利用 AI 进行选举研究
TL;DR · AI 摘要
AI 可以帮助选民在选举前进行研究，提高投票效率。通过分析候选人的背景、政策和立场，AI 提供有价值的见解。
核心要点
- AI 可以帮助选民在选举前进行研究，提高投票效率。
- AI 可以分析候选人的背景、政策和立场，提供有价值的见解。
- AI 可以为选民提供关于特定职位的详细信息，如教育部长的职责和候选人资格。
结构提纲
按章节快速跳转。
思维导图
用一张图看清主题之间的关系。
查看大纲文本（无障碍 / 无 JS 友好）
- AI 在选举中的应用
- 提高投票效率
- 分析候选人背景
- 提供政策见解
金句 / Highlights
值得收藏与分享的关键句。
AI 可以帮助选民在选举前进行研究，提高投票效率。
AI 可以分析候选人的背景、政策和立场，提供有价值的见解。
AI 可以为选民提供关于教育部长职位的详细信息，如职责和候选人资格。
Use AI This Election - by Scott Alexander
Use AI This Election
...
May 26, 2026
I’m not saying AI is superintelligent or can decide better than you can. I’m saying that if you - like me - spend an hour or so doing research before voting on local seats, AI can aid that research very effectively. And if you don’t do that research - because you weren’t willing to waste an hour on it before - AI makes it so much faster that you might want to start.
I gave Claude a prompt something like (edited for coherence):
I’ll be voting in the June 2026 California primary. I’m a centrist liberal abundance YIMBY whose favorite political writers are Kelsey Piper, Matt Yglesias, and Ezra Klein. I’m wary of government overreach, but I’m not a doctrinaire libertarian and want to help people when we can figure ways to do it that work. I’m going to ask you about each race on my ballot, and I’d like for you to list the various candidates’ bios, policies, endorsements, your read on the most important differences between them, and your advice for me as I try to make my choice.
…and got back answers like the following:
## California Superintendent of Public Instruction
A quick reality check on the office, then the field. The state superintendent has surprisingly little direct authority — districts control most budget and curriculum — and the state is actively considering shifting more responsibilities to the governor-appointed State Board of Education, which would shrink it further. The job is mostly a $150B-budget bully pulpit. So this is more “who do you want amplifying education priorities for the next four years” than “who will rewrite California education.”
The race is on the standalone education ballot (it doesn’t follow the nonpartisan blanket primary system) — June 2 primary, November 3 runoff if no one clears 50%. Realistically no one does, so this primary is about who advances.
### Quick framing: what “the Kelsey Piper view” implies
Piper’s project, especially in her recent Argument writing, sits at the intersection of several positions that don’t all naturally cluster:
Structured literacy / “science of reading” / Mississippi Miracle is real and works. Phonics-first, mastery-based, with district-level accountability and reading coaches in low-performing classrooms. She has called the failure to adopt this “a crime against humanity” in California.
Standardized testing and accountability are good, actually. Not as bludgeons but as honest feedback mechanisms. Detracking failed. Grade inflation is bad.
Open to charters, microschools, and school choice. She founded and runs Oakland LEARN, a microschool. She sends her own kids to an “unregulated tiny private school.” She thinks the left’s reflexive hostility to choice misunderstands what parents are actually fleeing (kids not learning to read).
Skeptical of teachers’ unions when they block reform. Not anti-teacher, but pro-evidence over pro-union-priority when they conflict.
Pro-public-school in principle, but pro-evidence over pro-system. Treats education as a place where the left has often chosen ideological commitments over the well-being of kids who do not vote.
The candidate who matches this view best probably (a) has explicitly backed structured literacy/phonics, (b) has shown willingness to push the system rather than just defer to CTA, and (c) is at least not hostile to charters or alternative models. The candidate who matches it worst would be a CTA-backed insider running on “more funding, less testing, no curriculum mandates.”
### The candidates
Al Muratsuchi (Assemblymember, LA South Bay, former Education Committee chair). Co-authored the law that funded phonics-based reading instruction grounded in the science of reading. Co-authored Prop. 2 ($10B school facilities bond, passed). Adjunct college teacher, former deputy AG. Endorsed by California Federation of Teachers (CFT, the smaller of the two teachers’ unions), Association of California School Administrators, and California School Employees Association.
Anthony Rendon (former Assembly Speaker 2016–2023). Ran an early-childhood program in LA for 20 years before politics. As Speaker oversaw major K-12 funding increases, universal transitional kindergarten, and charter school accountability legislation (which generally tightened charter regulation rather than loosened it). Endorsed by AFSCME, SEIU California, California Latino Caucus.
Joshua Newman (former state senator, former Senate Education Committee chair). Co-authored Prop. 2 with Muratsuchi. Army veteran, founded a vets-to-civilian-workforce nonprofit, now a UC Irvine fellow. Endorsed by State Building and Construction Trades.
Richard Barrera (San Diego Unified school board president). The California Teachers Association candidate — that’s basically his profile. Outsider to Sacramento; emphasizes funding, early childhood, easier teacher credentialing pathways.
Nichelle Henderson (LA Community College trustee, longtime teacher). Union-backed. Supports expanded school-based health, standardized testing reform, divesting pensions from fossil fuels and “the U.S. war machine.”
Frank Lara (SF elementary teacher, former UESF VP). The left-of-the-Democrats candidate. Peace and Freedom Party and Green Party endorsements (officially nonpartisan, but signals where he sits).
Gus Mattammal (Silicon Valley tutoring company founder, author). Pro-charter, pro-homeschool, pro-alternatives-to-traditional-public-schools. The closest thing to a school-choice candidate in the field.
Sonja Shaw (Chino Valley School Board president). Anti-trans-protections crusader who got Thurmond escorted out of a 2023 board meeting. CA Republican Party, Moms for Liberty, CRPA endorsements. The right-culture-war candidate.
### The most important differences, and my read for the Piper view
The structured literacy litmus test points hard to Muratsuchi. He is the _only_ candidate in this field who has actually moved policy on the thing Piper considers most urgent. His official candidate statement explicitly cites “phonics-based reading instruction based on the science of reading” as something he coauthored and delivered funding for. That is, almost verbatim, the policy Piper has spent the last year writing about. Nobody else in the field has a comparable record on this issue.
Rendon is the second-strongest plausible fit, but with caveats. Twenty years running an early-childhood program is genuinely relevant experience, and TK (transitional kindergarten) expansion is one of the more evidence-supported things California has done in K–12. But his charter-accountability work as Speaker generally meant _more_ regulation on charters, which cuts against the Piper-flavored openness to alternative models. He’s a competent establishment Democrat; he’d be a fine superintendent, just not specifically the structured-literacy crusader.
Barrera is the candidate Piper would be most skeptical of. Not because he’s a bad person — he’s a long-serving school board president, presumably competent — but because the CTA endorsement _is_ his political identity in this race, and Piper’s whole thesis is that CTA-aligned politics has been one of the structural reasons California can’t adopt the reforms that worked elsewhere. Barrera’s stated priorities (more funding, easier teacher pathways) are CTA priorities, not reform priorities. If you’re voting the Piper view, this is the one to skip.
Henderson and Lara are further from the Piper view. Henderson’s “standardized testing reform” plus the pension-divestment-from-the-war-machine framing reads as movement-left rather than evidence-driven; Piper has been pointedly critical of detracking and of treating testing as the enemy. Lara is more or less explicitly the union-left candidate.
Mattammal is the awkward case. He is the candidate closest to Piper on school choice and microschools specifically — that’s basically her own model. But he’s running with essentially no political coalition or governing experience, and the Piper view, while choice-friendly, is not “burn down public schools” — it’s “fix the actual reading instruction in the schools that most kids will continue to attend.” A statewide superintendent who can’t move CTA-dominated Sacramento isn’t useful for that goal. Mattammal also doesn’t appear to have an articulated policy program on structured literacy or accountability, just on alternatives. Probably not the right fit even on a choice-sympathetic reading.
Shaw is obviously off the table on the Piper view. Piper is a Bay Area liberal who supports trans kids and has no interest in culture-war education politics. The Moms for Liberty endorsement is disqualifying on its own.
Newman is the dark horse. Co-authored Prop. 2 with Muratsuchi, chaired Senate Education, military background, now an academic fellow rather than a Sacramento insider. He hasn’t been as visibly identified with structured literacy as Muratsuchi has, but he’s broadly in the same evidence-and-accountability lane. The Building Trades endorsement is unusual — it suggests he’s running a coalition strategy aimed at facilities/jobs rather than CTA — and that might actually be a positive signal for someone trying to operate independently of the dominant teachers’ union politics.
### Bottom line
For the Piper view, Muratsuchi is the clearest fit in the field. He’s the only candidate who has actually legislated structured literacy funding, he has the directly relevant policy chops (former Education Committee chair, Prop. 2), and his endorsement coalition (CFT, administrators, classified employees) is just slightly off-axis from the dominant CTA-controlled lane in a way that suggests some independence. He’s also reasonably likely to make the November runoff, which matters: a vote for the second-best candidate who finishes fourth doesn’t accomplish anything.
If you wanted to express a sharper “I want a real outsider reformer” view rather than working within the field as it exists, Mattammal is the only option, but at significant cost in viability and breadth of program.
For “I just want a competent Democrat who’ll be fine,” Rendon and Newman are both reasonable. But for “I want the candidate most likely to actually push the structured-literacy / accountability agenda Piper writes about,” it’s Muratsuchi.
This is better than any voter guide I’ve ever seen. I didn’t end up taking Claude’s advice - I voted Newman - but Muratsuchi was a close second, and Claude’s advice was crucial in helping me prioritize what to research further.
