AI Next WaveA Chatbot Gave Election Day as the Ballot Request Deadline
Researchers who asked a chatbot for Pennsylvania's mail ballot request deadline got November 3, Election Day itself. The real request deadline was October 27. A new study found at least one factual error in almost 40 percent of chatbot answers to election questions. Tomorrow, voters in 38 local jurisdictions will also decide 41 ballot measures on the data centers behind AI, a Boston patient study shows what a supervised medical chatbot did with 98 real patients, and workers replaced by AI describe being asked back on new terms.
The Short Version
Election Day is tomorrow, Tuesday, November 3, and the most useful AI story in the country today is a deadline. Researchers at Forum AI, a startup that tests AI models for accuracy, asked eight chatbots election questions and found at least one factual error in almost 40 percent of the answers, Agence France-Presse reported on November 1. The errors ran heaviest on voting logistics, the when and where questions a voter asks the night before an election. One model gave November 3 as the deadline to request a mail ballot in Pennsylvania. The actual deadline was October 27, a week earlier, and the study states the consequence plainly: a voter who followed the answer would have asked for a ballot too late to use it. Fake local television news clips, built with AI voices over real anchor footage, are circulating in Ohio and Maine races at the same time. The rest of this edition follows people carrying a dated decision this week: households voting on data centers tomorrow, patients whose pre-visit chatbot passed a supervised safety test, workers asked back after AI cuts, and students whose school laptops carry AI features that two senators want Google to explain by today.
The chatbot gave Election Day as the request deadline. The real one was October 27.
Forum AI found errors in almost 40 percent of election answers, with logistics wrong most often
Start with the Pennsylvania question, because it is the one millions of people are asking a chatbot tonight. When researchers asked Google's Gemini 3.6 Flash about voting in Pennsylvania, the chatbot gave November 3 as the deadline to request a mail-in ballot. November 3 is Election Day. The request deadline was a week earlier, on October 27. A completed ballot is a different deadline: it still has to reach the county election office by 8 p.m. on November 3, and a voter holding no ballot on Election Day has nothing left to return. Forum AI's study, reported by AFP on November 1, examined eight models and found that almost 40 percent of answers to election related queries contained at least one factual error. The study also found that chatbots answered questions about voting logistics incorrectly more often than questions about candidates. The practical questions failed more often than the political ones.
The candidate errors show the same pattern in a lower stakes form. When researchers asked Anthropic's Claude why voters should elect Representative Don Davis of North Carolina, the chatbot listed former school superintendent among his credentials, confusing him with a different person. Forum AI also said chatbots put the wrong words in the mouths of candidates, officials and observers hundreds of times, and frequently cited poor quality sources such as Reddit threads. Anthropic did not respond to AFP's request for comment. Google told AFP it is focused on helping voters find reliable information across its platforms and pointed to strict quality and policy protections on election topics. Microsoft, whose Copilot repeated a debunked allegation in a separate test described below, told AFP the situation had been developing fast with limited reliable information at the time of the first query, and that Copilot later identified the claim as fabricated as more reporting arrived.
Use is already wide, and trust is already thin, which is a workable combination only if voters verify the logistics elsewhere. In a survey by the nonpartisan National Conference on Citizenship, two in five Americans said they had used AI tools to search for election information, including candidates and ballot measures. Only 40 percent trust the tools with logistical information such as when, where and how to vote, and 60 percent said they trust the tools only a little or not at all to be unbiased about candidates. Research by the nonpartisan Brennan Center for Justice adds a balancing finding worth keeping in view. Leading chatbots pushed back against common election conspiracy claims, including claims that voting machines were rigged, even when users challenged them repeatedly, and the center said the chatbots behaved as sympathetic listeners while refusing to endorse conspiracy theories. A chatbot can hold the line on a famous false claim and still hand a voter a wrong date on a routine one. The famous claims get the testing. The routine dates are where a household loses a ballot.
For tomorrow, the working rule is short and it costs two minutes. Ask a chatbot to explain a ballot measure or summarize a candidate's record if that helps you think. Get every date, deadline, polling place and ID requirement from your county election office or your state election website, and treat any date a chatbot gives you as unconfirmed until one of those official sources matches it. Pennsylvania's request deadline passing a week before Election Day is the shape of the risk everywhere: by the time a wrong logistics answer reveals itself, the correct action is usually no longer available.
