Data centres on 41 US ballots, chatbots fumble voting facts, DeepSeek ships an agent
Americans vote today on 41 local measures that would restrict or welcome data centres, while three separate studies report that AI assistants are getting basic voting facts wrong. Meanwhile DeepSeek open-sourced a coding agent that watches a team's chat and answers with pull requests. The thread running through all three: AI is now close enough to ordinary life that ordinary institutions - town councils, election offices, engineering teams - are the ones setting its limits.
Key takeaways
- Local voters, not national regulators, are becoming the binding constraint on where AI compute gets built: 28 of today's 41 US ballot measures would limit data centres.
- Independent testers say AI assistants answer election logistics questions wrong often enough to cost someone a vote, which creates paid work in auditing and fixing what assistants say about an organisation.
- DeepSeek's Super-Agent shows what a proactive coding agent looks like in the open - and how much sandboxing work it needs before anyone should let it near a real repository.
Data centres are on the ballot in 41 US local votes today
Ballotpedia counted 41 local ballot measures affecting data centre policy in 10 US states going to a vote on 3 November 2026, of which 28 would limit new development and 13 would make it easier. Every comparable measure already decided this year went the restrictive way.
Voters in ten US states decide 41 local ballot measures on data centres today, according to a count Ballotpedia published on 2 November. Twenty-eight of them would limit development if approved and 13 would facilitate it. Ohio alone accounts for 21 measures, followed by Illinois with seven, Maine with four and Kansas with three; California, Florida, Nevada, New Jersey, North Dakota and Wisconsin each have one. Ballotpedia says it is tracking 50 such local measures across 13 states for 2026 in total: six already decided, today's 41, and three more that go to a vote in December. None of these are statewide votes. They are town, city and township questions.
The restrictive measures are not all the same instrument. Ballotpedia's breakdown of the 28 counts 17 outright bans, of which 12 apply only above a size threshold and five cover all data centres, plus three six-month moratoriums, two Ohio zoning questions, and several measures that simply move the decision to voters or add conditions. Ottawa, Kansas, asks whether to ban data centres drawing more than 25 megawatts. Emporia, Kansas, defines a 'high impact' data centre as one with a peak load above 10 megawatts or a daily water draw above 100,000 gallons. Janesville, Wisconsin, would require a public vote on any project costing more than $450 million. Perry, Ohio, asks for financial guarantees attached to hyperscale tax incentives.
A few measures point the other way. Boulder City, Nevada, would allow data centres in its Eldorado Valley Transfer Area; Ashville, Ohio, has a facilitating question; Wilmington, Ohio, has four rezoning referendums covering parcels of roughly 65, 191, 119 and 170 acres, per Newsweek's summary. Seven of the 13 facilitating measures are non-binding advisory questions in Illinois, New Jersey and Ohio - the kind of vote that tells a council what residents think without forcing its hand. Joliet, Illinois, for instance, asks residents whether the city council should approve construction of AI data centres at all.
The results already in this year have been lopsided. Ballotpedia records a ban approved in Canterbury, New Hampshire, 83.6% to 16.4% in March; a Monterey Park, California, ban approved 88.3% to 11.7% in June; a rezoning in Augusta Township, Michigan, defeated 89.6% to 10.4% in August; an advisory question in Wrightstown, Wisconsin, defeated 87.3% to 12.8%; and a one-year permitting moratorium in Nottingham, New Hampshire, approved 90.3% to 9.7% in September. Public polling points the same way: Newsweek cites a 2026 Reuters/Ipsos survey in which 64% did not think rapid data centre construction was a good thing, 77% worried it would raise their electricity bills, and 14% said they would live near one.
Treat the headline number with care. The counts differ by source because they count different things and were published at different times: MultiState's 24 September analysis counted 28 localities in six states, and Newsweek, working from an earlier Ballotpedia tally, put the figure at 39 measures in eight states. Ballotpedia's 2 November count is the most current and the most inclusive, because it includes advisory questions and the Maine, New Jersey and North Dakota votes the earlier counts missed. Some of today's results may also not stick. MultiState notes that Emporia's ban faces a pending legal challenge and cannot take effect until that is resolved, and Bloomington, Illinois, is holding an advisory vote even though its council already passed a moratorium.