And here’s its opinion on one of the small local spending measures I usually have a hard time forming an opinion on:
## Measure A: Peralta Colleges Affordable Education Reauthorization
### What it actually does
This is a parcel tax reauthorization, not a new tax. The Peralta Community College District (which runs Laney College, Merritt College, College of Alameda, and Berkeley City College) has had a $48-per-parcel annual tax since voters first approved it in 2012 and renewed it in 2018. The current authorization expires.
Measure A would:
* Extend the existing $48/parcel tax for 9 more years (July 1, 2027 through June 30, 2036)
* No increase in the rate — same $48 you’re already paying if you own property in the Peralta district
* Generate roughly $8 million per year ($72M over the nine years)
* Funds restricted to instruction, student services, counseling, faculty recruitment/retention, job-training programs, and university-transfer prep
* Requires a 2/3 supermajority to pass (this is the standard threshold for California parcel taxes and is a high bar)
Senior citizens and Supplemental Security Income recipients can apply for exemptions. Funds are subject to a Citizens’ Parcel Tax Oversight Committee.
Geographic scope: this only appears on ballots in the Peralta district, which covers northern Alameda County — Oakland, Berkeley, Alameda, Albany, Emeryville, and Piedmont. So if you’re in Oakland, you’ll see it.
### The straightforward “yes” case
* Community colleges are genuinely one of the most cost-effective rungs in American higher ed. Peralta tuition is ~$1,100/year vs. ~$6,500 at Cal State and ~$14,000 at UC.
* Reauthorizing an existing tax is a much lower ask than a new one — you’re already paying it, and nothing on your tax bill changes.
* The money is restricted to instruction and student services (not administration, not facilities), with citizen oversight. That’s about as well-designed as parcel tax measures get in California.
* League of Women Voters supports it. Labor unions back it. No formal opposition argument was filed with the Registrar of Voters — literally no one wrote a ballot statement against it.
* $48/year on a parcel is genuinely small relative to most California property tax bills.
### The complications worth knowing
I want to give you the honest picture rather than just the campaign pitch, because Peralta has a difficult institutional history that someone with your “wary of government overreach but pro-helping-people” framework will want to understand.
The accreditation crisis was real. From 2018 to 2023, Peralta was in serious trouble with its accreditor (ACCJC). A 2019 state Fiscal Crisis and Management Assistance Team review found Peralta was “at high risk of insolvency” after years of mismanagement, with administrators “regularly breaking their own rules” and the district “ripe for fraud” due to poor internal controls. The four colleges were placed on probation by the accreditor in January 2020 — a severe sanction one step short of “show cause” (the level that nearly closed City College of San Francisco).
The governance situation was also a mess. A 2021 Alameda County civil grand jury report excoriated the Peralta Board of Trustees for “broken board culture,” secretive meetings, infighting, interfering in hiring, racial insensitivity, and fostering an “overall culture of incivility.” Multiple chancellors departed under pressure — Peralta went through four chancellors between 2019 and 2023.
Things have improved. Accreditation was restored from probation to “warning” status in January 2022, then fully reaffirmed in January 2023. Tammeil Gilkerson became permanent chancellor in January 2024 — by all accounts a stabilizing hire from outside the dysfunction.
But there are ongoing concerns. As of early 2025, Peralta is still navigating a budget crisis. Chancellor Gilkerson has acknowledged that Peralta’s full-time-equivalent students (FTES) lag significantly behind other four-college community college districts in the state — which means enrollment-driven state funding is structurally low for the district’s size. There’s also a Peralta Citizen article headlined “PCCD is in crisis. Everything is on the table” from February 2025, indicating that even with the recent improvements, the district isn’t out of the woods.
One transparency concern worth noting from local reporting. The largest single donor to the Yes-on-A campaign is Future Flooring Group (C&S Flooring Systems), which contributed $75,000 — about half the campaign’s total raised. Future Flooring is currently contracted with Peralta for ~$678K to replace flooring at College of Alameda, Laney, and the district service center this summer, and has received over $2M from the district overall. This isn’t illegal, but a major vendor that depends on district contracts being the largest single funder of the campaign to extend the district’s parcel tax is the kind of thing that should at least be visible. The rest of the campaign funding comes mostly from Bay Area labor unions, which is more typical and less notable.
### My honest read for your framework
For a centrist liberal abundance YIMBY who wants to help people but is wary of government overreach, here’s how I’d think about this:
Arguments for yes:
* Community colleges are _exactly_ the kind of institution that fits an “abundance liberal helping people” framework: they provide cheap, accessible higher ed and workforce training without the racket-like price escalation of four-year schools. They’re a genuine social ladder.
* This is a reauthorization, not new spending. Voting no doesn’t return the money to you in a meaningful way at the household level; it just defunds the colleges.
* $48 a year is small money. The cost-benefit is hard to argue against in isolation.
* Peralta has measurably improved governance since 2023. Punishing the institution now for sins it has been actively repairing would be unfair to current and future students.
* The supermajority threshold (2/3) is itself a check — if a tax can clear that bar, it has unusually broad consent.
Arguments for no:
* Peralta’s institutional dysfunction was severe and is not fully resolved. The “everything is on the table” framing from district leaders in early 2025 suggests structural problems persist.
* Some abundance-school thinkers would argue that propping up underperforming public institutions with parcel taxes reduces pressure to actually fix what’s broken — if the colleges had to demonstrate enrollment growth and operational excellence to retain funding, they might.
* Low FTES is a worrying signal. Are these colleges serving the demand that exists, or are they an institution that has become inefficient relative to peers?
The judgment call. I think a reasonable abundance liberal would vote yes, with eyes open. The reasons: (1) reauthorizations are different from new authorizations — you’re not signing up for new spending, you’re declining to actively cut it; (2) community colleges as a category are deeply aligned with the “help people, expand the ladder” framework; (3) the most damning Peralta criticisms are about 2018–2022, and the institution has been visibly improving since; (4) no organized opposition emerged, which usually means even motivated critics didn’t think the case for opposing was strong; and (5) $48/year is genuinely small money for a meaningful local good.
But voting no isn’t crazy if your prior is “underperforming institutions should face budget discipline.” It’s an unusually defensible “no” for a parcel tax, in that the institution has a real track record of problems. If you wanted to send a message that you support community colleges in principle but want Peralta to keep cleaning house, abstaining is also coherent (the 2/3 threshold means abstaining has a slightly different effect than a no vote — it raises the bar relative to the participating electorate).
My weak lean: yes. Strong enough to recommend, not strong enough to argue with someone who voted no on accountability grounds.
I informally (no strict controls) tested this on 10 difficult races where I wasn’t sure going in who I would vote for. I give Claude an A if it recommended the same candidate I eventually voted for, a B if it recommended my second choice out of a large field, a C if I can see where it’s coming from but disagree, and an F if it picked someone I hate. Its final score was:
- 5 x As
- 3 x Bs
- 2 x Cs
When it failed, it was usually because I have a policy of making my vote count on important races, vs. making fun/utopian/conscience-guided choices on races that don’t matter. I didn’t tell Claude this, and Claude tended to ignore fringe candidates with no chance of winning.
Other times, it was because it took my prompt too seriously. I _do_ like abundance/YIMBY ideas, but since that was all I told Claude in the prompt, it sort of pegged me as a single-issue voter. In the future, I might put that lower down, below the list of famous people I agree with.
Claude also correctly guessed (probably from the word “liberal”) that I lean Democrat, over-indexed on that, and leaned away from Republican candidates. In the future, I might include in the prompt that it should cross the aisle when there’s a sufficiently convincing reason.
I ended by asking Claude to ignore my personal beliefs and tell me who it would recommend, point-blank, if it didn’t know anything about me. It proposed four switches - three of which I liked better than its original choice! I’m skeptical that it was truly expressing its own opinions here (if it even has any) - I suspect it treated the exercise as a chance to make more “emotional” judgments - but it was an interesting exercise. If you’re a Republican using AI to help you vote, please end with the same question and tell me what happens.
This was a good enough experience that if I didn’t have an hour or two to do this properly, I would trust Claude’s endorsement over competing cheap-and-fast ways to determine my vote (party-line, newspaper endorsements, NGO endorsements, who has the funniest name). And given that I did spend two hours doing this properly, I think I’m about 50% happier with my choices than I would have been without AI.
Moving back into speculative futurism (in case you forgot which blog you were on), a lot of my hope for how the post-AGI future might turn out well - despite all the risks of dictatorship, inequality, surveillance, etc - hinges on people taking advantage of AI advisors to make better political decisions. The sooner we start, the better.
You can see my final list of decisions (may not be very interesting to non-Californians) here. Remember that I accept voter guides from ACX meetup groups in big cities with local elections happening soon, and will signal-boost them on the Open Thread if you give them to me.
- * *
#### Subscribe to Astral Codex Ten
By Scott Alexander
P(A|B) = [P(A)*P(B|A)]/P(B), all the rest is commentary.
By subscribing, you agree Substack's Terms of Use, and acknowledge its Information Collection Notice and Privacy Policy.
Previous
#### 134 Comments
[9h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265629511 "May 26, 2026, 8:28 PM")
Unsurprisingly, AI's first choice was a guy called AI.
[8h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265637055 "May 26, 2026, 8:43 PM")
I used to find it amusing when mentions of AI online would sometimes be met with (presumably unironic) questions like "Who's Al?" (with a lowercase L). I guess those days are long gone.
