AI Training & Change Management in Sacramento: AI Governance for State and Civic Systems

Start with data transparency: What populations does the AI apply to? Build a dataset of 500–1000 historical applications disaggregated by: citizenship status (citizen vs. noncitizen), country of origin, language of application (English vs. Spanish or other), immigration status (documented vs. undocumented). Run the AI system on this dataset and compare denial rates: Do noncitizens have higher denial rates than citizens? Do applicants whose primary language is not English get denied at higher rates? Does immigration status affect approval? Document any disparities explicitly. Then ask governance questions: 'If the AI denies noncitizens at 20% rate and citizens at 10% rate, why? Is that because noncitizens have fewer assets (legitimate), or because the model is discriminating (problematic)?' Investigation often reveals that proxies in the data (address type, employment status, family structure) correlate with immigration status, and the model is picking up those proxies. Once you identify the disparity source, you can decide: rethink the model, adjust thresholds, or require additional human review for flagged cases. Transparency about disparities is the starting point; action is required.

Cautiously, and with extensive validation. A dispatch AI that accurately predicts call urgency 85% of the time in testing might perform worse in production when it encounters novel scenarios or community members it was not trained on. Best approach: Deploy AI in 'advisory mode' first — the system makes recommendations, but dispatchers are explicitly trained to validate and override. Track: When did dispatchers override the AI? Why? What was the actual call severity? Over 3–6 months, you gather data on system accuracy and failure modes. Only expand AI authority after validation. Also, establish clear protocols: 'If there is ANY doubt, treat the call as high-priority.' A false negative (treating an urgent call as low-priority) is worse than a false positive (treating a routine call as urgent). The system bias should always be toward caution in life-safety contexts. Build that into training: 'This AI system is a tool to help you prioritize. It is not the final word. Your judgment and gut instinct matter. If you doubt the AI's recommendation, override it.'

A multi-step process: (1) Identify affected roles 12–18 months before AI deployment. Job-impact analysis shows which roles will change.; (2) Work with agency and union to design redeployment pathways: 'Clerical staff can transition to data-analysis support or customer-service roles. Here are the new job descriptions and pay ranges.'; (3) Offer training: 12–24 week programs preparing workers for new roles. Offer during work hours (not off-hours) so participation is realistic.; (4) Support placement: Career counseling, mentorship, and active placement support. A worker who completes training should have a new position waiting.; (5) Protect pay: If a worker's new role has lower pay than their previous role, protect their previous pay level for 12 months while they gain proficiency.; (6) For workers who cannot or will not transition: Offer voluntary severance (typically 1 month per year of service), enhanced retirement benefits, or outplacement services. Expect 50–70% of affected workers to successfully transition to new roles, 10–20% to choose severance/retirement, and 10–20% to remain in existing roles (maybe with modified duties). The total cost is significant (training, wage protection, severance) but protects workers and maintains morale. California's union workforce will not accept silent automation; transparent, supportive transition is required.

Establish decision tiers based on impact. (1) Low-impact decisions (internal workflow automation, routine data processing): AI can make decisions autonomously, with human audit trails. (2) Medium-impact decisions (eligibility for services, resource prioritization): AI recommends, humans decide. An AI system can say 'approve this welfare application,' but a social worker makes the actual determination and has override authority. (3) High-impact decisions (denying critical services, police dispatch, criminal-justice recommendations): AI provides analysis, humans decide based on multiple inputs including AI assessment. No autonomous AI decisions in this tier. That tiering acknowledges that efficiency matters but not at the cost of accountability. California state agencies should codify these tiers in policy and train staff and leadership on the framework. Also ensure transparency: When a high-impact decision is made (deny benefits, dispatch decision), the process should be documentable: 'The AI assessment was X. The human decision was Y. The reasoning was Z.' That documentation enables audit and appeal, which are essential for public trust and legal defensibility.

Loss of public trust and eventual forced system removal. California residents and advocacy organizations scrutinize government AI use intensely (due to history of discriminatory enforcement, invasive surveillance, and inequitable service delivery). If the state deploys AI to welfare, criminal justice, or emergency dispatch without transparent communication about how it works and what safeguards are in place, media coverage and advocacy pressure will mount. Eventually, a problem surfaces (disparate impact discovered, algorithm error harmed someone), and the public calls for removal. The system gets paused or eliminated after expensive deployment. Transparency upfront (months before deployment): 'Here is the AI system we are testing. Here is how it works. Here are the equity and accuracy concerns we have tested for. Here is how you can provide feedback.' That transparency prevents the backlash and builds legitimate public confidence. Sacramento state agencies should treat transparency as a project requirement, not an afterthought.