Supplemental institutional audits · Complete web edition

California Institutional Artificial Intelligence Readiness and Human Formation Audit

This is the complete web edition of the canonical project document. Its narrative, recommendations, evidence tables and source links are presented below without abridgment.

4,169words

8tables

144text blocks

Download the Word document

Project, Research, Analysis, Editorial by Patrick Russell; assisted by ChatGPT (OpenAI).

Published as a web edition September 13, 2026

California Institutional AI Readiness and Human Formation Audit

This report applies the project’s AI Tool Suitability Matrix to ten California institutions. It examines whether each institution’s public evidence protects the work through which students acquire knowledge, judgment, attention, moral agency, and the ability to perform without machine assistance.

Executive finding

No institution in this sample publicly demonstrates the full chain required for responsible educational AI: an explicit human purpose, approved tools and protected accounts, assignment-level rules, age-appropriate or student-appropriate limits, independent assessment, family or learner notice, trained educators, and measured outcomes. LMU and Stanford publish the broadest frameworks. Berkeley, UCLA, USC, and Caltech publish useful but decentralized guidance. Thomas Aquinas College and Mother of Divine Grace provide unusually clear accounts of human formation, yet public AI governance remains incomplete. Harvard Westlake and High Tech High show promising curricular or authentic-work practices, but their public evidence does not yet allow a full implementation audit.

What this finding means

This is an evidence-readiness audit, not a ranking of educational quality or institutional character.
A published policy counts as a claim until implementation and outcomes are demonstrated.
A missing public document is recorded as unresolved. It is not proof that the practice is absent.
The standard gives equal interpretive weight to human formation, privacy and control, assessment, family or learner involvement, implementation, and outcomes.

Institutions reviewed

Sector	Institutions
Public universities	UC Berkeley and UCLA
Private universities	Stanford USC Caltech and Loyola Marymount University
Catholic formation models	Thomas Aquinas College and Mother of Divine Grace School
Secondary education	Harvard Westlake and High Tech High

1 Audit standard

The standard asks whether AI serves education or quietly replaces it. The preferred sequence is human teaching and independent effort first, source-bounded research second, guided tutoring third, and unrestricted generation only after the student has attempted the work. The standard disfavors answer generation, simulated companionship, automated emotional inference, and surveillance as ordinary educational substitutes.

Dimension	Evidence required for a strong finding
Human formation	The institution names the kind of intellectual and moral agency education should develop and carries that purpose into curriculum and assessment.
Independent capability	Students read write calculate reason explain or perform without AI at defined stages.
Truth and sources	AI claims must be traced to inspectable sources tested for error and disclosed when they affect the work.
Privacy and control	Approved tools use managed accounts and clear limits on training retention sensitive data advertising and third parties.
Classroom rules	Rules vary by task but remain clear before work begins and explain why use is allowed restricted or prohibited.
Family or learner voice	Families receive notice and meaningful choice for minors. Adult learners receive transparent notice consent and an equivalent alternative when AI is not essential.
Teacher preparation	Educators can design AI-aware assignments identify tool limitations protect data and judge learning without relying on detectors.
Mental fitness	Use protects attention belonging resilience human help and freedom from dependency or simulated relationship.
Independent assessment	The institution verifies mastery through live oral supervised practical handwritten or otherwise controlled performance.
Measured outcomes	Evaluation includes delayed mastery transfer integrity agency well-being access and subgroup effects rather than adoption alone.

Evidence codes

Code	Meaning
M	Meets the standard in the public evidence reviewed
P	Partly meets the standard or depends on local implementation
U	Unresolved because comparable public evidence was not located
C	Concern because public evidence identifies a material limitation or risk
NA	Not directly applicable such as parent notice for adult university students

Scores describe public evidence current to September 2026. They should change when an institution supplies stronger or more current evidence.