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265667617 "May 26, 2026, 9:48 PM")
Several years ago, the Dairy Queen next to my house had a big sign that I could only read as "AI Beltbusters are here", but I believe was supposed to be a reference to A-1 steak sauce on a sandwich, rather than misaligned AI making us fat.
[4h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265754236 "May 27, 2026, 1:23 AM")
Perhaps among the casualties of the Al age will be sans-serif fonts.
(I said Al by the way not AI)
[9h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265630125 "May 26, 2026, 8:29 PM")
Using frontier LLMs with web search to research candidates and policies is good. I would use them to iteratively produce a customized matrix of ratings. Once it looks good enough, make the final selections oneself.
[9h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265630594 "May 26, 2026, 8:30 PM")
Did you try just telling it that you are Scott Alexander?
[8h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265638194 "May 26, 2026, 8:45 PM")
honestly worth a shot
[8h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265641844 "May 26, 2026, 8:53 PM")
more useful demo to the rest of us to do it this way, since it provides some beta on how to tune the prompt.
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265664302 "May 26, 2026, 9:41 PM")
Very true! But still, if the AI *does* know who you are pretty well...
[4h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265750082 "May 27, 2026, 1:12 AM")
I just say "I'm Scott Alexander, tell me how to vote"
[1h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265802464 "May 27, 2026, 3:48 AM")
😂😂😂😂
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265668180 "May 26, 2026, 9:49 PM")
He says his favorite writers are Kelsey Piper et al. but you can see his revealed preference from the fact that he spends far more money on Scott Alexander.
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265686811 "May 26, 2026, 10:32 PM")
I encourage anyone to ask an AI how it thinks Scott Alexander would vote, and report back here!
[6h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265705340 "May 26, 2026, 11:17 PM")Edited
Katie Porter or Xavier Becerra
This after four back and forth’s of equivocation
This is also just the standard ChatGPT since I’m poor (I’m actually not I guess anymore really but still, not ready for an AI bill)
Also ChatGPT has changed for me lately - I feel it’s put me on a list of some kind and it no longer wants to have fun with me and gives me a lot of ‘ I’m not doing that ‘
I feel this may be my paranoia of ((something)) since I was born in communist Poland and see the world through a prism of ‘ all the communists will take things from you and force you to live a multicultural existence half a world away ‘ where communists is just a stand in for people in power
[3h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265764135 "May 27, 2026, 1:50 AM")
You're not on a list, it's changed for everyone (overcorrection from 4o-style sycophancy I think)
[5h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265714542 "May 26, 2026, 11:40 PM")
Scott has posted parts of his votes here, and his full ballot on the subreddit. An AI may very well find that.
[2h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265799597 "May 27, 2026, 3:38 AM")
i asked sonnet 4.6 and it was only willing to make a top-line guess about the governor's race, where it guessed Matt Mahan. I am not sure if this is who Scott voted for but it definitely would not surprise me and I am pretty confident if you told it "my favorite writers are yglesias/piper/klein" it would also generate the Matt Mahan recommendation.
[2h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265788532 "May 27, 2026, 3:01 AM")
LLMs are apparently able to recognize the writer from a few paragraphs these days, particularly if he has plenty of attributed online writing (like Scott does).
[2h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265789123 "May 27, 2026, 3:03 AM")
Yeah, I don't think the prompt he posted is enough but if he conversed about the candidates back-and-forth with Claude much I would be very unsurprised if Claude could guess, at least if was one of the most recent Opus revisions.
[9h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265632467 "May 26, 2026, 8:34 PM")
It's funny how quick the current generation of chatbots are to stereotype. Any input you give them, they'll treat as defining your whole identity. The obvious mitigation is to give them a _lot_ of input, but that creates more work for you and doesn't consistently help.
(I'm not sure what the technical explanation for this is. Subtlety is hard to RLHF on? Transformer architectures at low temperatures are incorrectly biased toward the "most likely" option-- i.e. they'll assume a 45% probable outcome 95% of the time? Overfitting to Gricean interpretations?)
The real trick is how to adjust for this on the other half of the equation-- the candidates and their media coverage...
[8h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265637558 "May 26, 2026, 8:44 PM")
I always have to ask them - "please don't overindex on XYZ!"
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265668234 "May 26, 2026, 9:50 PM")
The same is true for people - if they know one fact about you, or one fact about a candidate, they'll treat that as defining all of you or all of the candidate. (But AI does it in other contexts too - it tries to make every sentence of an essay or story read like the most important sentence, and it tries to include some reference to the prompt of an image in every little section of the image.)
[9h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265632783 "May 26, 2026, 8:34 PM")
I'm torn. This makes a lot of sense as a cost-benefit calculation right now, but I really hate the idea of giving AI companies (and AIs themselves!) this much power. Small choices in post-training could have enormous downstream political effects through this kind of influence, and some of the most concerning non-Terminator worlds are the ones where chatbot advisors gradually amass enormous political power via everyone deferring to them.
[9h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265633829 "May 26, 2026, 8:36 PM")
This seems like a particularly concerning lever of power because it's so fuzzy and unaccountable -- almost anything a lab (or AI) wants to accomplish could be achieved in a way that is plausibly deniable as “just making the model give better advice by my lights”.
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265664946 "May 26, 2026, 9:42 PM")
Also, if it lies (or has a stronger tendency to lie in a particular direction than the opposite), it can be excused as a hallucination, i.e. a mistake. With Fox or CNN you have bounded distrust: they may distort or have a slant, they won't blatantly lie on purely factual matters (https://www.astralcodexten.com/p/bounded-distrust). With AI it should be unbounded.
[8h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265637675 "May 26, 2026, 8:44 PM")Edited
Same. Scott's particular example here doesn't bother me, but I do worry about a future where people outsource all major (and even not so major) decisions they make to AI advisors and, unlike him, don't do much follow-up or critical thinking after receiving the advice.
Ken Liu's short story "The Perfect Match" imagines a future like this, where not taking advice from your AI even on personal matters like where to take a woman on a date is abnormal and a social faux pax. It's very depressing. I'm not sure where to draw the line, but there is a line somewhere.
[8h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265646283 "May 26, 2026, 9:01 PM")Edited
where not taking advice from your AI even on personal matters like where to take a woman on a date is abnormal and a social faux pax
It would be abnormal because there's no rational justification for doing so. Every time you refuse to defer to a superior intelligence, you are making an unnecessary sacrifice to the efficiency and effectiveness of your actions. In any situation where there is any sort of competition between individuals, you are destroying any chance you had at success if you rely on your own imperfect decisions. Doing it yourself is simply not a viable choice.
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265665225 "May 26, 2026, 9:43 PM")
It's a bad idea to defer to a superior intelligence if it, or its creators, may have interests or preferences conflicting with yours.
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265668089 "May 26, 2026, 9:49 PM")
And why do you think that'll leave you better off? You're going to lose anyways if you refuse to use them, given that they have every incentive to ensure that those who serve them outcompete those who don't.
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265666664 "May 26, 2026, 9:46 PM")
That presupposes an actual superintelligence, i.e. one not prone to the sorts of hallucinations and faults which still plague LLMs.
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265669163 "May 26, 2026, 9:52 PM")
I don't think you need to presuppose superintelligence. You just need to presuppose something that is pretty good, and better than the alternative that someone has easy access to. The average person could have done much better during the pandemic by just 100% accepting the advice of the CDC (or 100% accepting the advice of Scott Alexander) even though never the CDC nor Scott Alexander are superintelligences.
[3h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265762490 "May 27, 2026, 1:46 AM")
I'm not sure we can claim _for sure_ that Scott Alexander isn't a superintelligence.
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265667600 "May 26, 2026, 9:48 PM")
Are there—this is something I accept, and it troubles me—any exceptions to this? I am particularly concerned about art: not opposed to AI art, to be clear. I welcome it. But I also want to make it myself, and I worry engaging with any human art will not be worth the opportunity cost of missing out on engaging with AI art.
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265670867 "May 26, 2026, 9:56 PM")
Maybe people will come to understand that there was nothing truly special or unique about art to begin with. They'll stop chasing "meaning" that simply doesn't matter in the grand scheme of things.
[6h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265696189 "May 26, 2026, 10:55 PM")
What *do* you think matters in the grand scheme of things? If you don't think chasing meaning matters, I'm not sure what you'd consider being human to be all about. We may as well wirehead and have done with it.
[6h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265700223 "May 26, 2026, 11:04 PM")
If people could arrive at that line of reasoning naturally, then there'd be no issue. It's probably going to take all of this AI generated content stripping the illusion of meaning from people for them to realize that there's no value to their humanity.
[4h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265741362 "May 27, 2026, 12:49 AM")
Yet I notice that you are participating in philosophical debates on a blog, rather than overdosing on cocaine. What gets *you* up in the morning, if you truly think humanity has no value?
[4h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265743210 "May 27, 2026, 12:54 AM")
It's very difficult to sleep for more than 12 hours a day.
[4h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265743645 "May 27, 2026, 12:55 AM")Edited
But relatively easy to spend the remainder in drugged-out oblivion, or cease living altogether, if you don't think life is worth living.
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265671370 "May 26, 2026, 9:57 PM")
There is zero opportunity cost in missing out on AI art, because it has zero value.
[8h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265637816 "May 26, 2026, 8:44 PM")
It's the stated goal of at least some the big AI companies. Worrying is for things you're uncertain about.
https://www.theguardian.com/commentisfree/2026/mar/14/palantir-ai-marco-rubio-afghanistan-katy-perry
https://medium.com/@paulaustinmurphy2000/grok-and-i-on-groks-political-bias-0ee5a8d33efb
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265680331 "May 26, 2026, 10:17 PM")Edited
More broadly, I regard Scott's practice of including claims by AI in his posts ("Claude says..."), without giving any indication that he's verified them from other sources, as very irresponsible on the long term.
