeley and UCLA

Berkeley Center for Teaching & Learning: course policies for generative AI — course-specific permission and pedagogical rationale.

Berkeley Academic Senate: generative-AI guidance — academic integrity and authorship guidance.

UCLA HumTech: citing generative AI — acknowledgment guidance.

UCLA Digital & Technology Solutions: generative AI guide — integrity and responsible-use resources.

Stanford, USC, and Caltech

Stanford Teaching Commons: AI and your learning — student-facing expectations and outside-help analogy.

Stanford SCALE: research publications — research on AI tutoring and human support.

USC Office of Academic Integrity — institutional integrity guidance.

USC AI resources — institutional teaching, learning, and tool resources.

Caltech Hixon Writing Center: generative AI — course-specific learning framework.

Caltech CTLO: teaching in the age of AI — assessment and classroom-design options.

LMU and secondary schools

LMU: AI in teaching and learning — policy design, privacy, access, and alternatives.

LMU Provost: generative-AI task force — mission-aligned AI literacy and governance.

LMU Bellarmine College: AI and the humanities — philosophy, bioethics, theology, and humanitarian framing.

LMU School of Education: Ai-DEAL — AI education for educators, administrators, and parents.

Mother of Divine Grace School — Catholic classical distance learning, individualized support, and parents’ educational role.

Mother of Divine Grace: Teaching the Art of Learning — developmental stages and intellectual formation.

Thomas Aquinas College — Great Books curriculum and discussion method.

Thomas Aquinas College: AI and Human Flourishing — first-principles symposium on AI, wisdom, work, and the human person.

Harvard-Westlake curriculum guide — curriculum and interdisciplinary offerings.

Harvard-Westlake School — mission, programs, and school context.

High Tech High — project-based learning, internships, student work, and governance resources.

Limits of this case file

The audit relies on public institutional evidence available during the review. It does not claim that every relevant practice is published, that a named offering reaches every student, or that policy language predicts classroom behavior. COVID, mental fitness, family dynamics, and tool suitability require additional direct evidence and will be treated as explicit workstreams—not speculative conclusions.