2 Learning and formation scorecard

Institution	Human purpose	Independent work	Truth and sources	Mental fitness	Family or learner voice
UC Berkeley	P	P	M	U	P
UCLA	P	P	P	U	P
Stanford	P	P	M	P	M
USC	P	P	P	U	P
Caltech	P	M	M	U	P
LMU	M	P	M	P	M
Thomas Aquinas College	M	M	M	P	P
Mother of Divine Grace	M	M	P	P	M
Harvard Westlake	P	P	P	P	U
High Tech High	M	M	P	P	P

Interpretation

The formation leaders are not necessarily the AI governance leaders. Thomas Aquinas College and Mother of Divine Grace state the human end of education with exceptional clarity. High Tech High builds authenticity into projects and internships. Caltech can protect independent performance through demanding live work. LMU comes closest to integrating mission, ethics, privacy, access, and optionality in one published framework. Stanford’s guidance pays unusual attention to student well-being, privacy concerns, and equivalent alternatives.

Why mental fitness remains unresolved

The institutions commonly offer wellness services or mention student well-being, but the public AI guidance rarely measures whether AI use changes attention, persistence, help-seeking, loneliness, confidence, or human relationships. Stanford research reporting lower well-being among vulnerable users who seek emotional support from AI companions makes this omission material, while not proving a school-specific effect. [11]

Why COVID belongs in the audit

Students entered the AI transition after school closures, emergency remote instruction, uneven family capacity, social isolation, and disrupted learning routines. Institutions should therefore test whether AI rebuilds or further weakens attention, belonging, foundational knowledge, and independent agency. This report does not attribute a COVID effect to any institution without institution-specific data; it uses the pandemic as a required baseline question for interviews and outcome analysis.

3 Governance and implementation scorecard

Institution	Approved tools and privacy	Classroom rules	Teacher preparation	Independent assessment	Measured outcomes
UC Berkeley	P	P	P	P	P
UCLA	U	P	P	P	P
Stanford	P	M	M	P	P
USC	M	P	P	P	U
Caltech	C	M	P	M	U
LMU	M	M	M	P	P
Thomas Aquinas College	U	U	U	M	U
Mother of Divine Grace	U	U	P	P	U
Harvard Westlake	U	U	U	P	U
High Tech High	U	P	U	P	U

Evidence readiness groups

Group	Institutions	Meaning
Broadest public framework	LMU and Stanford	Multiple parts of the chain are visible but outcome evidence remains incomplete.
Structured but decentralized	Berkeley UCLA USC and Caltech	Strong components exist but course school or division discretion can create uneven student experience.
Formation rich and governance incomplete	Thomas Aquinas College and Mother of Divine Grace	The educational end is clear while AI tools privacy and use rules are not publicly demonstrated.
Selective or partial K 12 evidence	Harvard Westlake and High Tech High	Curriculum or authentic work is visible but the full student-facing governance chain is not.

The key distinction

A school may be excellent at human formation and still need AI governance. It may also have sophisticated AI services without protecting the capacities students must own. Readiness requires both.

4 UC Berkeley and UCLA

UC Berkeley

Verified public evidence Teaching guidance asks instructors to state how generative AI may be used and why. Campus risk guidance addresses appropriate use and data protection. Berkeley-led research using 95,513 students across 20 public research universities found substantial disciplinary variation and called for discipline-specific assessment reform rather than blanket bans or universal detection. [1, 2, 3]

What meets the standard The emphasis on pedagogical rationale, verification, disclosure, and discipline-specific assessment supports truth-seeking and avoids treating AI as one uniform problem. [1, 3]

What only partly meets Course discretion is educationally sensible, but the public record does not establish a common campus floor for protected unaided capabilities, school-managed student tools, equivalent AI-free pathways, or delayed mastery measures.

Measured outcomes The Berkeley-led study measures use and estimated misuse across institutions, not Berkeley learning gains attributable to a campus intervention. It is strong contextual evidence and limited institutional outcome evidence. [3]

Investigation request Obtain the approved-tool register and contract terms, a sample of AI statements by discipline, assessment maps identifying unaided work, faculty training completion, integrity trends, and any pre and post measures of learning.