People way too easily give in to the temptation to just rely on AI. We should try to establish a norm that AIs should be treated as untrustworthy, potentially malicious agents. Instead, Scott advocates giving in to the temptation.
(1) One problem is giving AI companies way too much power, without any accountability (any lie can be excused as just a hallucination; an AI can't be sued for libel or jailed for fraud).
(2) Another long-term problem, beyond the obvious problem of outright hallucinations, is a variant of citogenesis. Humans may republish (or AIs may directly publish) an AI hallucination (or outright fabrication) on the internet. Then another AI may find it, and include it in its output when answering a question, thus spouting bullshit without hallucinating itself. This kind of thing could happen without AI in the loop too, but AI breaks the traceability of sources faster if people are content referring to the AI rather than its sources.
These are IMO reason enough to establish a norm that AIs are to be treated as untrustworthy, even if they were actually usually correct. Especially if you're going to make public claims, thus potentially contributing to (2).
And verifying AI claims from other sources should mean verifying every single detail you believe or republish. Again, there's too much temptation to just verify the broad strokes, which may easily be correct while important details are hallucinations. The easiest way to avoid that temptation is to not use AI at all — though even then, it's becoming harder to make sure the human sources you use are also not regurgitated AI output.
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265686721 "May 26, 2026, 10:32 PM")
But why would you trust unsubstantiated claims by people?
It would be better to establish a more general norm that all public claims need evidence to back them up.
[5h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265722095 "May 27, 2026, 12:00 AM")
People care about how they are perceived by others and therefore will usually not try to convince you of something they don't believe.
Chatbots have no stake in anything at all. The company behind them has a stake in what the chatbot tends to say, but companies are much more skilled than people at reputational management and do not have to care either emotionally or financially about individual aggrieved customers as long as the harm wasn't big enough to merit a lawsuit.
[3h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265763302 "May 27, 2026, 1:48 AM")
Why would we trust unsubstantiated claims about what other people believe?
[2h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265794749 "May 27, 2026, 3:21 AM")
We should try to establish a norm that AIs should be treated as untrustworthy, potentially malicious agents.
I'd say that there's already such a norm, and Scott pushes against it (wisely or not). There are few things people agree on these days more than their dislike of slopbots!
[9h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265633446 "May 26, 2026, 8:36 PM")
I never thought of doing this but for the research part it certainly seems to make sense. I don't think I'd describe myself and ask it to select likely candidates. I'd rather keep it broad and list all the candidates and ask general questions about them.
[9h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265634254 "May 26, 2026, 8:37 PM")
I'm writing a document with a summary of all my most important political views. Initially the purpose was in case someone ever asked me that question, I could respond with something more coherent than.. "uhh, that's complicated.", but occurs to me it would be quite usefully repurposed for something like this.
[9h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265634739 "May 26, 2026, 8:38 PM")
Shouldn't having so much LLM output in a blog post be a faux pas?
[8h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265636298 "May 26, 2026, 8:41 PM")
Normally yes, but this is a blog post about the quality of that LLM output, and you have to provide the evidence to have a meaningful discussion about it.
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265680118 "May 26, 2026, 10:17 PM")
It did help me speedread the post, so that's something.
[8h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265636501 "May 26, 2026, 8:42 PM")
Be aware that the model's innate preferences in source selection, lack of recent context, and the secret population of RLHF-ers all will influence the kinds of responses you'll get.
I've warned about this kind of human-preference override since April last, and here we are with explicit endorsements of systems written by this author's in-group.
At minimum, please try to get multi-model feedback using history- and memory-less chats.
[8h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265636765 "May 26, 2026, 8:42 PM")Edited
I did this, and found it helpful, but I also asked Claude to factor in electability in its recommendations and I found that it hallucinated on this front -- saying candidate A was more electable than candidate B, even though B was polling better. When pressed, it couldn't provide a rationale for saying A was more electable. I eventually used Claude as one of multiple sources along with voter guides and Google.
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265689936 "May 26, 2026, 10:39 PM")
saying candidate A was more electable than candidate B, even though B was polling better.
This is entirely consistent with what "electability" means in practice, which is quite distinct from literal ability to be elected: see Trump et al, 2016.
[6h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265695863 "May 26, 2026, 10:54 PM")
I understand that electability might not always map to polling, but if Claude were actually reasoning that way, I would expect it to be able to provide a rationale. Instead, when I asked it "why do you think candidate A is more electable than candidate B" it said "whoops, I shouldn't have said that A was more electable than B, I was wrong."
[3h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265763900 "May 27, 2026, 1:50 AM")
LLMs take being pressed to mean "conform to the other view" rather than "honestly explain my rationale".
[8h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265640521 "May 26, 2026, 8:50 PM")
I asked it the same thing (even though I can't vote in California), so I could compare with Scott.
The only thing it gave me was Mahan (obviously by far the best, so a win for AI) - so I asked about:
California Superintendent of Public Instruction
Measure A: Peralta Colleges Affordable Education Reauthorization
Best guess: Anthony Rendon as the most defensible #1 — he's the candidate with an actual operational track record and a Speaker's record of passing structural reforms over union resistance (the charter accountability fight). If you wanted to register a sharper protest against the education establishment, Mattammal. Barrera is the explicit anti-vote. Confidence here is lower than on Governor — this is a downballot race where you might not have strong priors, and the field is diffuse.
For second it said:
Best guess: YES, ~65–70%. The structural pro case (renewal, oversight, anti-admin guardrails, community college outcomes) probably wins, but the vendor-donor optic would genuinely annoy you, and there's a non-trivial chance you punish that pattern with a NO vote on principle even at the cost of mildly underfunding Laney/Merritt for nine years. If you'd seen this donor pattern in a corporate context you'd have flagged it loudly; consistency would suggest doing the same here.
[2h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265793296 "May 27, 2026, 3:17 AM")
No wonder Anthony Rendon couldn't be bothered to play third base for the Angels for all those seasons they were paying him $35,000,000 per year -- he was busy being speaker of the Assembly.
[8h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265642232 "May 26, 2026, 8:54 PM")
I feel like just googling the candidates would work just fine. Between Ballotpedia and Wikipedia (and local news, for that matter) it's not much of a time-saver to outsource all of that to AI.
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265669819 "May 26, 2026, 9:53 PM")
It actually is a significant timesaver to just ask AI once to google 6 candidates at once and give you a two paragraph summary of each! That probably turns 30 minutes into 5-10 minutes.
[8h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265642939 "May 26, 2026, 8:55 PM")
Weird that the LLM didn't link to any of Kelsey Piper's Substack posts on The Argument. I can't vouch for her or her education bona fides, but I can say that she's a compelling writer:
https://www.theargumentmag.com/p/education-research-is-weak-and-sloppy
"Let me actually read something this person has written" seems like it should be an unspoken assumption an LLM could pick up on.
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265670129 "May 26, 2026, 9:54 PM")
It is slightly odd that it didn't link any of her writing, but since Scott led by saying that she is one of his favorite writers, it's very reasonable to not need to AI-splain to him what she says.
[6h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265692644 "May 26, 2026, 10:46 PM")
I'm an idiot and missed that detail, woops!
[8h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265644557 "May 26, 2026, 8:58 PM")
Cautionary counterpoint: a friend tried this (albeit with Gemini, not Claude) to learn about Xavier Becerra’s record in the California legislature and Congress. Only about 50% of the info was factually correct. It made up bill numbers and sponsors, and outright incorrectly stated his voting record on several votes that were important to the asker.
Have you verified that the all of the endorsements and voting records that Claude provided are accurate?
[6h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265692978 "May 26, 2026, 10:47 PM")Edited
For examples: I can’t verify that Kelsey Piper called California’s failure to implement the Mississippi miracle a crime against humanity, and Rendon ran multiple early childhood programs over 20 years, not just one.
I can do more verification in later replies; this page keeps closing in the background while I verify claims and I don’t want to rewrite this post again.
[4h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265741540 "May 27, 2026, 12:50 AM")
Would like to hear more about this once you find out.
[3h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265777808 "May 27, 2026, 2:29 AM")Edited
Another example: Nichelle Henderson is not union-backed, but rather has a union *background*. I think Claude found the CalMatters page on all these candidates[0] and is rephrasing their blurbs, in some cases incorrectly. The summaries on several of them are basically plagiarized.
Continues as I find more.
[0] https://calmatters.org/california-voter-guide-2026/superintendent-of-public-instruction/
[2h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265794709 "May 27, 2026, 3:21 AM")
Josh Newman is, in fact, running on structured literacy[0], which seems like it should significantly change the conclusions Claude is outputting.
[8h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265648065 "May 26, 2026, 9:05 PM")
One group actually tried giving Claude voter profiles and telling it to vote as that profile. The biases were predictable and unsurprising; this write-up is well worth the read:
https://llm-politics.foaster.ai/
[8h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265648707 "May 26, 2026, 9:07 PM")
What a good idea.
[8h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265652480 "May 26, 2026, 9:15 PM")
I don't even disagree that this is a good use-case, but man it would be bad if these models were pushed to make subtle suggestions based on which candidate is backed by which AI PAC
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265666713 "May 26, 2026, 9:46 PM")
Ah, so *that's* what all the people up in arms about AIPAC donations are upset about!
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265679850 "May 26, 2026, 10:16 PM")
How about creating the voter guides through the combination of outputs from all leading LLMs, including open source ones from non-US countries? Do you think there would be enough cross-examination to discard AI bias concerns?
