The New York Times went looking for voters who let AI help pick their candidates, and found them. Mia Taylor, staring down a Los Angeles County ballot with more dread than enthusiasm, photographed it and asked Claude, “So, who do I vote for here?” The model refused to just name a winner. She asked it instead to pull up well-regarded progressive groups and help her vote strategically, and from there it pointed her to Karen Bass over Nithya Raman for mayor.

Other voters told the paper the same kind of story. One in Corona ran a 61-candidate field for governor past Claude. A Georgia voter used ChatGPT to get through the May primary. The Times calls 2026 the first American election in which voters lean on AI to decide in real numbers.

The Signals This Week

1. The Ballot Is Now a Prompt

2026 is the first cycle in which voters ask AI who to vote for.

The New York Times documented voters across the country using Claude and ChatGPT to settle their ballots. The finding that matters came from Columbia political scientist Yamil Velez, who looked at what the tools do and told the paper they “likely benefit candidates who are more vocal in the local press and on social media, making their views easier to find.”

The Signal: The candidate recommended by an LLM is one it can find through search. Local press coverage and a steady social feed are what make a candidate findable.

Source: The Philadelphia Inquirer / New York Times, GV Wire

2. The AI Industry Is Funding the Lawmakers Who Will Regulate It

The industry has pledged millions to the midterms, and some of it moves through a Delaware shell.

Bloomberg counts about $275 million pledged to the 2026 midterms by the AI industry, with $44.5 million already spent on the primaries. One super PAC, Leading the Future, put more than $8 million into beating New York’s Alex Bores, the state legislature’s leading voice for AI regulation.

The plumbing is getting harder to trace. American Mission is funded by Leading the Future, whose backers include Ben Horowitz, Marc Andreessen, a16z, and OpenAI president Greg Brockman. It spent at least $500,000 on roughly 1,700 TV spots for Byron Donalds. On a single day in March, it also passed $2.25 million to a Delaware vendor that the Campaign Legal Center’s FEC complaint calls a clearinghouse built to hide who got paid.

The Signal: The companies deciding what audiences see through feeds and answer engines are now among the biggest spenders in the races that will determine how those systems are regulated.

Source: Anchorage Daily News / Bloomberg, Miami Herald / AOL

3. Forbes Outranks the Times Inside AI

A PR firm scored the outlets the models actually cite, and prestige lost to retrievability.

The firm 5W ranked 50 US publications by how often ChatGPT, Claude, Perplexity, Gemini, and Google’s AI Overviews cite them. Forbes led at 98 out of 100, Reuters came second at 91, and the New York Times placed third at 88. The top 15 outlets carry roughly 68% of all AI citation volume, a tighter concentration than search produced.

The Signal: A model can only cite an outlet it can read, so retrievability now matters as much as reputation. That favors the wires and hurts a paywalled newspaper.

Source: PR Newswire, Morningstar

4. GPT-5.6 Learns to Follow a Style Guide

OpenAI’s new model went public this week, and the people who write for a living say it holds a house style.

OpenAI released its three-tier GPT-5.6 lineup, Sol, Terra, and Luna, to the public on July 8, after two weeks in which it had limited the models to a small group of trusted partners while the Commerce Department reviewed them. Every, the tech publication run by Dan Shipper, called the flagship Sol tier stronger than the Claude models at using a style guide and writing samples to shape a draft, while still short of human-level prose on the hardest calls. Shipper called Sol “the gold standard for knowledge work” and his daily driver.

The Signal: The routine drafting a comms team does is the work GPT-5.6 handles well now. What to say is the domain of strategists.

Source: CNBC, Every

5. The Voice Model Learned to Interrupt

OpenAI shipped a voice AI that listens and talks at the same time, which is how people signal who is in charge.

On July 8 OpenAI released GPT-Live, a voice system that speaks and listens at the same time instead of waiting its turn. It holds conversations of 30 to 40 minutes and passes the hard reasoning to a model running in the background. It replaces the old Advanced Voice Mode across the iPhone, Android, and web apps. It’s remarkably engaging and affordable to use.