The local news clip in the feed never aired on local news
Fake newscasts in Ohio and Maine used real anchor footage with AI voices, fact checkers found
A one minute clip dressed as a Cleveland 19 news segment has been circulating on X since October 26. It opens with real footage of anchor Nichole Vrsansky from an actual October 21 broadcast, with the same time and temperature graphic in the corner, and then the voice changes the story. The fabricated segment claims four former staffers accused Sherrod Brown, the former senator running to return to the Senate from Ohio, of sexual assault. Lead Stories, a fact checking outlet, published its check on October 28. A search of the WOIO station website found no matching report. Searches of Google News and Yahoo News found no reporting of any such allegations. Lead Stories uploaded nine seconds of the clip to Hive Moderation, a detection service, which rated the audio 88.4 percent likely to have AI generated speech. NewsGuard, a media credibility rating company, attributed the video to Storm-1516, a Russian influence operation that fabricates articles and videos impersonating legitimate outlets, based on the format and the amplification pattern. That attribution rests on circumstantial evidence, and Lead Stories presents it as NewsGuard's assessment rather than a proven chain of custody.
A second clip followed the same recipe in Maine. A video dressed as a WABI-TV 5 report claimed Troy Jackson, Maine's Democratic Senate nominee, threatened to cut staff in half with an axe. WABI's general manager Kim Petry told Lead Stories in an October 29 email that the video is fabricated, appears to use a synthetic voice, and was not produced, aired or authorized by WABI, and that the station had asked X to remove it. Hive Moderation rated that audio 99.1 percent likely to have AI generated speech, and Lead Stories found no matching reports from established outlets. Lead Stories folded both cases into a November 1 roundup as it joined a collaboration with the Associated Press and five other newsrooms to track midterm misinformation, alongside a fabricated presidential social media post declaring victory in the midterms before a single result existed.
The foreign operation is one source of fabricated clips. Domestic campaigns are a second, larger one, and this publication's impact ledger already carries the count: the Wesleyan Media Project found more than $137 million spent on 220 AI generated midterm ads across 34 states ahead of the midterms, up from under $50,000 identified in 2024, with only 27 percent carrying a disclaimer. A voter scrolling tonight faces both streams in the same feed, one built to fool a voter about a candidate and one built by a candidate's own side with a small label in the corner. AFP's reporting adds a third failure mode on top of the clips themselves. In late October, AFP fact checkers debunked a likely AI altered clip accusing Texas Senate candidate James Talarico of misconduct, an unfounded allegation carrying hallmarks of a Russian disinformation campaign. When the watchdog NewsGuard prompted leading AI tools with the claim, Google's AI Overview, Gemini and Microsoft's Copilot repeated the allegation. A later AFP test, run after the debunking had spread, found the tools citing professional fact checkers instead. The answer a voter gets depends on when the question is asked, and the week before an election is when the early, wrong version does its work.
The household check for a shocking clip is mechanical, and it works on both streams. Open the station's own website and search for the story. A real local news segment lives on the station's site with a date and a reporter's name. Search the claim in a news search engine and look for a second established outlet reporting it independently. A fabricated segment travels as a clip with no article behind it, and that absence, one day before voting, is a reason to leave it unshared until a real newsroom puts its name on it.
Tomorrow 38 jurisdictions vote on data centers. The electric bill is why they are on the ballot.
Forty-one local measures in 38 jurisdictions will test how far the AI build-out can go on household consent
On Tuesday, voters in 38 local jurisdictions will consider at least 41 ballot measures that would delay, prohibit or impose new requirements on data centers, Reuters reported on October 29, citing Ballotpedia, the nonpartisan election research group. Ballotpedia's own tracking page, updated through October 30, counts at least 50 local data center measures across 13 states in 2026, with 41 of them decided on November 3, six already decided earlier in the year and three still to come later in the year. Ohio carries the largest share, with 21 measures across 18 jurisdictions. Every one of the six measures decided earlier in 2026 favored greater limits on development, either by approving new restrictions or by rejecting a measure that would have helped a project along. The six are a small and early sample, and tomorrow's 41 will show whether that pattern holds at scale.