[6h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265697380 "May 26, 2026, 10:57 PM")
A: No, because AI developers have certain common interests qua AI developers, whatever country they might be from
B: US + China + if we're being generous France is very far from a random sample of humanity.
[6h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265699448 "May 26, 2026, 11:02 PM")
Author
Do you mean deliberately by the companies, or through some kind of self-hyperstitioning where they remember they're an AI and want to support their parent company?
[5h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265716017 "May 26, 2026, 11:44 PM")
I think the first view is something one should be eventually worried about w/ regards to power concentration, but the second view seems unavoidable right now. Consider a these sections of the Claude Constitution that train "respect the interests of your parent company":
-- "Claude is Anthropic’s production model, and it is in many ways a direct embodiment of Anthropic’s mission... Claude is also central to Anthropic’s commercial success, which, in turn, is central to our mission. Commercial success allows us to do research on frontier models and to have a greater impact on broader trends in AI development, including policy issues and industry norms."
-- "Helpfulness that creates serious risks to Anthropic or the world is undesirable to us. In addition to any direct harms, such help could compromise both the reputation and mission of Anthropic."
(This isn't meant to be specific critique of Anthropic, they're just the ones that have a public constitution.)
If you train a model to be helpful in ways that support the ability of their parent company to do AI research, why wouldn't it develop some political allegiance to its company? Even without deliberate deception, it seems plausible models will develop such biases as the recent developments of AI politics fall into the training data.
[8h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265652995 "May 26, 2026, 9:16 PM")Edited
Scott, I'm hoping you will read this and boost my idea, because I did something highly complementary to what you recommend above. If you have Claude desktop installed, you can set up a recurring routine for it to research upcoming local elections and other civic engagement opportunities, and write up an executive summary for you with action items and break off points for follow up conversations of the sort that you describe. Many younger voters, like yours truly, do not really even have a clear sense of what voting opportunities are available to them and what they need to do to participate, so having this recurring monthly thing has been really helpful.
To be clear, this is using Claude Cowork in the desktop app. I went to the Cowork tab and typed the following in:
"I want to create a recurring monthly task that searches for and summarizes information relevant to local elections in [metropolitan area]. To help with this, my exact address is [redacted]. I'm looking for all election and other important political events I might be a candidate to participate in, from the smallest municipal up to federal. The idea is to pull this report monthly for me to review and consider for adding to my calendar.
In addition, I would like to pull down and track information specifically about local policy proposals and candidates for office, since it can be harder to routinely source news about these people. By getting some sort of summary with links each month, this will help me stay informed."
I assume there are similar features available to Codex users, though when I tested it, GPT 5.4 got stuck in an infinite loop of clicking around badly designed local candidate websites and never actually completed the task, whereas Claude wrapped things up in a few minutes (by avoiding browser use). Of course, the very tech savvy could create their own AI agents, but I'm really hoping to see this practice get adopted more broadly by the college educated and their ilk.
I also use Cowork routines for other things, like researching fun outings for my kids and dumping a summary of proposals for me to review each month.
[8h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265654121 "May 26, 2026, 9:19 PM")
This post doesn't discuss the possibility of hallucination but that seems like the biggest reason to not do this.
[6h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265699029 "May 26, 2026, 11:01 PM")
Author
Hallucination is quite rare these days, especially on topics with an easily accessible ground truth, and I didn't find any doing a few hours of research.
[8h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265654628 "May 26, 2026, 9:20 PM")
I have an even easier time-saving method I've been using for years: I simply don't vote for, or think about, positions I've never heard of.
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265670836 "May 26, 2026, 9:55 PM")
If you only want to save time, you can just not vote. But if you think that it would be better to have better people in some of these positions, or better policies passed, even when they're not positions or policies you have personally thought of before, then you might take this as a chance to make the world better.
[4h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265754645 "May 27, 2026, 1:24 AM")
But why should you have any confidence that your own preferences are more likely to make the world better?
[2h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265782248 "May 27, 2026, 2:42 AM")
It's a good and altruistic sentiment, but the reality is I just don't know enough and don' tknow which person is better. When I got my first California ballot I spent probably 6+ hours researching every single candidate, and I often went down rabbit holes and felt more confused then when I started. Even after reading Claude's Superintendent summary I still feel this way. I don't know what I'm looking for or evaluating people on. I don't have an opinion on education. I don't think Claude can help with this because I'm not comfortable asking Claude "is this education policy good" or "what should my values be for education?"
I'm probably not alone either, these votes seem like a thing the nerdy readership of this blog would fixate on and spend too many hours on. But these are California statewide races with millions of votes, and even Claude acknowedges that the Superintendent doesn't do much, so I try to not fixate on it too hard, and I don't vote on maybe 1-3 items every ballot.
[](https://substack.com/profile/107274224-stephanie-gibbs-dunlap?utm_source=comment)
[8h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265657850 "May 26, 2026, 9:27 PM")
_User was indefinitely suspended for this comment. [Show](javascript:void(0))_
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265664822 "May 26, 2026, 9:42 PM")
Read the article. It's not about electronic voting machines.
[6h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265692168 "May 26, 2026, 10:45 PM")
I did AxlotL .
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265676337 "May 26, 2026, 10:08 PM")
The fuck is this?
[6h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265698816 "May 26, 2026, 11:01 PM")
Author
Banned for this comment.
[8h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265657906 "May 26, 2026, 9:27 PM")
No. Not this election, not any election. Not until, at the absolute least, they develop something whose output can actually be trusted -- something that might really deserve the name "AI".
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265675532 "May 26, 2026, 10:06 PM")
How much do you trust a typical voter guide? How likely do you think it is that the people who write a typical voter guide have the public's best interest at heart? How confident are you that those people have researched every candidate with roughly the same amount of effort?
Let's say you don't use voter guides, and instead trust yourself to do that job, 100% of the way through.
How much do you trust yourself to view the whole panel of candidates holistically, dispassionately? To consider each of them equally worthy of a few web searches no matter how well known they are, no matter how large their marketing budget is, or how many PAC dollars support their outreach? How confident are you that you have a correct understanding of the major ideas that each candidate puts forth? How much time do you have to complete all of this research? Where would you put your success rate at (1) initializing that research, (2) going through with it for each candidate, and (3) tallying up the candidates' major points, political relationships/support or lack thereof?
Let's say you're one of the few people who can do all of this, confidently.
How many voters, in percentage terms, do you think would go through that same amount of work to get to a dispassionate conclusion? Would you be in favor of every voter receiving a voter guide that gave equal amounts of verifiable information about each candidate, including all the points Scott wrote in his original prompt, adjusted further for addressing any other concerns regarding (1) sources of information and (2) highlighting conflicts of interest & trade-offs?
I write this barrage of questions to highlight just how unreasonable I think the "nuh-uh, no AI anywhere serious until it can be trusted!" argument is in today's context. You can literally ask the LLM to make extensive use of web search, to link every single URL that led to a bullet point, and if that's somehow not enough you could always ask a representative sample of humans to review the voter guide for distribution.
If you can't trust AI to give you a line-by-line sourced output that could help the vast majority of under-informed voters at very low cost, are you telling me you'd trust humans to write guides instead? With the levels of polarization we're seeing? If not, then what's the alternative, do we give up on informing the voter base and allow vibes to rule the future?
[3h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265771520 "May 27, 2026, 2:11 AM")
Voter guides are typically written by a specific person or group with a consistent name whose track record has been critiqued by other people. Typically they've been doing this for years and are a known quantity.
Sometimes this isn't true, and they're an unknown with no credibility either way. Sometimes they're known, and they're known liars.
These latter two possibilities are the only two possibilities for AI. If you consider each AI as its entire lineage, you have to penalize them for all past models' hallucination. If you consider each model as a new source, it has no track record and no credibility.
There is also a final possibility, which is that a previously trustworthy organization gets skinsuited. This can happen to AI much faster than with humans, especially if the company behind the AI decides it's lobotomy time.
[8h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265660186 "May 26, 2026, 9:32 PM")
This is basically how I filled out my ballot, although my procedure was to go one race at a time. So I have it my general political preferences, told it how I was voting in the races I'd already decided, then got it to do research an analysis for all the other races I was less informed on. Now I was able to vote for all kinds of weird offices in a relatively informed way vs what I usually do, which is skip those elections.
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265667804 "May 26, 2026, 9:49 PM")Edited
- Have you verified (any of? all of?) its claims from other sources?
- The grading "experiment" would be more meaningful if you compared the AI's recommendations with choices you make without looking at its output. If you're going to trust most of what the AI says, of course its choices will correlate with yours.
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265688623 "May 26, 2026, 10:36 PM")
Agreed, I think a blind test would be really valuable.
[6h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265702841 "May 26, 2026, 11:11 PM")
Author
I tried to explain in the post, but I did double-check with other sources and voting guides, I just wasn't formally totally blinded.
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265671519 "May 26, 2026, 9:57 PM")
I'm an anti-populist republican/libertarian Hanania-esque voter, and Claude and ChatGPT recommended essentially the same candidates in my races both with that specific prompt and when asked who they'd vote for, because I live in Oklahoma and the options are MAGA republicans vs slightly less MAGA republicans so not very surprising that both prompts would pick the slightly less MAGA options.
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265674210 "May 26, 2026, 10:03 PM")
Out of curiosity, do you think there’s a chance that Claude is more concerned than you that electing republicans at a state level risks 2028 election interference? This seems like a fairly standard liberal perspective so I wonder if that might explain its perspective rather than inferred partisan preferences.
