“The problem with AI in California public schools is not that students cannot access it. The problem is that most people do not understand what AI is, what it can do, or what it cannot do.”
James L. DeBacco, MSW, DSW(c) | Doctoral Researcher, USC Suzanne Dworak-Peck School of Social Work. Founder & CEO, DeBacco Nexus LLC | Member, CalCompute Consortium. Patent Pending — USPTO 19/571,156 | April 2026
A note from the author: I manage hearing impairment, sight impairment, dyscalculia, and abstract reasoning dyslexia. I use AI daily as an assistive tool. I can attest from direct personal experience that governed AI — shorter, clearer, and constrained — reduces my cognitive load in ways that ungoverned AI does not. This paper is not written about a population studied from the outside. It is written from within it.
The Problem Is Not Access. It Is Knowability.
California has invested significantly in closing the digital divide — ensuring students in under-resourced schools have devices, connectivity, and technology. That work matters. But it does not address the problem that is actually in front of us.
The problem with AI in California public schools is not that students cannot access it. The problem is that most people — students, parents, educators, veterans, justice-involved individuals, and aging adults — do not understand what AI is, what it can do, or what it cannot do. They encounter it without context. And the AI they encounter is ungoverned: verbose, technically framed, and designed for users who already know how to engage with it.
This is the knowability gap. It exists before anyone touches a keystroke. And it widens every day that California deploys AI in schools without a governance standard.
What Ungoverned AI Costs Every User
The knowability gap is not only a problem for disadvantaged populations. It affects every person who uses AI regardless of education, income, or technical sophistication.
Ungoverned AI models generate comprehensive responses by default. A simple question receives a 400-word answer. A direct request receives hedged, qualified, verbose output. This is not a flaw in any individual model — it is the default behavior of a system with no constraint on output length.
Industry analysts describe what they call a “token consumption explosion”: what once returned 1,000 tokens now routinely returns 100,000 (Ikangai, 2025). For consumers on flat-rate subscriptions, this cost is invisible. As Deloitte has documented, when AI is consumed through packaged software, token consumption is abstracted entirely — users see a predictable subscription fee with no transparency into what is being consumed on their behalf (Deloitte Insights, 2026).
Most users have no understanding of what is being generated or at what cost. According to one industry analysis, most consumers pay a flat monthly subscription and default to the most powerful available model, unaware of the computational resources consumed on their behalf (WAV Group Consulting, 2025).
Whether this structural invisibility is an intentional design choice or an architectural consequence is a question California policymakers should ask directly of AI providers. What is documentable is the result: consumers at every level of sophistication consume more tokens than their tasks require, at costs they cannot see, producing cognitive and environmental waste that governance could reduce.
A Voice from the Classroom and the Profession
The knowability gap is not an abstract policy concern. It is a daily reality for educators being asked to integrate AI into their institutions.
“I don’t trust AI, but I still try and use it.” Dr. Kim DeBacco, retired Lead Instructional Designer at UCLA, made this statement after direct engagement with AI tools in an educational context. She consults regularly with colleagues across institutions who share her concern — faculty being asked to accept AI as a legitimate learning support tool while lacking any framework to evaluate what “governed” versus “ungoverned” means in practice.
Dr. DeBacco’s statement is not a rejection of AI. It is a precise description of how thoughtful, experienced educators engage with a tool they cannot fully trust because its behavior is unpredictable, its output is often excessive, and its accountability is unclear. That is a rational response to ungoverned AI. The problem is not the educators. The problem is the tool.
Who Bears the Greatest Cost
While ungoverned AI affects all users, the populations who bear the greatest cost are those least equipped to navigate its failures.
Students with Disabilities and Learning Differences
For students managing hearing impairment, sight impairment, dyscalculia, abstract reasoning dyslexia, traumatic brain injury, PTSD, anxiety, and other neurodivergent profiles, ungoverned AI does not reduce cognitive burden — it increases it. Verbose, hedged, comprehensive output assumes processing capacity that these students may not have in a given moment.
Governed AI — constrained, brief, and precise — creates the conditions in which AI becomes genuinely assistive rather than overwhelming. The architecture that enforces a hard output ceiling for an emergency response also makes a tutoring response processable for a student who cannot navigate 400 words to find the one sentence they needed.
Students in Under-Resourced Schools
A 2024 study cited by the California Department of Education found that young Black and Latino students use generative AI at higher rates than young white students (Flapan, 2026). The students most exposed to ungoverned AI are often the least equipped to navigate its failures. A student from a family without prior exposure to AI tools receives the same verbose, technically framed output as a student whose parents work in technology — with none of the context to interpret it.