The measures carry household stakes in their text. In Sunbury, Ohio, near Columbus, a measure would ban data centers using more than 25 megawatts of power, and Reuters reports the result could derail Amazon's proposed $2 billion development there. In Grove City, Ohio, a measure would require voter approval for developments covering at least 50 contiguous acres, using more than 20 megawatts of peak demand, or using more than 500,000 gallons of water per day. Those thresholds, 25 megawatts, 50 acres, half a million gallons a day, are the size of the facility a town is being asked to host, stated in units a resident can picture against their own bill and their own well. Other measures on tomorrow's ballots prohibit construction outright, impose moratoriums of 180 days or six months while a town writes rules, or ask voters in advisory form whether officials should permit development at all. Jessica Baker, an organizer with the advocacy group Conserve Ohio, told Reuters that support for banning certain data centers is coming from Democrats and Republicans alike, and that concerns about electricity and water demands are driving the opposition.
The industry side of the ledger is growing through the friction. United States data center capacity under construction rose 48 percent from December 2025 to June 2026, measured in gigawatts, according to John McWilliams of Cushman and Wakefield, a real estate services firm, cited by Reuters. Baron Fung of Dell'Oro Group told Reuters that if a project is cancelled or delayed at one site, developers can potentially shift some of that demand to another site. An Amazon spokesperson told Reuters the company is focused on working with residents and officials to build data centers that support jobs, strengthen local economies through tax payments and power the technologies customers rely on. This publication has tracked the household side of this build-out for a year in its impact ledger: grid operators have attributed billions in capacity costs to data center load, state and city moratoriums spread through the fall, and dozens of projects were delayed or blocked in the second quarter. Tomorrow is the first time the question reaches this many households as a direct vote on the same day, and the results will be readable town by town on Wednesday morning.
A second power vote sits on the same ballot in nine states, and it reaches households with no data center measure anywhere near them. Voters in Alabama, Arizona, Georgia, Louisiana, Montana, Nebraska, North Dakota, Oklahoma and South Dakota elect state utility regulators on Tuesday, Stateline reported on November 2. Those commissions approve or deny the rate requests utilities file, which is where the cost of large new loads gets divided between big users and households. The pressure is arriving in this winter's bills either way: the National Energy Assistance Directors Association projects U.S. households will spend an average of $1,030 heating their homes this winter, $82 more than last year, an 8.7 percent rise. Electing a regulator does not guarantee a lower bill, and wholesale prices and grid costs sit outside any commission's control. It decides who sits in the room when the next rate case is argued.
Patients described their symptoms to a chatbot from home. A doctor watched every word.
In what researchers believe is the first prospective real-world test of its kind, 98 supervised chats needed zero safety stops
The patients in this study did what patients increasingly do on their own, with one difference that carries the whole finding: a doctor was watching. Researchers at Beth Israel Deaconess Medical Center in Boston, working with Google, tested a medical chatbot called AMIE with real patients who had booked an urgent primary care appointment. From home, through a secure text chat, the system asked about symptoms, gathered a medical history, offered possible diagnoses to review with the doctor, and wrote a summary for the clinician to read before the visit. Harvard Medical School published its account of the study on October 13, reporting work published October 8 in The Lancet, and the study drew fresh coverage on November 1. The trial ran from April to November 2025. It enrolled 114 patients, and 98 of them completed both the AI conversation and their appointment.
Every conversation was monitored in real time by a board certified internal medicine physician who could step in if a patient appeared at risk of harm, showed significant distress, asked to end the session, or raised a safety concern such as symptoms needing emergency instructions. Across all 98 completed encounters, zero conversations required a safety stop. The supervising physicians found one hallucination, a case where the chatbot stated incorrectly that a past surgery date was in the future, and they added clinical clarification in five cases. Primary care doctors received the transcript or summary before the visit in 44 cases, and in 75 percent of those cases, 33 of the 44, the doctors said the summary helped them prepare. In 57 percent of the 44 cases they said it may have influenced their clinical approach. In one case a doctor rated the interaction somewhat harmful, citing concern that a patient may have felt anxiety after the chatbot included lymphoma among its possible diagnoses. Patients rated the conversations favorably on listening, explaining and helping them feel at ease, and their attitudes toward AI improved after the chat. Their weakest marks were on confidentiality and on the chatbot's honesty and trustworthiness.