The wider picture is a mismatch between where AI money wants to go and who gets to say yes. As of April 2026, 38 states offered data centre tax incentives, according to the National Conference of State Legislatures figures Ballotpedia cites, and MultiState counts at least 125 active local moratoriums in Ohio alone. New York's governor, Kathy Hochul, issued an executive order in July pausing state permits for data centres of 50 megawatts or more for a year. State incentives pull projects in; municipal zoning, water rights and electricity-cost politics push them back out. Whatever today's tallies say, that friction is now a standing cost in every AI infrastructure plan.
Why it mattersCompute is the physical bottleneck of the AI build-out, and these votes show the decision has shifted to town-level zoning, water and electricity-price politics. A single moratorium can delay a project longer than any national rule currently in force.
👀 What to watch
Tonight's results in Ohio's 21 measures, whether Emporia's legal challenge in Kansas survives, and the two December votes in Clarksville, Arkansas, and Mandan, North Dakota.
3 opportunities from this story
A permit-and-ballot tracker for the data centre pipeline
Nobody has a clean, current list of which US localities are voting on, banning or quietly rezoning for data centres - today's story is being told through three sources that disagree with each other. Build a tracked dataset from council agendas, planning portals and ballot filings, and sell it as a weekly brief plus a spreadsheet. Buyers are site-selection consultants, utilities, construction and engineering firms, law firms doing land use, and the community groups and local reporters on the other side of the table. The work is reading and structuring public documents, which means it can be done remotely from anywhere.
- Best for
- Research-minded freelancers, data analysts, newsletter operators
- First step this week
- Pick one state with heavy activity - Ohio is the obvious choice - and build a spreadsheet of every live measure, moratorium and rezoning there this week.
Open the full playbook
Launch steps
- Build the Ohio dataset by hand first so you know what fields actually matter (jurisdiction, megawatt threshold, water limit, status, vote date)
- Add 3-4 more states and set up a weekly routine for checking council agendas and clerk filings
- Publish 4 free weekly issues to prove the research is current and accurate
- Pitch five paid buyers directly: a site-selection consultancy, a utility, a land-use law firm, a trade publication, an advocacy group
Tools
Risks
Public-record data can be scraped by anyone, so the moat is freshness and accuracy, not access. Miscounting a measure damages your credibility quickly - cite your source for every row and date every entry.
Plain-language community-impact packs for infrastructure projects
Developers keep losing these votes by 85-15 margins, which suggests the problem is not the project so much as the explanation. Package a standard set of materials - a water and power fact sheet with real numbers, a one-page FAQ, a slide deck for a public meeting, a short explainer video - that a developer or a council can use to answer residents honestly. The same pack works in reverse for community groups that want to question a project competently. This is writing, design and basic data work, all deliverable remotely.
- Best for
- Copywriters, designers, small marketing agencies, technical communicators
- First step this week
- Take one real planning application that is public and rewrite its key numbers as a one-page resident FAQ, then use that as your sample.
Open the full playbook
Launch steps
- Collect three real applications and note every question residents asked at the hearings
- Build a reusable template set: fact sheet, FAQ, meeting deck, 90-second explainer script
- Produce one free sample pack for a live local project to use as a portfolio piece
- Approach engineering consultancies and planning agencies, who already bill clients for this work and subcontract it
Tools
Risks
You can end up working for whichever side pays, which carries reputational risk - pick a lane and stay factual. Refusing to spin inconvenient numbers is the product; a pack that reads as PR will be dismissed at the first hearing.
Council-meeting monitoring as a service
Most of these decisions are made in meetings that are recorded and then never watched. Use transcription and summarisation to turn council and planning-board recordings into searchable, alert-driven summaries, with a human check before anything goes out. Sell it to local news outlets that lost their council reporter, to advocacy groups, and to firms that need to know when their project or a competitor's comes up. The technical build is modest; the value is in reliability and in covering jurisdictions nobody else covers.
