Ballots arrived in California mailboxes carrying fourteen statewide propositions, twelve statewide races, and a stack of local measures thick enough to prop a door open. One voter in Menlo Park looked at the pile, looked at his calendar, counted the evenings between now and Election Day, and did the math on a free evening he did not have and was not going to find. Then he did something no voter could have done two months earlier: he handed the whole ballot to his agent.
Muse is Meta's personal AI agent, launched September 8, 2026, and it runs on its own computer with a browser, a memory that persists between conversations, and connections to the apps a person actually lives in, which is the part that matters for this story. This voter asked it for two things, personalized recommendations for every contest on the ballot and a one-page cheat sheet he could fold into his pocket and carry to the polls, and he asked for both in a single message sent from his kitchen table. What came back was useful, and it was also a warning about the difference between an agent that reads your ballot and an agent that reads it right.
Start with what the agent did, because the workflow is the story, and the workflow is simple enough to copy, which is precisely why its failure modes deserve a careful look rather than a shrug. Muse read the official voter guide, worked through the propositions one by one, checked the candidates against the voter's stated positions, and produced a single page: a line per contest, a recommendation, and a sentence of reasoning. It did the first pass in minutes, while reading the official analysis for fourteen propositions at a careful 200 words a minute runs roughly 70 to 90 minutes for the propositions alone, before a single candidate statement. That is the pitch in one number: an agent compresses an evening of reading into a page you can check in fifteen minutes. It worked. Mostly.
But the fifteen minutes matter more than the evening, because the check is the only thing standing between a useful summary and a confident error, and nobody writes a protocol for the part they skip. A cheat sheet is a filter, and a filter hides everything the agent chose not to mention. So the voter checked the sheet against the guide, line by line, and that check is where the story gets interesting, because the sheet was right on the propositions and nearly missed the one thing that mattered in exactly one race.
The race no voter guide contains
In the attorney general race, the agent recommended the incumbent, Rob Bonta, and gave a clean reason, experience and endorsements and a record the voter could check in an afternoon, which is exactly the kind of recommendation that feels complete until you know what it left out. Then, because this agent keeps a memory of the voter's life, a memory built from months of requests and files and standing instructions, it added something no voter guide would ever print and no search engine would volunteer. Bonta's office had sent a July 29 notice backing the voter's own land-use project under AB 712, a fact that lives in the voter's files and nowhere in the voter guide, which is why no amount of searching would have surfaced it without an agent that already knew the voter. And on October 27, Bonta's office joined a federal lawsuit, Oregon v. Bessent, seeking to restore beneficial-ownership reporting rules that would apply to the voter's own project company, which meant the same office was simultaneously defending the project and suing to regulate it. Same office. Same voter. Two directions.
Read the complaint and the tension holds up. Oregon v. Bessent, filed October 27, 2026 in the U.S. District Court for the District of Oregon as case number 6:26-cv-2359-MTK, was brought by twenty states and the District of Columbia, with California as a plaintiff. It challenges federal rules exempting domestic companies from reporting who owns them, and it asks the court to block the planned purge of ownership data already collected, a purge scheduled for February 10, 2027. If the states win, the reporting comes back, and it comes back for companies like the voter's, and if they lose, it does not. Either way, the voter's pick for attorney general was also a pick about his own company's paperwork, and he had not seen that until the agent put the two facts on the same page.
That is the strongest case for a personal agent over a generic chatbot, and nobody should miss why, because the difference between knowing the candidates and knowing the voter is the difference between a summary and a conflict surfaced. A chatbot answering cold knows the candidates, while an agent with a memory knows the voter. It can find the conflict that exists only in the voter's own history, which is exactly the conflict a voter guide cannot print and a search engine will not surface unless you already know to ask. It is also the strongest warning in this story, because the same memory that finds the tension is a memory the voter has to trust, and trust is the whole question.
He kept the pick, and he kept the tension too, written on the sheet next to the recommendation, which is the honest version of the story. An agent that only tells you what you want to hear is a yes-man with a browser. This one earned its keep by telling him something uncomfortable.
The numbers that should scare you
Now the part that should slow everyone down. Two independent studies published in fall 2026 tested chatbots on election questions, and both found error rates that would be disqualifying in any other product a person trusts with a consequential decision.