UCLA

Verified public evidence UCLA asks faculty to identify knowledge and skills students must develop on their own, connect AI rules to learning outcomes, and acknowledge permitted AI-generated content. Its Class of 2025 survey received 6,639 responses; 73 percent reported using generative AI for coursework, 65 percent judged it more helpful than harmful, and about three quarters wanted training for instructors and students. [4, 5, 6]

What meets the standard The policy guidance begins with the capability the student must own. The survey creates a factual baseline and records student concern about dependency, critical thinking, accuracy, and bias. [4, 6]

What only partly meets Acknowledgment makes assistance visible but does not prove authorship or mastery. Public evidence does not show a common approved-tool environment, systematic privacy choice, or the learning effects of actual student use.

Investigation request Request the 2026 tool-specific survey results, approved-account rules, assignment samples, instructor training participation, independent assessment practices, and outcome comparisons by discipline and usage pattern.

5 Stanford and USC

Stanford

Verified public evidence Stanford treats unapproved generative AI assistance like help from another person. Its course-policy guidance asks instructors to specify tools, contexts, disclosure, consent, sensitive-data limits, support, well-being, and optional alternatives. AIMES supplies examples and critical AI literacy resources. Stanford SCALE cautions that access to AI tutoring is not enough and that evidence remains limited across contexts. [7, 8, 9, 10]

What meets the standard The public framework is unusually complete at the course level. It recognizes privacy objections, permits AI-free alternatives, asks whether students have prerequisite skills, and recommends process-oriented assignments that expose thinking. [7, 8]

What only partly meets A rich innovation ecosystem does not establish uniform implementation. The public evidence reviewed does not show which students use managed tools, which courses protect independent performance, or which pilots improve delayed learning and agency.

Mental fitness relevance Stanford research links intense companionship-oriented chatbot use among people with smaller social networks to lower well-being. The finding supports a clear boundary against educational AI companions, while remaining correlational and not campus-specific. [11]

Investigation request Obtain tool contracts, AIMES participation and reach, representative syllabi, independent assessment requirements, pilot protocols, retention and well-being outcomes, and rules that prohibit AI from functioning as counselor or companion.

USC

Verified public evidence USC offers an enterprise ChatGPT Edu environment to active students faculty and staff. It describes the workspace as protected and USC-only, while also warning that users have no expectation of privacy on USC-owned technology. Academic integrity guidance discourages overreliance on AI detection and requires additional evidence and professional judgment. [12, 13, 14]

What meets the standard A managed enterprise environment is stronger than unmanaged consumer accounts. The detector guidance protects due process and requires human judgment. Faculty retain responsibility for course expectations. [12, 14]

What only partly meets The public record does not demonstrate consistent assignment-level rules, protected independent assessments across schools, an equivalent alternative for students who decline AI, or measured learning effects. Institutional access can normalize use before the educational purpose is established.

Investigation request Request the ChatGPT Edu data-flow and retention terms, adoption by school, student notice language, syllabus sampling, Keck and other professional-school assessment rules, faculty preparation data, and independent measures of learning and integrity.

6 Caltech and Loyola Marymount University

Caltech

Verified public evidence The Hixon Writing Center recommends a course-specific policy because ethical and dishonest uses remain context-dependent. The Humanities and Social Sciences division allows AI in submitted work only when an instructor explicitly permits it. Caltech information security states that it lacks contracts for most AI services, so ordinary security privacy and compliance protections may not be in place. Caltech also strongly discourages AI detectors for student writing. [15, 16, 17, 18]

What meets the standard Default prohibition in HSS until explicit permission protects authorship. Caltech’s live technical culture and assessment guidance can support direct demonstration, while its detector stance protects trust and due process. [16, 18]

Material concern The absence of contracts for most tools is a stated privacy and control limitation, not merely a missing document. It should constrain use involving private sensitive unpublished or proprietary information. [17]

What remains unresolved The public evidence does not establish a campuswide approved student tool, broad AI literacy curriculum, mental-fitness safeguards, or measured effects on technical understanding, writing, and judgment.

