California Legislature Directory & 2026 Staff Audit: How AI Sees Power

This project has two halves that work together. The first is a directory: a clean, structured map of the California Legislature—people, roles, committees, and decisions. The second is a 2026 staff audit, Inside the Golden State: A Data-Driven Analysis of Influence and Inclusion, which uses that structure plus public payroll data to surface patterns of pay, representation, and power. Together, they show what responsible AI in government can look like—and what it cannot.

Modern AI is not an all-knowing brain. It is closer to a statistical inference and deduction engine: given structured data, it finds patterns, estimates probabilities, and generates summaries. It cannot see the future, read minds, or fix injustice on its own. What it can do—when we give it clean inputs—is help humans see the system they are already living in with more clarity.

Download the Tools

What the Directory Contains

The directory is organized the way a working legislature operates:

• Senate/ — Member folders by district (e.g., Allen_Benjamin_SD24), committees, and floor sessions.
• Assembly/ — Member folders by district (e.g., Addis_Dawn_AD30), committees, and floor sessions.
• Governor/ — Divisions of the Governor’s Office (Appointments, Cabinet, Communications, Legal Affairs, Legislative Affairs, and more).
• _Shared/ — HR assets, resumes, job descriptions, legislation references, district data, and extraction scripts that keep everything in sync.

Each member folder is built for real work:

• Legislation/ — Bills tracked by process stage (Ideas, In_Committee, Second_Reading, Third_Reading, In_Other_House, With_Governor, Enacted).
• District/ — Counties, major cities, constituent cases, and events.
• Correspondence/ — Incoming and outgoing letters, emails, and responses.
• Admin/ — HR, expenses, ethics filings, and compliance.
• Scheduling/ — Calendars and meeting requests.
• Press/ — Press releases, clips, and social media.
• Staff_Contacts.csv — Staff rosters built from salary data.

Under _Shared/Reference/ you’ll find staff salary data, legislative deadlines, rules, handbooks, district boundaries, and more—the kinds of reference materials that real staff rely on to track what is happening and when.

What the 2026 Audit Found

The 2026 audit takes that structure and asks a hard question: Who actually holds power inside this system? Using staff rosters, payroll data, and a retrieval-augmented AI pipeline, it surfaces a few core findings:

• The “Chair Premium” — Staff working for Committee Chairs earn about 30% more than staff for rank-and-file members. Committee leadership is the clearest path to higher pay.
• The Representation Gap — Hispanic/Latino residents are 40.3% of California, but make up only ~25% of staff and less than 20% of top executive staff in the Legislature and Governor’s Office.
• Pay Gaps Between Branches — Assembly staff have a higher median salary than Senate staff; Capitol staff are paid far more than District Office staff.
• Gender Paradox — The Governor’s Office and Senate executive tiers are majority female, while the Assembly’s executive tier is majority male, shaping how different parts of government may prioritize policy.
• Blind Spots — Black and Asian/API staff are significantly underrepresented in senior roles, and algorithmic inference from names may undercount them further.

These findings echo the project introduction: this is a transparency audit, not a verdict on any individual. The report is a probabilistic model’s best guess based on public records, meant to move conversations from anecdotes and rumor toward evidence and measurable gaps.

How AI Was Actually Used (RAG, Not Skynet)

The audit uses a Retrieval-Augmented Generation (RAG) pipeline instead of asking an AI model to “invent” answers. A single JSON file (staff_data.json) holds the calculated counts, medians, and percentages for each branch, party, and hierarchy tier. That JSON becomes the source of truth.

• AI is given the structured data first, then prompted with questions like: “Which branch has the highest disparity between executive and support staff for Hispanic representation?”
• The model can summarize, compare, and explain—but it cannot change the underlying numbers.
• All insights are traceable back to the directory-style structure and the staff_data JSON.

This is what responsible AI in government looks like in practice: retrieval first, generation second. Facts are anchored in verifiable records. Narrative is layered on top, not substituted in.

What AI Is (and Is Not) in This Project

Used this way, AI behaves much less like science fiction and much more like an ultra-fast research assistant that only works with what you give it:

• AI is pattern detection. It infers gender and ethnicity from names, then highlights gaps between staff and population. It does not “know” anyone’s identity; it works in probabilities.
• AI is structured summarization. It turns hundreds of pages of tables into a few paragraphs about the Chair Premium, pay gaps, and glass ceilings.
• AI is scenario testing. You can ask, “What happens if we focus recruitment on underrepresented tiers?” and get back reasoned projections grounded in existing numbers.

At the same time:

• AI is not an all-knowing oracle. It cannot see beyond the data you feed it. If communities are undercounted in the inputs, they will be undercounted in the outputs.
• AI is not moral authority. It can describe disparities; it cannot tell you what justice requires. That belongs to people, law, and politics.
• AI is not a replacement for lived experience. An algorithm can flag a pattern; only humans inside the institution can explain why it exists and how to change it.

Enforced Sameness, Fake News, and Why Structure Matters

One part of the report looks beyond numbers to a concept called “Enforced Sameness”—institutions that look diverse on paper but quietly punish dissent, direct communication, or different moral vocabularies. In those environments, people learn to self-censor long before formal rules are broken.

AI can either make that problem worse or better:

• If AI is trained on choreographed talking points, it will reinforce choreography.
• If it is grounded in real staff structures, pay data, and documented gaps, it can surface tensions that are already there.

Fake news thrives where no one can see the underlying tables: who is in the room, who gets promoted, who gets paid what, and which communities are left out of leadership. The combination of this directory and this audit moves in the opposite direction—toward:

• Visible hierarchies (Executive, Senior, Legislative, Support)
• Measurable gaps (e.g., 40.3% Hispanic population vs <20% in executive tiers)
• Checkable claims (every chart in the PDF ties back to staff_data and public records)

From Files to Democratic Practice

The scripts bundled with the directory—PDF extractors, district setup tools, and LLM pipelines—exist for one reason: to turn messy, scattered public records into something you can actually work with. That work is not glamorous, but it is where “AI in government” really lives.

• Extract unstructured documents into consistent text.
• Normalize names, districts, committees, and dates.
• Store them in predictable places with predictable formats.
• Use AI to help search, summarize, and compare—but always against the underlying source documents.

If a human can’t clearly categorize it, a machine definitely can’t. The California Legislature Directory and the 2026 staff audit are built from that premise. They are not a new brain for the state—they are a clearer mirror. Used well, they can help staff, advocates, and the public separate fact from fiction in how power and inclusion actually work in Sacramento.