Institute for Strategic Dialogue ran 2,400 prompts across six chatbots, including models from OpenAI, Anthropic, Google, Meta, xAI, and DeepSeek, covering ten states, and the design matters because it tested the systems voters actually use rather than laboratory versions of them. Twenty-nine percent of English-language responses about voting were incomplete, inaccurate, or outdated, and Spanish-language responses were 16 percent less likely to be accurate than English ones, which means the voters with the least access to help got the worst answers. Weak spots clustered on logistics, the questions voters most need answered precisely, and some systems drew on outdated state government pages from the 2024 election, which is the kind of error that looks authoritative precisely because it cites a real government source for the wrong year.
Forum AI tested eight chatbots and found at least one factual error in almost 40 percent of answers to election questions, Agence France-Presse reported on November 1, and the study matters because it measured the questions voters ask the night before an election rather than the ones researchers ask in a lab. Errors ran heaviest on voting logistics. One model told a voter that Pennsylvania's mail-ballot request deadline was November 3, Election Day itself, and stated it with the same confidence it would use for a correct answer, which is the whole problem in one example. Real deadline was October 27, a week earlier. One wrong date. One lost vote.
A California study by the Young People's Alliance and CITED tested five chatbots on 72 questions about the state's 2026 races and propositions, and it found the failure modes that matter most for a voter who trusts a summary, including models that answered questions about the wrong year's measure. DeepSeek described the wrong or prior-cycle ballot measure in roughly half its responses to two of the six propositions studied, while Claude fabricated that an active Los Angeles mayoral candidate had already lost her primary, which is not a rounding error but a candidate erased from the race. When models cited ideologically identifiable sources, 26 of 27 responses cited only one side of the spectrum, which means the research looked thorough and was, in fact, a single perspective wearing a bibliography. None of this is a rounding error; it is the normal output of systems asked to summarize contests they half remember.
People are doing it anyway. Pew Research Center found in February 2026 that 49 percent of American adults use AI chatbots, up from 33 percent in 2024, a sixteen-point jump in two years that tracks the arrival of agents capable of doing things rather than merely answering questions. A September survey by the American Association of Political Consultants Foundation found about one in five voters consult chatbots for election news and voting, and that figure is the one that should worry anyone who has read the accuracy studies, because one in five is no longer an early adopter. A National Conference on Citizenship survey found two in five Americans used AI tools to search for election information, while only 40 percent trust the tools with logistical information like when, where, and how to vote. Wide use, thin trust. That combination only works if voters verify the logistics somewhere else.
How to use an agent on a ballot without getting burned
Here is the protocol, and it is the whole point of this article. Four rules. Do not ask the agent what a measure does. Point it at the official voter guide at voterguide.sos.ca.gov and require a citation to the guide for every recommendation, because a recommendation without a citation is an opinion, and an opinion is not what you asked the agent to produce. Muse can be given a source document, told to read it, and made to show its work in a file you can check, which is the one advantage it holds over a bare chatbot answering from memory. If the agent cannot cite the guide, the recommendation does not go on the sheet.
Second, check the logistics yourself, every time, on the Secretary of State's own site, because a deadline is the one fact on a ballot where being wrong by a week costs you the vote itself rather than just a worse choice. Deadlines, polling places, and drop boxes are the exact category where the studies found the worst errors, and a wrong deadline costs a vote in a way a wrong summary of a proposition does not. Third, read the sheet against the guide before you vote, line by line, which took this voter about fifteen minutes, and fifteen minutes is the price of the whole arrangement, so if you will not pay it, do not make the deal. Skip that check and the time you saved is an illusion. You did not save an evening; you outsourced your vote to a system that is wrong a third of the time.
Fourth, ask the agent what it does not know. Make it list the contests where the guide was thin, the candidates with no record to check, and the measures where the funding data was incomplete, because the gaps are the map of where your own judgment has to do the work instead of the agent's. A sheet that admits its gaps is worth more than a confident one, because the confident one is the one that fabricated a candidate's loss and cited one side of the spectrum 26 times out of 27.