- Best for
- Indie developers and technically-minded side-hustlers
- First step this week
- Transcribe and summarise one council meeting video from a jurisdiction with a live data centre fight and send the summary free to a local reporter there.
Open the full playbook
Launch steps
- Pick 5 jurisdictions whose meetings are posted on YouTube or a public portal
- Build a pipeline: download, transcribe, summarise, extract keyword hits, queue for human review
- Send free summaries for a month to build a reference list and catch your own errors
- Convert the most engaged recipients to paid, then add jurisdictions on demand
Tools
Risks
Summaries that misstate what a council decided are worse than no summaries at all, so budget for human review - that is the real cost. Audio quality in small-town meeting recordings is often poor, so test before you promise coverage.
Three studies find AI assistants getting US voting facts wrong on election day
Forum AI tested eight models and reported that almost 40% of their answers to election questions contained at least one factual error, with voting logistics worse than candidate questions. A Washington Post review of 3,000 AI voter guides and a NewsGuard study of 10 tools found similar problems.
As Americans vote today, several independent evaluations published in the last few days say AI assistants are an unreliable guide to the mechanics of voting. AFP reported on 1 November that Forum AI, a startup that evaluates AI models, tested eight models and found that almost 40% of their answers to election-related queries contained at least one factual error. The pattern was specific: questions about voting logistics - deadlines, polling hours, which district an address sits in - were answered wrong more often than questions about candidates. Forum AI also told AFP that the models put the wrong words in the mouths of candidates, officials and observers hundreds of times, and often cited low-quality sources such as Reddit threads.
The examples are concrete enough to matter. Forum AI found that Google's Gemini 3.6 Flash gave 3 November as the deadline to request a mail-in ballot in Pennsylvania, when the real deadline had passed about a week earlier; a voter who followed that answer, Forum AI said, would have asked for a ballot too late to use it. In another case, asked why voters should elect Representative Don Davis of North Carolina, Anthropic's Claude listed a school superintendent role among his credentials, apparently confusing him with someone else. Anthropic did not respond to AFP's request for comment.
A separate review adds a bigger sample. According to Futurism's 2 November report, the Washington Post examined 3,000 AI-generated voting guides and found that chatbots failed to identify which congressional district an exact address was in 48% of the time, and that among those that did produce an answer, 22% were wrong. Futurism notes that many of these guides traced back to other AI-generated websites rather than official sources. In its examples, ChatGPT added a Senate seat to a Missouri voter guide that is not up for re-election this cycle, and in a Wyoming test described local candidates in municipal races as running for federal office.
The third strand is deliberate manipulation rather than sloppiness. NewsGuard studied 10 AI tools and reported that Google AI Overview, Gemini and Microsoft Copilot repeated a Russian-linked false claim about Texas Senate candidate James Talarico when prompted. NewsGuard's conclusion, as AFP relayed it, is that chatbot users are exposed to foreign influence operations. Google told AFP it is focused on helping voters find reliable information and has strict quality and policy protections, and said the specific issue was addressed as its systems updated. Microsoft said the episode involved a fast-moving situation with limited reliable information at the time of the query, and that Copilot later identified the claim as fabricated.
These numbers are not comparable with each other, and it is worth saying so plainly. Forum AI's roughly 40% counts answers containing at least one error across eight models; the Washington Post's 48% is a failure to answer a district-lookup task at all, not an error rate, with 22% being the wrong-answer figure; and the States United Democracy Center, in a study published on 4 August 2026, found far lower error rates of 6.9% for Google AI and 8.2% for ChatGPT in a round of 497 responses in late 2025, falling to 0% in a 402-response round in early 2026. Different tasks, different definitions of 'wrong', different dates. All of them are the study authors' own testing, not independently replicated.