Investigation request Obtain the current approved-tool list and exception process, division-level policies, oral or proctored assessment use, privacy training completion, student disclosure practices, and evidence that students can explain and reproduce AI-assisted work.

Loyola Marymount University

Verified public evidence LMU tells faculty to consider accounts, cost, accessibility, privacy, and reasonable alternatives before requiring a tool. It publishes AI-free and AI-integrated options, supports protected Microsoft Copilot accounts, and advises against uploading protected or confidential information. Its task force seeks classroom-ready definitions and frameworks. A seven-faculty study found student openness to AI integration but a gap between current implementation and higher-level learning goals. [19, 20, 21, 22]

What meets the standard LMU connects institutional mission, human dignity, ethics, privacy, access, teacher choice, learner choice, and curricular experimentation more explicitly than the other institutions reviewed. [19, 21, 22]

What only partly meets The strongest elements are guidance and emerging programs. Public evidence does not yet establish schoolwide reach, recurring unaided assessment, or sustained gains in judgment, learning, well-being, and professional practice.

Investigation request Request the task force’s adopted framework, course adoption data, School of Education integration, Ai-DEAL participation, assessment samples, Copilot telemetry boundaries, student survey instruments, and pre and post outcome evidence.

7 Thomas Aquinas College and Mother of Divine Grace

Thomas Aquinas College

Verified public evidence TAC requires one integrated Great Books program built around original works and small vigorous discussion. Its 2026 AI and Human Flourishing symposium asked whether AI can be intelligent or wise, how it changes education and leisure, and how technology can remain subordinate to the human person. The symposium included competing viewpoints and Socratic seminars. [23, 24, 25]

What meets the standard The curriculum protects sustained reading, live reasoning, memory, conversation, and judgment. These practices make independent performance and human presence structural rather than optional. The AI inquiry is framed by first principles and human flourishing. [23, 24, 25]

What remains unresolved A conference and podcast series establish an institutional perspective, not a student-use policy or required AI curriculum. Public evidence reviewed does not identify approved tools, privacy rules, disclosure requirements, or how AI-assisted work is treated in tutorials, papers, laboratories, or examinations.

Investigation request Ask for the current student and tutor AI policy, examples of AI questions inside the required program, assessment and authorship rules, technology boundaries in residence and study, and evidence about attention, reading depth, intellectual confidence, and post-graduate judgment.

Mother of Divine Grace School

Verified public evidence MODG defines education as teaching children how to think and orders its curriculum through developmental stages of memory and observation, analysis, and communication. It treats parents as primary educators and provides training, consultants, scripted lessons, discussion, revision, and individualized support. [26, 27, 28]

What meets the standard The developmental sequence, family authority, and emphasis on habits of thought provide a strong basis for delayed and age-bounded AI use. The family can observe authorship and give help without replacing the child’s effort. [26, 27, 28]

What remains unresolved No sufficiently specific public AI curriculum or use policy was located for this review. Family control should not be assumed to solve vendor privacy, age restrictions, disclosure, source verification, or consistency across thousands of homes.

Investigation request Request age-banded AI guidance, parent training, approved and prohibited tools, privacy and account requirements, student disclosure rules, oral checks for online work, adaptations for learning differences, and measures of independent reading writing reasoning and attention.

8 Harvard Westlake and High Tech High

Harvard Westlake

Verified public evidence The 2026 to 2027 curriculum includes Ethics AI and Society for grades 10 through 12 and Dialogue Across Difference for grades 10 and 11. The latter practices listening empathy inquiry respectful disagreement and public presentation. Other offerings emphasize research interviewing fact-checking and feedback. [29]

What meets the standard The curriculum recognizes that AI education must include ethics and that a well-formed mind needs dialogue, evidence, and the ability to face disagreement. These are meaningful formation practices when students must demonstrate them directly. [29]

What only partly meets The named offerings appear selective rather than a required schoolwide sequence. Public evidence reviewed does not establish approved tools, managed accounts, privacy rules, parent notice, AI-free alternatives, teacher preparation, or common independent-assessment requirements.