The other side, at full strength
Fairness demands the strongest version of the case against all of this, stated at full strength and not strawmanned, and it is a strong case, stronger than most people who have already handed an agent their ballot would like to admit. An agent should not be anywhere near a ballot. The error rates are not edge cases: twenty-nine percent incomplete or wrong, forty percent with at least one factual error, and the failures cluster on logistics, the exact questions that decide whether a vote counts. Worse, the failure modes are invisible: a confident summary of the wrong year's proposition looks exactly like a confident summary of the right one. A fabricated candidate loss reads like news, and a citation list that quietly draws from one side looks like research.
A voter who delegates the reading delegates the thinking, and a one-page cheat sheet is a filter that hides everything the agent chose not to mention, which is the part of the deal nobody reads until the count is over and the result is certified. Most voters will not run the verification protocol in this article, because a protocol is work and the whole point of the agent was to avoid work, which is exactly how a convenience becomes a vulnerability. They will take the sheet, vote the sheet, and never know which line was wrong, because a wrong line on a cheat sheet looks exactly like a right one until someone checks it against the guide. Democracy does not survive that at scale, and no amount of convenience justifies it. That argument deserves to be heard in full before anyone hands an agent a ballot, and it is not answered by saying the agent is usually right, because usually right is not the standard for a vote.
It is also incomplete, because it assumes the alternative is a voter who reads everything carefully, which is not the alternative anyone actually has. Most voters do not read fourteen propositions; they skip the down-ballot races, guess on the local measures, or leave them blank. California's own count shows how many ballots are still being processed weeks after the polls closed, 8,906,017 counted and an estimated 4,816,722 still outstanding as of November 5, with county certification due December 3 and the Secretary of State certifying December 11. Proposition 40, the one-time 5 percent tax on billionaires, was defeated 52.8 percent to 47.2 percent, a margin of more than 600,000 votes, and the local Measure P in Menlo Park sat at No 2,886 to Yes 2,668 in early returns with the count still open. People voted on all of it, mostly without reading the guide. An agent with a verification protocol is not competing against a careful voter, because the careful voter who reads all fourteen propositions does not exist in any numbers worth counting. It is competing against a skipped race, and on that comparison it has a real chance, but only if the voter does the fifteen minutes, which is the entire thesis of this article in one sentence, and the sentence only works if the fifteen minutes actually happen.
What I could not verify
Honesty requires the gaps. This is one voter's case study, not a test across voters, states, or ballot lengths, and the workflow has not been measured at any scale, so everything in this article about the protocol is a design argument, not a measured result, and should be read that way. The error-rate studies tested bare chatbots, not Muse working from a specified source document, so Muse's error rate under the verification protocol in this article is unmeasured, and I am not claiming it is lower, only that the protocol is designed to catch errors rather than hide them. Proposition 40 and Measure P figures are semi-official as of November 10 and 11, and certification is not due until December 11, 2026. The July 29 AB 712 notice comes from the voter's own records, while the Oregon v. Bessent facts are verified from the filed complaint itself, which anyone can read, and that difference in sourcing is worth keeping in mind. No claim is made about how Muse would handle a ballot in a state with a different voter-guide format, and the time math uses stated assumptions, 200 words a minute and 70 to 90 minutes for fourteen propositions, which your ballot will not match exactly.
What to do next time
First, give the agent the official voter guide as its source and demand a citation for every line on the sheet, because a sheet without citations is a list of opinions, and opinions are not what you asked for. Second, verify every deadline, polling place, and drop box on the Secretary of State's site yourself, because logistics is where the models fail most, and a deadline is the one fact where the error does not just mislead you but disenfranchises you. Third, spend fifteen minutes reading the sheet against the guide before you vote, and keep any tension the agent surfaces written next to the recommendation rather than resolved away, because a tension you can see is a decision you own, and one you cannot see is a decision the agent made for you. Fourth, ask the agent to list what it could not verify, and treat those contests as yours to decide rather than its, because an agent that cannot show its work on a race has no business picking a winner in it. Fifth, remember that the count does not end on election night, and a close local measure can sit undecided for weeks, which is a reason to vote early and a reason to check the result after certification, not just on the night.
Your ballot is the most consequential document an agent will ever touch, more than your inbox or your calendar or your shopping list, because it is the one place where a confident error becomes a decision you cannot take back. It deserves a better deal than a question typed in a hurry. Give the agent the source, make it show its work, and keep the deciding for yourself. That is the deal. Then let it do the reading. You do the voting.