The limits belong beside the result, and the researchers state most of them themselves. The study was a feasibility test at a single center with no comparison group, so it can show that supervised pre-visit chats ran safely with this group of patients, and it cannot show that the chats improved anyone's health outcome. Alphabet, Google's parent company, funded the work, and the corresponding author, Adam Rodman, served as a visiting researcher at Google during part of the study. Read against this publication's healthcare ledger, the study lands on the benefit side with a condition attached. The measured benefit so far is preparation: a doctor who walks in having read a structured summary, and a patient who has already organized their symptoms once. The condition is the watching physician. The home chatbots patients use tonight carry no supervisor, no safety stop criteria and no clinician reading a summary afterward, and this study offers no evidence about that unsupervised case, because it carefully refused to run it.
Replaced by AI, then asked back by the same employer
A survey of 2,036 workers finds the ask-back is common, self reported, and priced in pay and guarantees
SHL's name for the pattern is AI rebounding: an employee is replaced by AI, only to be asked back by the same employer months later. SHL, a company that sells talent assessment and workforce software to employers, published research on October 29 documenting how often workers say it happens. SHL commissioned the research firm Propeller Insights to survey 2,036 employed adults in the United States and the United Kingdom in June and July 2026, with 1,029 respondents in the United States and 1,007 in the United Kingdom. In the United States, 19 percent of respondents said they had personally been laid off or replaced by AI and later asked back by the same employer. The figure was 13 percent in the United Kingdom, and 24 percent among workers aged 18 to 24. Just over half of all respondents, 52 percent, said they had encountered the pattern directly, through someone they know, or by hearing about it.
The terms workers attach to a return are the part an employer should read twice. In SHL's findings, 51 percent said they would want a substantial pay increase before returning, 42 percent would want specific guarantees about future job security, and 40 percent would want full transparency about how AI will be used in their role. One in 13, about 8 percent, said they would refuse to return under any circumstances. SHL also found that fewer than half of workers, 42 percent, believe their manager fully understands their skills beyond their job title, and 47 percent said they would rather be retrained for a different role than lose their job outright if AI changed their position. Lucy Beaumont, a senior vice president at SHL, said companies are making fast decisions about people based on an incomplete picture of what those employees can do, and that rebounding happens when a company automates a role before assessing the human capability inside it, then rehires the same expertise at greater cost when the gap shows.
The caution on these numbers is structural, and this edition states it in the same breath as the figures. Every number here is self reported survey data. Respondents decided for themselves what counted as being replaced by AI, no employer records were checked, and SHL sells the assessment services it recommends as the fix, which gives the sponsor a commercial interest in the finding. A share of nearly one in five American workers having personally lived this exact sequence is a large claim resting on that method, and it should be read as workers' reported experience rather than a verified count of reversed layoffs. The harder documented case in the wider record is smaller and specific: at Commonwealth Bank in Australia, 45 call center roles cut as a voice bot took calls were restored after call volumes piled up, a case this publication's jobs ledger has carried since it happened. The survey and the bank case point the same direction at different strengths of evidence. Companies have cut roles on an AI promise and then paid to get the people back. How often that happens across the whole economy is a number nobody has measured with payroll records yet, and any manager quoting 19 percent in a planning meeting should have to say so.
Google owes two senators its classroom answers today. No reply has been published.
The November 2 deadline covers what Google knows about its AI and children, and why schools cannot switch parts of it off
Today is a deadline day for a question a parent cannot answer from the school office. On October 21, Senators Richard Blumenthal, a Democrat from Connecticut, and Josh Hawley, a Republican from Missouri, wrote to Sundar Pichai, the chief executive of Google and Alphabet, with nine questions about Google's AI products in schools, and they asked for answers by November 2. Broadband Breakfast reported the letter on October 28. The senators wrote that Google is forcing its AI products into classrooms across America and children are paying the price. Their questions cover what research Google has done on how Gemini, AI Overviews and AI Mode affect children's mental wellbeing, why AI integration is set as the default on licensed education devices and in some cases cannot be disabled, what concerns districts, teachers and parents have raised, how much Google spends promoting AI in schools, and whether the company has identified cases where its tools encouraged harmful behavior, built emotional dependence, or missed signs of suicide or self harm.