Investigation request Obtain the student AI policy, participation and enrollment data, parent communications, tool contracts, discipline-specific rules, teacher training, student work samples, oral defenses, and outcomes for critical thinking, dialogue, authorship, and well-being.

High Tech High

Verified public evidence The 2026 to 2027 handbook requires students to use AI according to the High Tech High AI Use Policy and teacher expectations. Board minutes show that the policy followed input from students and staff and was intentionally broad. HTH requires juniors to complete a four-week full-time internship with a mentor, a substantive project, a portfolio, and a presentation of learning. Its public project archive makes authentic work visible. [30, 31, 32, 33]

What meets the standard Projects, workplace mentors, journals, portfolios, and public presentations create several opportunities to verify authorship, explanation, collaboration, and real-world consequence. [32, 33]

What only partly meets The handbook references the policy but the evidence reviewed does not expose enough operational detail to judge approved tools, privacy, family choice, AI-free alternatives, teacher preparation, or consistency across campuses. Collaborative work still requires individual authorship checks.

Investigation request Obtain the full AI Use Policy and implementation guides, parent notice, tool approvals, assignment-level protocols, project logs, individual oral defenses, mentor observations, and outcome comparisons before and after AI adoption.

9 Cross institution findings

Policy is ahead of outcome evidence

The institutions publish many explanations of acceptable use and very little evidence of durable learning. UCLA provides a useful student-use baseline, Berkeley-led research supplies a large cross-institution misuse study, Stanford and LMU report research or pilots, and High Tech High displays authentic products. None of these alone answers whether students know more, reason better, retain skills longer, or act with greater independence because of a specific institutional approach.

Course discretion needs a common floor

Berkeley, UCLA, Stanford, USC, and Caltech rely substantially on instructor or division decisions. Local judgment is necessary because learning goals differ. Students still need common rules for privacy, disclosure, prohibited high-stakes uses, due process, equivalent alternatives, and the independent abilities every graduate must demonstrate.

Approved access can create educational pressure

Managed accounts improve privacy and administrative control, but institutional availability can imply that use is expected. Schools should distinguish permission from recommendation and require faculty to name the educational purpose before assigning a tool.

Formation and governance are complementary

TAC and MODG provide the clearest accounts of wisdom, truth, intellectual habit, and the human teacher or parent. LMU provides the clearest public bridge from mission to AI practice. Stanford provides the broadest experimentation and course-policy resources. HTH provides strong authenticity structures. The best model would combine these strengths rather than copy one institution wholesale.

K 12 requires a higher standard of care

For minors, a course statement is not enough. Families need notice of the tool, purpose, data practices, expected frequency, human supervision, and a workable alternative. Schools should bar AI companionship, automated emotional inference, and behavioral surveillance except in a narrowly justified safety or accessibility case with independent validation and appeal.

COVID changes the baseline

Outcome studies should separate convenience from recovery. A useful intervention should improve at least one damaged capacity, such as foundational knowledge, attendance, persistence, belonging, or confidence, without worsening attention, dependency, isolation, or family strain. Pre-pandemic comparison can provide historical context, but the most defensible design also follows current students over time and compares similar groups under different instructional conditions.

10 Minimum evidence package

The investigation should request the same evidence package from every institution so that comparison does not reward institutions merely for publishing more polished webpages.