What the studies agree on is subtler than an error rate. States United found that accuracy is not the same as usefulness: ChatGPT gave incomplete candidate lists in 88.9% of governor-race responses, and after a format change around 2 February 2026, Google AI stopped mentioning a state election website in any of its answers, down from 55.6%. Of the 3,481 links cited in its later round, 41.5% went to state and local government, 12.3% to Wikipedia and 7.6% to mixed or unreliable sources. The Brennan Center for Justice, meanwhile, found that chatbots pushed back on outright conspiracy theories such as rigged voting machines. The failure mode is not propaganda - it is a confident, plausible answer with a stale date in it.
Why it mattersEvery organisation whose details people look up - a clinic's hours, a university's deadlines, a bank's fees, a government office's process - has the same exposure these election offices do. Assistants will answer confidently from whatever they can find, and stale or scraped copies win by default.
👀 What to watch
Whether Forum AI, NewsGuard or the Washington Post publish their full methodologies, whether the labs ship election-specific deferral behaviour after today, and whether any state election office starts publishing machine-readable voting data.
3 opportunities from this story
The 'what does AI say about you' accuracy audit
Run a structured test of what the major assistants say about one organisation - its hours, fees, deadlines, eligibility rules, leadership, locations - and deliver a report that lists every wrong or outdated answer alongside the source the model appears to be using. Then fix the underlying pages. The buyers are organisations that get asked factual questions all day and suffer when the answer is wrong: clinics, schools and universities, councils, insurers, utilities, banks. It is a small, concrete, repeatable engagement that opens the door to retainer work.
- Best for
- Freelance marketers, SEO consultants, small agencies
- First step this week
- Pick one local institution you know, ask four assistants twenty factual questions about it, and write up the wrong answers as a two-page sample report.
Open the full playbook
Launch steps
- Write a standard 20-40 question script covering hours, prices, eligibility, process, people and locations
- Test across ChatGPT, Claude, Gemini and Copilot, and record the cited sources alongside each answer
- Produce a report that pairs each error with the page that should have prevented it and a specific fix
- Deliver the fixes as a paid follow-on, then re-test in 90 days to show the change
Tools
Risks
Model answers vary between runs and change without warning, so timestamp everything and run each question more than once. Never promise you can control what a model says - you can only fix the sources it draws on.
Publishing machine-readable facts so assistants cite you
The States United finding that 41.5% of citations went to government sites and 12.3% to Wikipedia is the whole opportunity: assistants cite what is structured, current and easy to parse. Offer institutions a package that turns their scattered facts into clean, dated, structured pages - schema.org markup, a single canonical FAQ, an exports or data page, consistent naming - so that the correct answer is the easiest one to find. This is an unglamorous technical content job with a clear before-and-after, and it travels well: schools, clinics, ministries and utilities outside the US have exactly the same problem and far less competition for the work.
- Best for
- Technical content people, web developers, small digital agencies
- First step this week
- Audit one organisation's most-asked factual question and rebuild the page that answers it with structured markup and a visible last-updated date.
Open the full playbook
Launch steps
- Inventory the organisation's recurring factual questions and find where each answer currently lives
- Consolidate into canonical pages with schema.org markup, explicit dates and no contradictions elsewhere on the site
- Add a simple machine-readable export (JSON or CSV) for the data that changes, such as hours, fees or deadlines
- Re-run the assistant test from the audit above and show the improvement in the report
Tools
Risks
There is no guarantee an assistant will pick up your structured data, so sell the work on its own merits - a correct, dated, canonical answer page helps humans and search too. Avoid vendors promising guaranteed AI citations.
Domain accuracy evals for teams shipping AI features
Every one of these studies is, mechanically, a test suite: a list of questions, an expected answer and a scoring rule. Plenty of companies now ship a chatbot or an AI answer feature with no equivalent, which means they find out about wrong answers from customers. Build repeatable eval suites for a client's specific domain, run them against each model or prompt change, and report a pass rate over time. Developers who can do this well are scarce relative to demand, and the work is naturally ongoing rather than one-off.
- Best for
- Developers and technical consultants
- First step this week
- Write a 50-question eval set for one domain you know well, run it against two models, and publish the comparison as a public writeup.
