Three days before the California primary, I sat down to research eight races I knew almost nothing about — insurance commissioner, controller, treasurer, BOE district, county superintendent, assessor-clerk-recorder, a US House seat, and a school parcel tax. I told the AI my axis (center-right, libertarian-flavored, no fringe candidates) and let it pull candidate positions, endorsements, and the consolidating-party signal for each race. Then I drilled in: for the superintendent race I added “anti-CTA, pro-testing, pro-school-choice, pro-parental choice but without the culture-war framing” and got a candidate who actually matched.

Three things were genuinely useful. First, the strategic frame: California’s top-two primary makes “don’t dilute the non-D vote” a real consideration, not paranoia — knowing which non-D candidate has the consolidating party endorsement matters more than my preference between two unknowns. Second, the data the AI surfaced on the local school parcel tax — per-pupil spending already 67% above the state median, +42% over four years, mediocre proficiency — turned an emotional ballot pamphlet into a quantitative question. Third, on a candidate I’d been told was “extreme,” the actual record showed a focused single-issue activist whose policy direction was anti-libertarian (more state intervention in family decisions, not less), which mattered more than the loaded label I’d been handed.

The AI didn’t choose for me. It made the choosing legible. Resources it leaned on: CalMatters voter guides, Ballotpedia, EdSource, Berkeley IGS and PPIC for the only state-level polling that exists, the Secretary of State and county registrar pages for ballot-measure text, and the local papers (Almanac, SM Daily Journal, Palo Alto Daily Post, KQED, LAist) that do the actual reporting on county races nobody else covers. Next time, start more than three days out — the research-and-clarify loop gets meaningfully better the more specific I get about my own preferences, and “more specific” turned out to mean asking myself harder questions than I’d asked in past cycles.