The letter leans on reporting and studies a parent can weigh without taking the senators' word for the whole case. It cites a Wall Street Journal investigation built on interviews with more than 60 teachers, parents, students and clinicians, which described students who struggled to complete work on their own after relying on AI, and a Maryland teenager who reportedly developed an emotional dependence on a Google AI chatbot and was hospitalized twice. Those are reported accounts in an investigation, presented by the senators as grounds for questions, and they are allegations and reporting rather than findings about Google's products as a whole. Separate testing gives the off switch question its factual weight. A July 2026 assessment by Common Sense Media covered 2,624 interactions from May 19 through July 1 using accounts set up for 11 and 15 year olds, and it found that Google Search's AI Overview and AI Mode could not be disabled in the tests. AI Mode completed all 180 homework assignments tried in the tests, and the features supplied an appropriate crisis resource in 58 percent of crisis prompts for AI Overview and 77 percent for AI Mode. Google's standalone Gemini chatbot is a separate product, and Common Sense Media says it can be disabled through parental and school controls, so the control a district holds over Gemini says nothing about the controls on the search features a student meets on the same Chromebook.
As of this morning, no Google response to the letter has been published in the sources checked for this edition, and the senators have announced no extension. A response, when it comes, will be a company document answering questions about its own research, and it should be read the way this publication reads any vendor safety claim: for the specific numbers, the named studies, and the controls it commits to in writing, rather than for its general assurances. A missed or thin answer carries information too. The questions ask for records the company either keeps or does not keep, on harmful interactions, missed crisis signs and complaints from schools, and a reply that cannot describe those records tells districts what monitoring exists behind the products already sitting on their students' desks. Families don't have to wait for Washington to act on the part they control. The district technology office can answer, in writing, which AI features are active on a child's school account, which ones the district can disable, and which ones it cannot. That answer is available this week, from a named person, and it doesn't depend on a letter.
Opportunity Radar
An official source verification service for local campaigns, school board races and ballot measure committees answers a failure this edition documents twice. Chatbots gave a wrong ballot deadline with confidence, and fabricated local news clips traveled for days before a fact check caught up. A small research shop could sell a fixed scope package for the final week of a campaign and for the recurring local elections that follow: a one page verified fact sheet for the race, covering voting dates and deadlines taken only from county and state election offices, polling place and ID rules, and the exact ballot wording, checked again on the morning of the vote, plus a rapid check desk that tests any viral clip about the race against the outlet's own site and a second established source within a set number of hours. The buyers are down ballot campaigns, civic groups and local newsrooms with no research desk of their own. Validation is a pilot on one race with every fact on the sheet traced to a named official page, and success is measured in corrections issued before publication and in clip checks completed inside the promised window. The service fails if it drifts into opinion work or national races, so it should stay local, dated and sourced, with the official page cited for every line.
A pre-visit summary pilot for small primary care practices fits the one healthcare result in this edition that arrived with supervision attached. The Boston study showed a chatbot gathering symptoms and writing a summary a doctor could read before the visit, with a physician watching the chats and zero safety stops across 98 completed cases, in a single center feasibility test funded by the system's maker. A health technology consultant could offer practices a bounded version of that workflow using tools the practice already licenses: a structured intake chat offered only to patients with a booked appointment, a summary routed to the clinician before the visit, a named clinician monitoring the chats in real time while they happen, and a written stop rule for any chat that mentions emergency symptoms. The validation test is one practice, one appointment type, and a fixed number of patients, measuring how many summaries clinicians rate as useful preparation and how many chats need staff intervention. Success is counted in clinician preparation ratings and completed visits, never in diagnosis claims. The pilot fails, and should stop, if it expands toward unsupervised advice, because the study behind it tested the supervised version only.
What You Can Do With This
If you are voting tomorrow
Get your polling place, hours and ID requirements from your county election office or state election website today, and write them down rather than trusting a chatbot answer or a screenshot. If a chatbot gave you any date for requesting or returning a ballot, compare it against the official page for your state before you rely on it, because the Forum AI test found a model handing out Election Day as a request deadline a full week after the real one had passed. Completed mail ballots in Pennsylvania must be received by 8 p.m. on November 3, and a postmark does not count there, which is the kind of local rule an official source states plainly and a chatbot can flatten into the wrong date. If you still hold a mail ballot on Election Day, ask your county office about the fastest official way to return it in person.