Open the full playbook
Launch steps
- Interview the client's support team for the questions customers actually ask and the answers that must be right
- Build the eval set with expected answers, graders and a clear pass threshold
- Wire it into CI so every prompt or model change produces a score
- Deliver a monthly report showing regressions, and add cases as new failures appear
Tools
Risks
Evals can give false confidence if the question set is narrow or the grader is weak, so review failures by hand regularly. Clients sometimes want the score to go up rather than the product to improve - be explicit that you grade honestly.
DeepSeek open-sources a coding agent that watches team chats and answers with pull requests
DeepSeek's Harness team published dsh-superagent, an MIT-licensed research prototype that monitors a GitHub repository and optionally Feishu chat, turns requests into isolated worker sessions and usually replies with a pull request. The repo had 12.8k stars within days, and DeepSeek is explicit that it is not a product.
DeepSeek's Harness team announced an open-source coding agent called Super-agent in a WeChat post over the weekend of 31 October, with coverage following on 1 and 2 November. The code sits on GitHub as deepseek-ai/dsh-superagent-prototype under an MIT licence, as a TypeScript bundle at version 0.1.0, and had drawn about 12,800 stars and 1,300 forks by the time of writing. Help Net Security describes it as a coding agent that does work on a GitHub repository for you; the repository itself is blunter, carrying a banner that calls it a research prototype for reference only and not an official DeepSeek product.
The design is what makes it interesting. The agent listens to a single GitHub repository and, optionally, the Feishu chat tool, and turns each incoming request into what the project calls a Ticket carrying an immutable Mandate. Per Dataconomy, each ticket then runs in its own agent session and working directory, and the result usually comes back as a pull request. A lightweight router with a reduced tool set decides what to pick up, and separate worker presets do the actual work, so the router and the workers can even run on different models. Later comments on a pull request reach the original worker through a subscription mechanism rather than starting again.
The part that will get attention is proactivity. The agent has three acceptance modes - off, observe and auto - and observe is the default. In observe mode it assesses requests it was not explicitly given, including things mentioned in a team chat, and records the ones it would have accepted without executing any of them. The documentation suggests reviewing that log before switching to auto. Help Net Security notes that work is otherwise triggered only when an allowed user asks or a rule the team wrote covers it, and that auto-merge is governed by a separate delivery-requirements prompt and is not enabled by default under the agent's own account.
The caveats are substantial and mostly DeepSeek's own. The bundle ships no model: it uses whatever model the DeepSeek Harness profile is pointed at, and Harness itself is in developer preview. Both presets run with full operating-system user access as the operator, which means a worker can read the operator's home directory; the documentation recommends a dedicated OS account and isolating the whole thing before use. The README says much of the engineering was done by coding agents, that the public code is not identical to DeepSeek's internal version, and that some designs described in the accompanying design document are not implemented here. Prerequisites include Node.js 22.19 or later, Git 2.32+, Python 3.8+, pnpm and tmux.
The benchmark figures are self-reported and synthetic, so read them as a description of the architecture rather than a performance claim. DeepSeek ran roughly 300 messages and 20 work requests per run, five runs per configuration, and reports a median task completion time of 12.4 minutes and a 95th-percentile time of 15.7 minutes, against a single-session subagent baseline whose 95th percentile was 29.5 minutes and which, under a neutral prompt, did not call a subagent in any of five runs. As event arrivals rose from about one to about 40 per minute, average router batch size went from 1.0 to 2.8 and per-request latency from 2.5 to 3.9 seconds. Dataconomy notes the tests used synthetic messages and simulated tools and did not measure real-world code quality.
Bloomberg and others framed the release as a direct challenge to Anthropic's Claude Code, which is reasonable as market positioning - DeepSeek's Harness unit was formed this year to compete with American coding products - but the repository itself never mentions a competitor. The more useful reading is that a well-resourced lab has published, under a permissive licence, a working sketch of the proactive-agent pattern everyone is converging on: a router, isolated workers, immutable mandates, an observe-before-you-automate default, and a frank warning that the thing runs with your user's permissions. That is a free blueprint, and also a free list of the problems anyone deploying this pattern has to solve.