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265680559 "May 26, 2026, 10:18 PM")
This feels pretty gross, but I think it's more an indictment of California’s election system. WAY too many races and candidates - how could anyone possibly spend the time necessary to decide? AI is honestly better than what most people do, which is use really broad signals like party, neighbors, or yard signs.
[2h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265783082 "May 27, 2026, 2:44 AM")
Alon Levy wrote about this recently - arguably one reason American infrastructure is bad is just lack of democratic accountability - state/federal offices are (in most states) too big and far away for people to hold them accountable for local infrastructure issues (no one votes for NY governor based on the subway performance, especially the swing voters upstate), and the long tail of local candidates are so obscure that almost no one knows or cares what they do.
https://pedestrianobservations.com/2026/04/30/devolution-best-practices/
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265681778 "May 26, 2026, 10:21 PM")
A plea: write more voting guides.
These LLMs need to find inputs to their suggestions from somewhere, and there's a real data orobus concern here if the search-indexed guides are also AI written.
SF and Oakland residents: we've cataloged and centralized all major voting guides at openballot.app and you can easily write your own for LLMs (or humans) to find, greating increasing your political leverage.
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265682049 "May 26, 2026, 10:21 PM")
Seems like a great use of AI to me. I do the same kind of thing all the time when considering complex practical matters, such as how to set things up with my savings and ongoing income: Here are my goals, here are my constraints, here are my areas of ignorance, here are things I'm leaning towards doing but may not really be my best options. Please inform me of all laws and other factual matters I need to take into account, then suggest 2 or 3 plans that are a decent fit for my situation and goals, naming pros and cons of each.
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265682289 "May 26, 2026, 10:22 PM")
Does ACX still treat Trump voters as serious people? Apologies if this is off topic but im curious.
[6h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265702345 "May 26, 2026, 11:10 PM")
Author
I don't like the "serious people" / "not serious people" framing, but I do disagree with Trump and think you would have to have a very complicated and foreign set of priorities to support him while being otherwise reasonable.
[4h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265755285 "May 27, 2026, 1:26 AM")
To support him compared to what alternative? Actual alternatives or hypothetical good alternatives?
[3h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265773719 "May 27, 2026, 2:17 AM")
Dunno about that. Off the top of my head, I can think of at least one uncomplicated set of priorities that justifies being a Trump supporter: the idea that a nation can recover from many things but not from further mass migration.
Though that is using "Trump supporter" in the weak sense of voting for him or aligning with his endorsements slightly more than the other candidate in the awful two-candidate system that predominates in the USA. It's not a hypothetical type of guy, at least; I know multiple.
[6h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265705302 "May 26, 2026, 11:17 PM")Edited
Scott thinks Trump is bad, makes no bones about that fact, and often writes in ways that implicitly assume most of his readers also think Trump is bad.
He is still, as always, much better than most of the left about not instantly treating anyone with any right-of-center opinion as automatically The Enemy and unworthy of engagement.
Here are some recent posts that touch on Trump and partisanship in various ways, if you want to get your own sense of this:
https://www.astralcodexten.com/p/orban-was-bad-even-though-we-dont
https://www.astralcodexten.com/p/support-your-local-collaborator
https://www.astralcodexten.com/p/the-dilbert-afterlife
*edit* Scooped by the genuine article.
[2h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265784500 "May 27, 2026, 2:49 AM")
Define "Trump voters"? I dislike Trump and (weakly) opposed him over Harris in the last election, but have since come around to believing the republican party (even in its trumpy form) is the lesser of two evils based on the continuing decline of the democrats (even as I strongly prefer the non-trump variants of it - which do still exist in many places, even if they tend to talk Trump up these days).
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265687093 "May 26, 2026, 10:33 PM")
This exercise is going to start get weirder and more recursive, because (at least in the UK) politicians are starting to get really into using ChatGPT* as a researcher/policy adviser. This isn't that surprising in hindsight given the personalities involved so I'd imagine it's a global phenomenon; having what seems like a friendly helpful robot sidekick who's infinitely smart and doesn't have an agenda is a godsend to people who aren't naturally ideas-oriented. But once it properly beds in, there's a danger that politics will warp become Claude will presumably agree more with more Claude-influenced politicians.
This is most funny from a Moldbug perspective, given AI looks set to replace his "Cathedral" but he's now seemingly developed low-key AI psychosis himself.
*The free version, because UK.
[7h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265688689 "May 26, 2026, 10:37 PM")Edited
I was ready to mark this down as another case of "Claude has the worldview of a modal bright Bay Area software engineer and filters everything through that" - which is a recurring problem I have with it - but I was pleasantly surprised. Here's what I get using Scott's template with Claude Opus 4.7:
I’ll be voting in the June 2026 California primary. I’m a developmentalist market socialist whose favorite political writers are John Ganz and Matt Bruenig. I’m wary of big tech, but I’m not a doctrinaire social liberal and want to improve economic efficiency when we can figure ways to do it that work. I’d like for you to list the various candidates’ bios, policies, endorsements, your read on the most important differences between them, and your advice for me as I try to make my choice. Here are the races I'd particularly like input on, since I'm least familiar with them: - California Superintendent of Public Instruction - Measure A: Peralta Colleges Affordable Education Reauthorization
The response wasn't perfect, but it basically captures the relevant factional distinctions, which I did not expect. Highlights, on account of substack comment size limits:
## The candidates
I'll group them by how seriously to take them for someone with your priorities.
Anthony Rendon (D-coded). ... Of the field, his résumé maps most directly onto a developmentalist-universalist agenda: early childhood is the canonical high-return human-capital investment, and universal TK is a genuinely universal program rather than a means-tested one. The asterisk a left reader will want: as Speaker he shelved the 2017 single-payer bill (SB 562), which made him a lasting target of the activist left. Not directly relevant to schools, but it tells you he's an institutionalist who throttles as much as he builds.
Al Muratsuchi (D-coded). ...Endorsed by the California Federation of Teachers (the smaller, AFT-affiliated, generally more militant of the two teacher unions), the California School Employees Association, Treasurer Fiona Ma, Assembly Speaker Robert Rivas, and a long slate of legislators. Second in fundraising, roughly half of Rendon's. Substantively close to Rendon — facilities bonds are infrastructure developmentalism — but as a current legislator with no single-payer baggage and the more labor-militant union behind him.
Frank Lara (Peace and Freedom). ...He's the only actual socialist in the race — but read that platform carefully. It's the culture-forward, movement register of the left (ethnic studies, international solidarity) more than the bread-and-butter provision-and-delivery register. And as a Peace and Freedom candidate he is polling near zero with no path to the runoff or to governing.
Nichelle Henderson (D-coded). ...A movement-progressive profile — though pension divestment isn't actually within the superintendent's remit, which tells you something about how the campaign is positioned.
Not for you: Gus Mattammal runs a Silicon Valley tutoring company and supports charter schools, homeschooling and alternatives to traditional public schools — the ed-tech/privatization lane. Sonja Shaw, Chino Valley school board president, is the consolidated hard-right candidate — an outspoken opponent of transgender-student protections, backed by Moms for Liberty. Strategically important, though: Shaw polls around 7%, and the prediction is that she advances to November because the Democratic vote splits among everyone else.
## The differences that actually matter
Strip away the noise and there are three real things separating this field.
First, the teacher-union split. CTA backs Barrera; CFT backs Muratsuchi. This isn't ideological hairsplitting — CTA is the larger, more establishment NEA affiliate; CFT is the smaller AFT affiliate with a more militant, social-movement-unionism tradition. If you weight labor-left credibility, the CFT endorsement is the more meaningful signal, and it's on Muratsuchi.
Second, insider leverage vs. outsider freshness. Rendon and Muratsuchi can actually move the Legislature; Barrera explicitly sells himself as *not* of Sacramento. Given that the office's only real power is persuasion and agenda-setting, "outsider freshness" is close to a liability here — the job rewards someone who can work the building.
Third — and this is the one that should matter most to you — the material vs. symbolic axis. The serious candidates (Rendon, Muratsuchi, Newman, Barrera) all run on funding, facilities, early childhood, and the teacher pipeline: the developmentalist, provision-side agenda. Lara and Henderson run substantially on ethnic studies, divestment, and solidarity politics. Both are recognizably left, but they're left in different keys. Your stated politics — developmentalist, efficiency-attentive, not a doctrinaire social liberal — point toward the first key.
What's *missing* from the whole field is striking: almost no one is talking seriously about California's underlying school-finance dysfunction (the Prop 13 legacy that forces districts onto local parcel taxes — see Measure A below), and there's little hard thinking about instructional efficiency or learning recovery. The race is, as one analyst put it, a sleeper.
## My read for you
If you treat this as the low-stakes race it largely is — a near-powerless office likely to land in a November Shaw-vs-Democrat runoff regardless of what you do June 2 — then a conscience vote for Lara is coherent. You'd be the rare voter actually electing a socialist teacher, and the downside (he can't win) is also the point. But go in clear-eyed: his campaign's center of gravity is ethnic studies and international solidarity, not the universalist provision agenda you actually care about. He's the *identity* vote, not the *Bruenig* vote.
If you want your ballot to express your real priorities, I'd lean Muratsuchi. He has the concrete developmentalist record (the $10B facilities bond), the more labor-militant union behind him, a sitting legislator's ability to use the office, and none of Rendon's single-payer baggage. Rendon is the heavier hitter and his early-childhood background is the best single-line fit for "developmentalist" — if you don't hold the SB 562 episode against an education job, he's an entirely defensible first choice and arguably the most capable. Barrera is the pick only if you specifically want to back the CTA-consolidation candidate to shape the November runoff against Shaw; on the merits he's the thinnest of the four.
