“Public infrastructure built for residents needs to govern what comes out of AI models, not just what goes into them.”

The Problem Nobody Is Talking About

When I began studying how artificial intelligence affects the most vulnerable populations I work with — veterans, justice-involved individuals, people in crisis — I noticed something that the AI industry rarely discusses. The problem was not what AI said. It was how much it said.

A person calling a health hotline on a phone with 2% battery does not need a 300-word explanation of chest pain symptoms. They need four words: call 911 right now. A CalFresh recipient who speaks Khmer does not need a policy lecture. They need one clear sentence in their language telling them what to do next.

Ungoverned AI inference is a running faucet. It produces output until something external stops it. No architectural constraint governs how many tokens are generated, how much energy is consumed, or whether the response actually serves the person asking.

What the Prototype Data Shows

In prototype testing conducted on a consumer iMac — fully offline, across 23 California DHCS medical threshold languages — I measured what happens when a governance layer enforces a hard output ceiling before a response reaches the user:

  • Mean energy reduction of 17.4% on real hardware
  • Token reduction ranging from 82 to 94 percent across multiple runs
  • Every output delivered in the language the user actually speaks

These are early results. Prototype work. Not validated at scale. But they point to something California needs to take seriously as it builds CalCompute and prepares its framework report for the Legislature by January 2027.

The Governance Gap

The SB-53 framework asks the right questions about safety, transparency, and accountability. But those questions are being asked at the model development layer — where the most powerful AI systems are built. Nobody is yet asking them at the inference layer — where those systems actually touch people.

That is the governance gap California has an opportunity to close.

Public infrastructure built for residents — for health services, emergency response, education, and social services — needs to govern what comes out of AI models, not just what goes into them. It needs audit trails. It needs language equity. It needs output that serves the person in front of the screen, not the system generating the response.

California’s Opportunity

California is positioned to lead on this. CalCompute is the right context to ask the question that the private sector is not asking: what does governed inference look like at public scale?

I do not have the complete answer. But I know the question is real, the problem is measurable, and the 2027 framework deadline is closer than it appears.

James DeBacco is the founder of DeBacco Nexus LLC and a doctoral candidate at the USC Suzanne Dworak-Peck School of Social Work. He holds a provisional patent on AI inference governance architecture (USPTO 19/571,156). Contact: [email protected] · debacconexus.ai

Transparency: The 17.4% energy reduction figure is from prototype testing on an Intel iMac using powermetrics package power telemetry, 50 matched prompt pairs, April 3 2026 (chain hash: 159f6faad2bf9050). The 82–94% token reduction range is from multiple V8/V9 prototype runs on consumer hardware. All results are from prototype environments. Not validated at production scale. GPU-level validation is the identified next step. Patent pending — not yet granted.