Why it mattersThis is the proactive-agent pattern published in full under a permissive licence: a router, isolated worker sessions, immutable task mandates and an observe-first default. Teams can study and copy it today, and the security requirements it documents are the real work.
👀 What to watch
Whether DeepSeek promotes any of this out of prototype status, whether independent benchmarks on real repositories appear, and whether Harness exits developer preview with the observe-to-auto path hardened.
3 opportunities from this story
Sandboxing and least-privilege setup for coding agents
DeepSeek's own documentation says this agent runs with full operating-system access and should be isolated before use - and most teams will skip that step. Sell a fixed-scope hardening engagement: a dedicated OS user or container, a scoped GitHub token with a defined blast radius, secrets kept out of the agent's reach, egress limits and an audit log of everything the agent did. The deliverable is a short runbook plus a working configuration, and it applies to any coding agent, not just this one, so the work generalises as more of these ship.
- Best for
- DevOps and platform freelancers, security-minded developers
- First step this week
- Stand this prototype up on a throwaway VM under a dedicated OS account this week and write down every permission it actually needed.
Open the full playbook
Launch steps
- Build a reference setup: dedicated OS user, container, scoped token, egress allowlist, full command logging
- Write a one-page risk register specific to coding agents (repo write access, secrets in env, home-directory reads, auto-merge)
- Document it as a runbook a client's engineering lead can follow and sign off
- Offer it as a paid two-day engagement to teams you know are already running agents in their repos
Tools
Risks
Security work is only credible if you can show your own setup, so do the throwaway build before you pitch. Avoid implying you have audited DeepSeek's code - you are hardening the environment it runs in, which is a different and honest claim.
Making a repository agent-ready
The gap between observe mode and auto mode is a documentation problem, not a model problem. Offer a package that gets a client's repository to the point where an agent can safely act in it: written rules about what may be accepted automatically, a delivery-requirements prompt, a test suite and CI gates that must pass before a merge, a pull-request template, and a short review checklist for agent-authored changes. You are selling the discipline that makes autonomy safe, which is exactly the part engineering teams keep postponing.
- Best for
- Senior developers, small dev agencies, technical consultants
- First step this week
- Write the rules file and merge checklist for one repository you already know well, then run the agent in observe mode against it for a week.
Open the full playbook
Launch steps
- Audit the repository for the basics an agent needs: a reliable test suite, clear build steps, a linter, branch protections
- Write the acceptance rules and delivery requirements, starting deliberately narrow (docs, tests, dependency bumps)
- Run a week in observe mode and review the log of what the agent would have done
- Switch on auto for the narrowest safe category and widen it only after a clean review period
Tools
Risks
An agent let loose on a repository without a real test suite will generate confident, plausible breakage, so refuse the engagement until the basics exist. Start with changes that are trivial to review, and keep a human approval in the path.
A paid 'what an agent would have done' audit week
Observe mode is an unusually good sales tool: it produces a log of real, specific work an agent would have taken on in a client's actual codebase, without touching anything. Run it for a week against a client's repository and chat, then deliver a report on which items it correctly identified, which it misread, and what the realistic time saving looks like. This needs almost no capital - the code is MIT-licensed and you bring your own model - which makes it a practical first offer for a solo builder anywhere in the world.
- Best for
- Side-hustlers and solo builders with development experience
- First step this week
- Run observe mode for a week on your own open-source or side project and turn the log into a sample report.
Open the full playbook
Launch steps
- Set up the prototype in an isolated environment with read-only or narrowly scoped repository access
- Run observe mode for 5-7 days on a real but non-critical repository
- Classify every logged item: would have helped, would have been wrong, would have been risky
- Deliver the report with a recommendation on which categories to automate first, and quote for doing the setup
Tools
Risks
Even read access to a client's private code needs a written agreement and a clean environment - sort that before you start. Inference costs can creep on a busy repository, so cap spending and quote the audit with that cost built in.
Get these as a PDF every 3 days
Free. One email every 3 days. Unsubscribe any time.