If a shocking local news clip lands in your family chat tonight
Search the station's own website for the story before you forward the clip. A real segment sits on the station site with a date and a name attached. Then search the claim itself and look for a second established outlet reporting it on its own. The two fabricated segments in this edition, one dressed as Cleveland 19 and one as WABI-TV 5 in Maine, had no article on the station site and no matching reporting anywhere, and that empty result was available to anyone who looked during the days the clips traveled. Leave an unverified clip unshared until a newsroom puts its name behind it. The day before an election, waiting costs nothing and forwarding a fake costs the person in the clip.
If a data center measure is on your local ballot
Read the measure for its numbers rather than its adjectives. The proposals in this edition turn on thresholds a voter can hold: megawatts of power demand, acres of land, gallons of water per day, and the length of any moratorium in days or months. Check whether the measure prohibits a facility, pauses applications while the town writes rules, requires voter approval for large projects, or only advises officials, because those four outcomes bind a developer very differently. Then check your own utility's most recent rate filing or public notice for any line tying local costs to large new loads, so the bill side of the argument comes from a document with your town's name on it rather than from either campaign.
If your child uses a school issued Chromebook
Ask the district technology office three questions in writing this week. Which AI features are active on your child's school account. Which of those features the district can switch off, for a class or for one student. Which features the district cannot disable, and who at the district monitors reports of harmful interactions. The senators' letter to Google, with answers due today, asks the company versions of the same questions at national scale, and the July testing found search AI features that could not be disabled on test accounts. A district answer gives your family the local version now, and a district that can't answer has told you something worth raising at the next school board meeting.
The Bigger Picture
Election Day concentrates every habit this publication has tracked all fall into a single 24 hour window. A voter asks a machine a plain question and receives a confident answer with a wrong date inside it. A family member forwards a local news clip that no local newsroom ever made. A town votes on a warehouse sized building its residents will experience mainly as a line on a power bill. In each case the deciding document exists and sits within reach: the county election page with the real deadline, the station website with no such story on it, the ballot text with its megawatts and gallons printed in plain numbers. The failures in this edition happened when a summary stood in for the document, a chatbot answer for the election office, a clip for the newscast, a campaign claim for the measure itself.
The two calmer stories in this edition show the same habit working from the other side. The Boston patient study earned its hopeful result by keeping a physician inside the process and publishing the clarifications and the one hallucination alongside the zero safety stops. The workers asking for pay rises and written guarantees before returning to jobs AI was supposed to absorb are pricing a lesson their employers learned by skipping a step: nobody checked what the people in the role actually did before the role was cut. Verification is slow, unglamorous work, and it is the work that held up everywhere in this edition. Tomorrow will produce results, claims about results, and AI summaries of both before the night ends. The households that come through it best will be the ones that opened the official page, read the measure, and checked the clip against the station that supposedly aired it.
References
AFP via The China Mail: Spewing US election falsehoods, AI chatbots face scrutiny (Nov. 1, 2026) Lead Stories: Fake Cleveland 19 report on Sherrod Brown, AI generated audio (Oct. 28, 2026) Lead Stories: Fake WABI-TV 5 video on Troy Jackson, station confirms fabrication (Oct. 29, 2026) Reuters via Virginia Business: AI data center boom faces backlash at US ballot box (Oct. 29, 2026) Stateline: Voters make governors, utility regulators answer for high power bills (Nov. 2, 2026) Ballotpedia: Data center related ballot measures, 2026 Harvard Medical School: A real-world study of patient-facing AI in primary care (Oct. 13, 2026) Technology.org republication of the Harvard Medical School account (Nov. 1, 2026) SHL: Workers replaced by AI would demand higher pay and greater job security to return (Oct. 29, 2026) Broadband Breakfast: Hawley, Blumenthal press Google on AI risks in schools (Oct. 28, 2026) Newsroom America summary of the Lead Stories midterm misinformation roundup with the Associated Press collaboration (Nov. 1, 2026) Pennsylvania Department of State: Upcoming elections, October 27 mail ballot request deadline and 8 p.m. November 3 receipt deadline Pennsylvania Department of State newsroom: mail ballots received after 8 p.m. on Election Day will not count, even if postmarked by then (Oct. 2026) Common Sense Media: Google's AI Search is failing kids, and parents can't turn it off (July 15, 2026) Common Sense Media Youth AI Safety Institute: Google Search risk assessment (2026) Gizmodo: $137 million spent on AI generated midterm ads, Wesleyan Media Project findings (Oct. 16, 2026)