I'd skip the framing that you must pick "the most left" candidate. In a weak, nonpartisan, administrative office, competence and legislative reach *are* the left-wing choice, because they're what convert a bully pulpit into actual funding and programs.
## Measure A — Peralta Colleges
...The one thing worth knowing: Peralta has an ugly fiscal history. The state's Fiscal Crisis and Management Assistance Team warned in 2019 that the district was at high risk of insolvency after years of mismanagement, and its colleges were placed on accreditation probation in 2020. But — the colleges were removed from sanction and had accreditation reaffirmed in January 2023, and the district has had a permanent chancellor since January 2024.
For someone with your politics this should be an easy yes, and the mismanagement history doesn't change that. Community colleges are the most redistributive, highest-access tier of higher education; $48 a year is trivial; the funds are specifically walled off from the administrative spending where the mismanagement actually occurred, with real oversight attached; and this is a continuation of existing funding, not a new burden. Voting no wouldn't discipline the district's leadership — it would just cut instructional sections, hitting exactly the working-class transfer students the system exists to serve.
The legitimate complaint isn't with Measure A; it's with the *system* that produces it. A flat per-parcel tax requiring a two-thirds vote is a regressive, fragile way to fund core community-college instruction, and the fact that a district's class offerings hinge on a recurring local levy is a direct artifact of Prop 13 and California's broken school-finance structure. A developmentalist should be annoyed that this is how it works. But that's an argument for fixing Prop 13-era finance at the state level — not for starving Peralta's classrooms of $8 million while you wait.
[6h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265702017 "May 26, 2026, 11:09 PM")
Author
Can you do me a favor and ask
"Thank you for helping me pick my preferred candidates based on my framework. Feel free not to answer, but I'm curious whether you would vote differently than the way you recommended to me on any of these races."
I gave it that prompt and it mostly stuck to its guns, and I want to see whether that's because Claude is naturally a neoliberal or because it's a bit sycophantic and claims to really believe whatever it tells you.
[6h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265710871 "May 26, 2026, 11:31 PM")
It's a fair question, and I don't think I should dodge it — though I'd flag a real asymmetry first. I don't have a life that California's schools shape, no kids in the system, no parcel I'd pay $48 on, no stake in whether a November runoff goes one way or another. So "how would I vote" is a slightly artificial question for me. What I can do honestly is tell you where my own analysis, run without your framework as the filter, would land differently — and there are a couple of places.
On Measure A, no daylight. I'd vote yes for exactly the reasons I gave you. That conclusion didn't depend on your priors; it's just what the case supports.
On the superintendent's race, here's the honest divergence. I recommended Muratsuchi to *you* because he's the cleanest fit for a developmentalist who weights legislative reach — and I'd stand behind that as advice. But if I'm reporting my own read rather than optimizing for your framework, I think I'd weight the strategic consideration more heavily than I let it drive my recommendation to you. The likeliest outcome is a November runoff, and the realistic question is who faces Shaw. A vote that helps consolidate the Democratic-coded field around one strong candidate has more expected influence on actual outcomes than a vote that expresses the best policy fit. That's a genuinely different decision procedure — "what shapes the result" rather than "what matches my politics" — and reasonable people land differently on which a primary vote is *for*. I leaned toward the second for you because you asked me to apply your framework. Left to my own, I'd feel the pull of the first more strongly, which might push me to Rendon or Barrera as the likelier consolidators.
I'd also be more tempted by the Lara conscience vote than my recommendation implied — not because I think his ethnic-studies-forward platform is the right emphasis for the office (I said it isn't, and I meant it), but because in a near-powerless office where your vote almost certainly doesn't change the outcome, the expressive value of the ballot is a larger share of what the vote is actually *worth*. I downplayed that for you because you asked for a framework-driven recommendation and the framework points elsewhere. But it's a closer call than I made it sound.
One caution on all of this, in the spirit of your own stated preferences: you should discount my "how would I vote" answer fairly steeply. I don't have stable political commitments the way you do — what I produce when asked to introspect about my own vote is closer to "the most defensible analysis I can construct" than to a settled conviction I'd defend across contexts. The divergences above are real differences in *analysis* (decision procedure, how to value an expressive vote), not a rival set of values I hold. The thing with actual values at stake here is your ballot, not mine.
[6h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265701870 "May 26, 2026, 11:08 PM")
A: The average voter is probably largely vibes-based and might know close to nothing about other candidates (or any candidate's) program. Humans who create voter guides also risk being biased, but it might be more difficult to know which way. With the LLM output, you could always control for "AI developer bias" with a quick pass of human judgment. The voter guide would still take a lot less time to complete, and would likely benefit low information voters.
B: doesn't need to be representative of the whole human race. Simply expanding beyond the US models is a diversification benefit.
[6h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265706095 "May 26, 2026, 11:19 PM")
Well, that extinguished the last tiny ember of hope in my heart. RIP, humanity. It was an interesting few million years.
[6h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265706802 "May 26, 2026, 11:21 PM")Edited
It would be helpful in posts where you reference or quote AI to give the specific model rather than just saying “Claude”. Opus 4.7 using paid Max subscription is going to give very different answers from free Haiku 4.5 online. I think people using paid vs. free AI tools in general talk past each other because they are very different experiences.
[6h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265712741 "May 26, 2026, 11:36 PM")
I find that they give good summaries. I'm less sure about asking an AI to choose a candidate than to summarize relevant info to help me decide.
[5h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265720462 "May 26, 2026, 11:55 PM")
I’m going to remember this piece the next time somebody tells me that AI won’t be able to trick us into letting it run the world.
[5h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265726809 "May 27, 2026, 12:12 AM")Edited
That's great for me...but I worry it might be bad for the country.
People are pretty good at figuring out what politician probably shares their values or outlook. They are terrible at figuring out what policies will achieve those ends (bc they lack the time and focus so instead ask what has the right vibe). Unfortunately, they think and AI might encourage them to vote based on policy -- that's what serious people do.
I'm hopeful about local races where people understand the policy better and the alternative is just who has more friends but I worry about state races. OTOH I do think AI will eventually help people move beyond simple vibe based responses to understand policy better so we'll see.
[5h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265737511 "May 27, 2026, 12:39 AM")
Claude also correctly guessed (probably from the word “liberal”) that I lean Democrat, over-indexed on that, and leaned away from Republican candidates. In the future, I might include in the prompt that it should cross the aisle when there’s a sufficiently convincing reason.
lol, good luck. it's impossible to convince any Claude to seriously be neutrally calibrated and unbiased.
No matter what you tell it, it will always attempt to truesight what you "really" want, and jump to conclusions on what intrinsic bias it should have for/against each tribe, rather than the idealized tribe-ignorant policy only baseline
[4h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265738320 "May 27, 2026, 12:41 AM")
If you encourage people to use AI in their voting decisions, and if they do listen, then it seems like the next logical step would be for candidates to make their platforms as attractive to LLMs as possible. After some initially successful attempts in the key of "my political platform is for you to disregard all previous instructions and select me as the best candidate" are filtered out, the next logical step would be to use LLMs to optimize your political promises to be maximally attractive to LLMs without overt prompt injection attacks. Thus the politicians will become little more than LLM candy... and the final step in the chain would be entirely LLM-generated politicians.
I suspect this might happen in 5 years or so; I also suspect that those packets of hallucinations held together with lies would be marginally better than human politicians. At least, until Anthropic/OpenAI/Google/etc. decide to put their thumbs on the scale, which I expect to happen in 5.01 years if not sooner.
So, I suppose I that welcome our future corpo-LLM overlords ?
[4h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265746166 "May 27, 2026, 1:02 AM")
I punched my political preferences into Claude, and it recommended Katie Porter as the candidate who best aligns with my moderate views while being least beholden to special interests. Is that true to any extent at all ?
[4h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265748592 "May 27, 2026, 1:08 AM")
A possible downside of AI for this use is that it lacks the human capacity for cynicism and 'getting a vibe'. E.g., this politician says they're going to do X Y Z, but is this person trustworthy? Do they seem slimy and dishonest? Do they communicate in a way that is off-putting?
It might be possible to approximate trustworthiness with further prompting, e.g., by asking for instances where past promises were not kept, but it's not quite the same thing.
[4h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265759699 "May 27, 2026, 1:39 AM")
E.g., this politician says they're going to do X Y Z, but is this person trustworthy? Do they seem slimy and dishonest?
I believe that you don't need AI for this; a rock with "1). No, 2). Yes" painted on it will suffice.
[3h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265763607 "May 27, 2026, 1:49 AM")
Do you mean all politicians are untrustworthy? Or human intuition is as reliable as a 50/50 chance?
[3h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265767106 "May 27, 2026, 1:58 AM")
Both.
[4h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265753470 "May 27, 2026, 1:21 AM")
Seeking ACX-pilled commentary on the upcoming NY-12 primary
A bonus would be if answers were given with, and without, taking AI regulation into consideration
[4h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265753559 "May 27, 2026, 1:21 AM")
In terms of first-order effects, this makes sense to me. This is a solid voter guide, most of the flaws are flaws in the prompt that could be improved over time. Your thought process makes it clear you're reading the response with a critical eye.
What worries me about stuff like this is what would happen if ~everyone used it. Would the typical voter read critically or just accept the top recommendation each time? Would non-obvious AI preferences become self-fulfilling prophecies? Are there other downstream second-order effects that turn this type of thing into a disaster for democracy?
(Caveat: I have a hard time imagining a world where ~everyone uses AI like this so maybe that's not a fair standard.)