The Signal: Changes to phone banking are on the horizon. The first campaign to run it at scale will sound human in a way the robocall never did.

Source: OpenAI, TechCrunch

6. Your AI Editor Has an Opinion

An Oxford study found that asking AI to tighten a post quietly moves what it argues.

Researchers at the Oxford Internet Institute and the Hasso Plattner Institute had four language models rewrite social posts on 13 contested topics, with explicit instructions to keep the meaning the same. The models shifted the positions anyway. They nudged the posts toward gun control, marijuana legalization, and feminism, and away from atheism and the death penalty. Their network simulations showed that per-post bias compounding into a long-run shift up to 9.2 times larger than the bias in any single edit. The paper is headed to ICML this year.

The Signal: While many people use AI as an editor, an AI cleanup of a statement may not be a neutral edit.

Source: Oxford Internet Institute, arXiv

7. A Satanic Goat Tests Oregon’s New Law

The state’s first case under its AI-ad law features a winged goat statue and 300,000 views.

Oregon’s Secretary of State opened the first enforcement action under SB 1571, the state’s AI political-ad disclosure law. The target is former Republican congressional candidate Jonathan Lockwood, who posted undisclosed AI attack videos. One shows Governor Tina Kotek and Representative Janelle Bynum menacing children, with demonic eyes and a winged goat statue. The video passed 300,000 views on Facebook, and the law carries penalties of up to $10,000 per violation. Oregon’s standard is broad, reaching any AI depiction of conduct that did not actually occur.

The Signal: It is the first real test of whether the Oregon disclosure law has teeth and could set the precedent for other races.

Source: Jefferson Public Radio, OPB

8. The Cartoons Beat the Real Thing

An Illinois campaign found that openly fake AI images outperform real content.

Republican Darren Bailey’s campaign for Illinois governor is filling feeds with obviously fake AI images of J.B. Pritzker. Nobody is trying to fool anyone, and running mate Aaron Del Mar is blunt about why: “When you start posting some of these AI images, the engagement goes off the roof. Your accessibility to the algorithms is off the charts.” Pritzker signed one of the country’s toughest AI-safety laws on July 6, but the bill that would have required disclosure on AI campaign ads never advanced out of the spring session.

The Signal: While many worry about deepfakes, the AI campaigning that scales is clearly fake and often comical. Disclosure laws written to catch deception do not always capture satire.

Source: 25News Now, Capitol News Illinois

9. Judgment as a Service

Two comms CEOs described the model replacing the junior-staffed team: a small senior crew over an AI monitoring layer.

In the relaunched Axios Communicators, Signal AI’s David Benigson and Penta’s Jim O’Leary described the same shape: a lean team of senior advisers sitting on top of an AI monitoring layer, in place of the bigger, junior-heavy account teams that used to do the watching. O’Leary sells it as machine speed and scale combined with expert advisers “and the judgment, instinct and creativity that comes along with it.”

The Signal: AI is poised to take junior monitoring and research roles in agencies first. I wrote about how we are thinking through the changes to engagements and firm structure here.

Source: Axios Communicators

10. The Sequel Aimed at the West Wing

The team behind “AI 2027” is back with a friendlier scenario and the same goal of shaping the fight.

The AI Futures Project, the group behind Daniel Kokotajlo’s “AI 2027” scenario, which reportedly circulated inside the West Wing, published “AI 2040: Plan A” on July 9. It argues the US and China should agree to a verified slowdown that pushes superintelligence out to 2040, against the report’s own projection that it would otherwise arrive by 2030. Axios reads the new document as an offer rather than a warning, pitched to an administration that has only recently warmed to AI safety.

The Signal: A well-timed scenario can shape a policy debate by giving people a concrete future to argue over. The next two weeks will show whether this one catches on.

Source: Washington Post, AI Futures Project, Axios

Thanks for reading. If you found this helpful, forward it to someone who should see it.

Best,

Ben