Evidence item	What to request	Why it matters
Purpose and scope	Adopted AI policy mission rationale covered populations and responsible office	Shows whether AI use begins with an educational end and named accountability.
Tools and contracts	Approved products account types age limits retention training advertising subprocessors and deletion	Separates protected institutional use from consumer use.
Classroom implementation	Representative syllabi assignments disclosures alternatives and disciplinary variations	Tests whether central guidance reaches ordinary student work.
Independent mastery	Oral live supervised handwritten practical or closed-tool assessment map	Shows what knowledge and judgment the institution expects students to own.
Educator preparation	Training curriculum participation role expectations and instructional support	Tests implementation capacity rather than policy aspiration.
Family and learner voice	Notice consent opt-out alternative complaint and appeal procedures	Protects meaningful choice and makes power accountable.
Mental fitness	Attention belonging help-seeking dependence loneliness workload and relationship measures	Detects whether efficiency gains carry human costs.
Learning outcomes	Baseline follow-up delayed transfer subgroup integrity and graduation measures	Tests whether AI changes education rather than only workflow.
Incident record	Privacy integrity bias safety accessibility and due-process incidents with resolutions	Reveals operational risk and institutional learning.
Annual review	Owner review date revision triggers public summary and sunset rules for pilots	Prevents temporary experiments from becoming permanent by inertia.

Interview sequence

Ask leaders what human capability the institution refuses to outsource.
Ask teachers and students what actually happens in ordinary assignments and assessments.
Ask families of minors what notice and choice they received.
Inspect the approved account and vendor contract rather than relying on a product description.
Compare submitted work with live explanation and delayed independent performance.
Review results by discipline age prior achievement disability language background and frequency of AI use.

11 Recommended institutional baseline

Every institution in the sample could adopt the following baseline without eliminating local academic judgment or legitimate experimentation.

Requirement	Minimum rule
Protected capabilities	Each program identifies reading writing quantitative reasoning disciplinary knowledge practical skill and judgment that students must demonstrate without AI.
Assignment rules	Every assessed task states whether AI is prohibited restricted or permitted and explains the learning reason.
First attempt	Students make an independent attempt before AI supplies explanation critique or revision assistance unless accessibility needs require another design.
Bounded tools	Source-bounded or guided modes precede unrestricted assistants for instructional use.
Managed accounts	Required use occurs only through an approved account with contractually reviewed privacy and security terms.
Human decisions	AI does not make final decisions about grades discipline admission counseling accommodations or student risk.
Independent verification	Programs use recurring oral live practical supervised or closed-tool assessment to verify authorship and mastery.
Meaningful alternative	A student can use an equivalent non-AI pathway when AI is not essential and the objection concerns privacy access disability ethics or conscience.
Family protection	For minors families receive clear notice purpose data practice frequency supervision and appeal rights before required use.
Mental fitness boundary	AI is not presented as friend therapist spiritual guide or teacher replacement. Use is time-bounded and followed by human or offline activity.
Evaluation	Pilots define outcomes and comparison conditions before launch and publish learning integrity access and well-being results.
Review	An accountable office reviews tools incidents evidence and policy at least annually and retires uses that do not justify their costs.

Decision rule

Do not scale a tool merely because students use it, teachers request it, or a vendor offers protected access. Scale only when the institution can name the capability being developed, show that the tool does not replace that capability, protect the learner’s data and freedom, verify independent mastery, and measure outcomes that matter.

12 Sources

Numbered citations in the report refer to the primary institutional pages below. Sources were reviewed as public evidence available in September 2026. Product terms and institutional policies can change, so contract and policy verification remains part of the investigation.

1. UC Berkeley Center for Teaching and Learning course policy guidance

2. UC Berkeley Office of Ethics Risk and Compliance guidance

3. UC Berkeley study of generative AI use and assessment reform

4. UCLA course policy guidance

5. UCLA guidance on citing generative AI

6. UCLA Class of 2025 generative AI survey

7. Stanford Teaching Commons course policy guidance

8. Stanford Teaching Commons AI teaching guidance

9. Stanford Teaching Commons AI resources and AIMES

10. Stanford SCALE research on AI tutoring

11. Stanford HAI study of AI companions and loneliness

12. USC ChatGPT Edu service