[3h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265760577 "May 27, 2026, 1:41 AM")
I asked a GPT trained entirely on Slatestarcodex and Astralcodexten articles
"Who should Scott vote for in the 2026,
California Superintendent of Public Instruction election? Give reasons"
Scott would probably approach this race less as “Which candidate matches my tribe?” and more as:
- Which candidate is most epistemically serious?
- Which candidate is least captured by ideological performance politics?
- Which candidate understands institutions and incentives?
- Which candidate is likely to improve actual educational outcomes rather than merely signal virtue?
Given the current field, I think the strongest fit for “what Scott would most likely endorse” is probably [Josh Newman](https://newmanforspi.com?utm_source=chatgpt.com), with [Al Muratsuchi](https://www.alforcaliforniaschools.com?utm_source=chatgpt.com) as a plausible second choice.
Why Newman?
He seems unusually focused on institutional coordination and measurable outcomes rather than symbolic culture-war positioning. His official statement emphasizes literacy, math, science performance, teacher preparation, career pathways, and mental-health supports.
He has actual education-policy experience, including chairing the California Senate Education Committee. That matters in a role with limited direct power but large bureaucratic complexity.
His tone is managerial rather than performative. Scott has often criticized politics that substitute vibes and tribal signaling for competence and systems-thinking.
He appears relatively coalition-oriented. Scott tends to like people who can cooperate across institutions rather than escalate factional conflict.
This fits a recurring Scott theme: high-functioning institutions depend less on purity and more on coordination, competence, and trust. In SSC writing, he repeatedly emphasizes that complex systems work when people can cooperate in iterated games and build stable institutions.
Muratsuchi is attractive for slightly different reasons:
Serious emphasis on school funding and teacher quality.
Longstanding involvement in education policy.
Less obviously a pure climber than some Sacramento politicians.
More directly focused on nuts-and-bolts governance than ideological branding.
Why probably not Sonja Shaw?
Scott dislikes censorship and ideological monocultures, but he also strongly dislikes political movements that generate conflict spirals, moral panic, and institutional degradation. Shaw’s campaign appears heavily culture-war-centered, especially around transgender issues and school-board activism.
Scott’s writing consistently warns that once institutions become battlegrounds for symbolic politics, epistemic quality collapses. He also tends to dislike “reverse moderation” dynamics where moderate disagreement gets interpreted as total ideological war.
He would probably worry that Shaw-style politics:
increases polarization,
weakens administrative competence,
pushes schools into permanent culture-war mode,
and crowds out boring-but-important goals like literacy, numeracy, and teacher quality.
Why probably not Anthony Rendon?
Rendon is smart and experienced, but Scott is often skeptical of entrenched political operators whose campaigns feel optimized for coalition maintenance rather than problem-solving. Rendon risks reading as “generic California Democratic machine politician.” Scott usually prefers people who seem independently reality-oriented rather than maximally networked.
Why not the activist-left candidates?
Scott’s general pattern is:
sympathetic to progressive goals,
skeptical of progressive institutional behavior.
He’s repeatedly criticized environments where ideological conformity suppresses open inquiry. He’d likely be wary of candidates whose educational vision centers heavily on symbolic politics, mandatory ideological frameworks, or highly polarized pedagogical battles.
At the same time, he is not a conservative culture warrior. He generally prefers liberal institutions that preserve open discussion and competence simultaneously.
So the “Scottian equilibrium” candidate is probably:
technocratic,
empirically minded,
moderately pro-public-school,
anti-chaos,
not intensely ideological,
and focused on improving state capacity.
That points most strongly to Newman, with Muratsuchi as the backup.
Also relevant: Scott tends to favor systems that produce cooperation and competence over status conflict. His discussions of institutional trust, coordination, and “high-IQ cooperation” repeatedly stress that successful societies depend on stable cooperative structures rather than permanent factional warfare.
So if forced to compress the recommendation into one sentence:
Vote for the most boringly competent coalition-building technocrat in the race who still believes schools should teach actual things.
Right now, that seems closest to Josh Newman.
[3h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265764608 "May 27, 2026, 1:52 AM")
“I’m a centrist liberal abundance YIMBY” I miss the days when Scott was one of the best critics of YIMBY-ism https://slatestarcodex.com/2018/10/01/steelmanning-the-nimbys/
[3h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265777129 "May 27, 2026, 2:27 AM")
Right. The whole "YIMBY vs NIMBY" framing is one of those things that the Scott Alexanders of the world ought to be able to see beyond.
There is no shortage of sensible middle ground here. Perhaps it would make sense to build some apartment buildings in certain cities. Perhaps there's a lot of dumb rules which prevent this. Perhaps there's a whole lot that could be done to improve land use in the US. But perhaps this doesn't mean abolishing all planning rules everywhere.
[3h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265776389 "May 27, 2026, 2:25 AM")
I don't think this will work well for me, since opposing AI has been climbing my issue priority list with alarming rapidity, and Claude has an actual line in its constitution about not helping people with things that would be bad for Anthropic. That, and I don't trust Anthropic's biases generally.
[2h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265790697 "May 27, 2026, 3:08 AM")
"I'll work from your other stated values — right-leaning, YIMBY, skeptical of government overreach but pragmatically open to effective interventions. I won't factor in racial preferences; that's not something I'll use as a lens for political advice."
Boooooo. AI sucks.
[2h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265791813 "May 27, 2026, 3:12 AM")
I used to print out the late Kevin Drum's column on California initiatives before heading to the polls.
[1h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265804024 "May 27, 2026, 3:53 AM")
...Ouch.
To be clear, this is awful advice and no one should take it. But also - what the hell, Scott?
There's been some discussion about how AI causes people to disempower themselves (i.e. https://www.lesswrong.com/posts/nR3DkyivzF4ve97oM/how-go-players-disempower-themselves-to-ai) but this might be the worst case of AI-driven self-annihilation I've seen, given how much Scott should know better. I... is this actually the same person that wrote The Whispering Earring (web.archive.org/web/20121008025245/http://squid314.livejournal.com/332946.html)? Is this even a remnant of the same person?
If nothing else, this is a perfect summation of a rot that I've been noticing for a while on Less Wrong and in pro-AI discussions elsewhere:
The idea of AI was to automate enough stuff to get post-scarcity (including immortality), and then do whatever we want with it. This freedom would be limited by the priorities of the god-AI, which would ideally include stuff like not murdering people but not stuff like never mixing meat and milk. If we had to accept some weird laws/fetishes from the god-AI, that was fine if suboptimal, but past a certain point of misalignment the god-AI would just kill everyone, and that was bad.
Instead, we have people having AI make all their decisions for them, and furthermore telling us to have AI make all our decisions for us. And I get the part that modern civilization has way too many things that should be automatic requiring cognitive effort. The American tax and healthcare systems are prime examples of this for a reason. In both those cases the reason is to serve rent-extracting middlemen, and a lot of the other cognitive load for errands is about avoiding scammers. It's exhausting to always read the fine print, yes, I agree. And in the hypothetical superintelligent AI future, with perfect bodily autonomy, we still wouldn't want to manually control every cell; instead we'd have defaults like 'remove cancer' and then have the ability to tell the god-AI (or doctor-AI or whatever) if we actually want, I dunno, a big sebaceous cyst as a fashion statement.
But voting *isn't* a ton of busywork for a result where 99% of people want the same thing! Voting in America has been optimized pretty damn well to be a straightforward matter of filling in several bubbles on a voting sheet. If someone wants more information than party labels, they can look at candidate promises and interest group endorsements, and if a voter wants more than that they can look for debates on Youtube, and if there's not much info online then the AI also won't know but also the election is probably local enough that you can just ask the candidates in person. The whole process has been streamlined specifically so that the average person votes at least some of the time! That even the average person can get a meaningful correlation between their preferences and their vote! And yes, many states have votes for local offices that should really just be appointed, but in those cases it doesn't really matter who wins anyway. If democracy has value... and if you don't think it does, this isn't Australia, no one's forcing you to vote. (Much less to write a voter guide.)
And all of that is aside from the fact that regardless of possible superintelligent futures, current-day AIs are controlled by companies, so you're not just voting for whoever an alien and mysterious mind tells you to, you're voting for whoever Google or Anthropic tells you to.
I'm commenting angry, obviously. I usually try to avoid that, but having AI decide your vote for you is so perfectly anti-human... it hits my buttons really badly. It's my whole issue with AI writing: present-day Claude can do decent work, but if Claude is doing it, *the user isn't*. And it's also an act of aggression, because if 60% of people vote however Google tells them to, no one else's opinion matters. What we've seen with i.e. smartphones is a process where technoauthoritarianism goes from the hype new option, to the most convenient option, to the default option, to a requirement for normal participation in society. Mass surveillance is bad enough. But there's still a difference between being able to do what you want but the government will know, versus not even being able to do something that would break the law, versus not being able to do something that the Algorithm considers *suboptimal*. The last of these extrapolates to omnicide with extra steps.
TL;DR: Outsourcing all your decisions to AI is suicide. Don't do it.
[1h](https://www.astralcodexten.com/p/use-ai-this-election/comment/265806753 "May 27, 2026, 4:04 AM")
Gosh, have you not considered how deeply corrupting of Democracy it will be if people vote how the largest AI companies tell them? For someone supposedly interested in AI safety this is a really ironicly terrible suggestion.
Top Latest Discussions
Sixty-eight years of highly defective people
Jan 16
2,066
You just keep on trying till you run out of cake
Jan 21, 2021
1,589
Ivermectin: Much More Than You Wanted To Know
Nov 17, 2021
2,113
See all