Veterans, Justice-Involved Individuals, and Aging Adults
Veterans navigating benefits systems, justice-involved individuals re-entering society, and aging adults encountering AI for the first time share a common experience: AI as a foreign language spoken at full speed, with no accommodation for users who do not yet know the vocabulary.
Ungoverned AI assumes the user already knows how to engage with it. Governed AI does not make that assumption. That distinction is the difference between a tool that serves these populations and one that alienates them further from systems already designed without them in mind.
California’s Schools Have Already Experienced the Cost
These are not hypothetical concerns. California’s two largest school districts have already experienced consequential AI failures that demonstrate the cost of deploying AI without governance (CalMatters, 2024).
Los Angeles Unified launched an AI chatbot called Ed as a personal assistant for students, then shut it down within three months when the company that built it collapsed. San Diego Unified’s board signed a contract for curriculum that included an AI grading tool board members did not know was there.
In response, the California Legislature passed Senate Bill 1288 (Becker, 2024), mandating a statewide AI Working Group for K–12 schools, convened by the California Department of Education (California Department of Education, 2025). State Superintendent Tony Thurmond stated: “There is an urgent need for clear direction on AI use in schools to ensure technology enhances — rather than replaces — the vital role of educators” (GovTech, 2025).
That working group addressed data privacy, academic integrity, equitable access, and classroom integration. What it has not yet addressed is inference-level governance: how much AI should produce, for whom, under what constraints, and with what accountability. That is the gap this paper addresses.
Governed AI Benefits Every User
Governed AI is not remedial technology. It benefits every user — including those who are technically sophisticated.
A governed AI response is shorter, clearer, faster, and less computationally expensive. For an expert user, that means more efficient workflows and more precise output. For a student with a learning difference, that means reduced cognitive load. For a parent who does not speak English, that means a response matched to their language. For a veteran navigating a benefits claim, that means actionable guidance rather than comprehensive hedging.
The same governance that serves the most vulnerable user also serves the most sophisticated one. That is not a trade-off. That is the architecture.
Empirical prototype testing conducted by DeBacco Nexus LLC between January and April 2026 observed token reduction ranging from 67 to 98 percent across separate test tiers using different models and trial conditions, documented in timestamped logs with full methodology (DeBacco Nexus LLC, 2026). These results were observed in initial prototype testing and warrant further validation at scale. The reduction represents not only energy savings but a measurable reduction in cognitive load for every user regardless of prior knowledge.
A Hypothesis for California
Governed AI inference, applied as the default standard in California public schools and public services, will produce measurable improvement in engagement, comprehension, and trust among all users — with disproportionate benefit for students who are neurodivergent, learning-challenged, diagnosed with TBI or PTSD, or from disadvantaged and transitioning backgrounds — compared to ungoverned AI currently deployed without constraint or accountability.
This hypothesis is testable. The empirical foundation from prototype testing supports it. The documented failures in California’s largest school districts confirm the urgency. The passage of SB 1288 confirms the state’s recognition that a governance standard is needed.
CalCompute is positioned to build that standard — not as a remedial measure for vulnerable populations, but as infrastructure for every Californian who uses AI.
References
California Department of Education. (2025). California public schools: Artificial intelligence working group. SB 1288 (Becker, 2024). https://www.cde.ca.gov/ci/pl/aiineducationworkgroup.asp
CalMatters. (2024, August 6). California’s two biggest school districts botched AI deals. Here are lessons from their mistakes. https://calmatters.org/economy/technology/2024/08/botched-ai-education-deals-lessons
DeBacco Nexus LLC. (2026). Empirical research tier catalog: Inference governance module [Internal research documentation]. Patent Pending USPTO 19/571,156. Available upon request.
Deloitte Insights. (2026, February 6). AI tokens: How to navigate AI’s new spend dynamics. https://www.deloitte.com/us/en/insights/topics/emerging-technologies/ai-tokens-how-to-navigate-spenddynamics.html
Flapan, J. (2026, February 26). As cited in: AI images scandalized a California elementary school. CalMatters. https://calmatters.org/economy/technology/2026/02/ai-images-scandalized-a-california-elementary-school-now-the-state-is-pushing-new-safeguards
GovTech. (2025, September 2). California Education Department launches AI workgroup. https://www.govtech.com/education/k-12/california-education-department-launches-ai-workgroup
Ikangai. (2025, August 21). The LLM cost paradox: How “cheaper” AI models are breaking budgets. https://www.ikangai.com/the-llm-cost-paradox-how-cheaper-ai-models-are-breaking-budgets
WAV Group Consulting. (2025, December 16). AI token costs are invisible until they aren’t. https://www.wavgroup.com/2025/12/16/ai-token-costs-are-invisible-until-they-arent