> daily_signal(2026_06_25)

There is a video of a Senate candidate saying things he never said, running as a real campaign ad, and nobody broke a law.

PickBits Daily Signal · Thursday, June 25, 2026

By Mark Pickering · 9 min read · June 25, 2026

// tl;dr

A national party committee ran an AI-generated deepfake of a Texas Senate candidate, and there is no federal rule against it. The NRSC posted a synthetic video of James Talarico reading his own old posts back on camera, words he never said aloud, with only a faint "fake" watermark. The FEC that is supposed to write a rule is deadlocked; what exists instead is a patchwork of 28 inconsistent state laws, and the one federal action, an AI-voice ban in robocalls, does not reach a video ad.
A federal appeals court heard, for the first time, whether training an AI on copyrighted work is fair use. On June 11, the Third Circuit took up Thomson Reuters v. Ross Intelligence; the lower court found Ross infringed more than 2,000 Westlaw headnotes to build a competing legal tool, and the panel pressed Ross on the direct-competitor fact. A ruling could stay narrow (an older, non-generative tool) or move the floor under every model trained on scraped data.
The Authors Guild ran 10 human-written articles through five AI detectors, and the results indict the tools. Pangram flagged 0% AI across all 10; ZeroGPT swung from near zero to 76%, tagging a Joan Didion obituary at 66% and a Louise Erdrich Pulitzer note at 76%. Same human words, wildly different verdicts, and these tools increasingly gate students, writers, and job applicants.
An AI triage tool cut the wait from a suspicious mammogram to a biopsy from over two months to under ten days. UCSF and UC Berkeley's Mirai reads screening mammograms and flags the highest-risk ones for same-day workup; in a run of 4,100+ scans, it pulled about one in eight for immediate attention. It triages; it does not diagnose, and a radiologist still reads every case.

We have tracked the AI-training-copyright fight since CNN sued Perplexity last month, and today it reached an appeals court for the first time: in Thomson Reuters v. Ross, the panel sounded skeptical that training on work nobody paid for is fair use. That escalation, from district-court suits up to the appellate level, is the day's clearest signal that the demand for rules is getting stronger. So is the rest of it, with 28 statehouses writing their own deepfake laws and admissions offices reaching for AI detectors. What is getting weaker is the means to enforce any of it.

The FEC that should govern synthetic campaign ads is deadlocked, so a party committee ran a fake ad of Texas candidate James Talarico as a real ad and broke no federal law. The detectors meant to police "AI writing" flagged a Joan Didion essay as 66 percent machine. The one counter is the hospital, where an AI, aimed narrowly and kept under a human radiologist's supervision, cut a biopsy wait from two months to ten days. The push to govern AI is intensifying while the machinery to enforce it lags, and the technology behaves best exactly where a person stays in the loop.

Today: deepfake attack ads are running in the midterms with no federal referee, a federal appeals court heard the first fair-use fight over AI training, the tools that police "AI writing" flagged a Joan Didion essay as 66% machine, and an AI cut the wait from a suspicious mammogram to biopsy from over two months to under ten days.

1. A deepfake of a Senate candidate ran as a real campaign ad, and no federal rule says it can't.

The synthetic-political-ad era arrived with no referee on the field.

The National Republican Senatorial Committee posted a video of Texas Senate candidate James Talarico sitting at a desk, in his own face, reading his own old posts out loud and commenting on them. He never read them on camera. He never gave that commentary. The whole clip was AI-generated to make him look extreme, and it carried nothing but a faint watermark admitting it was fake. It ran as a real campaign ad.

Here is the part that should bother you, whatever your politics: there is no federal rule against it. The Federal Election Commission, which is supposed to write one, is deadlocked and has not done so.

What exists instead is a patchwork of 28 state laws, most untested and inconsistent with one another. The single federal action on synthetic media, the FCC's ban on AI voices in robocalls, does not even reach a video ad. So between now and November, the deepfake ad is not a glitch in the system. It is the system, until someone builds a referee.

Why this matters: This is the first national election conducted with a convincing generative video and no agreed-upon rules for it, and the absence of a federal standard does not prompt pause. It creates a free-for-all, decided 28 different ways. The candidate in the crosshairs today has a press team to push back; most targets of a synthetic clip will not. Action this week: Do the boring thing that actually protects you. When a political video makes you feel something instantly, whether rage or disgust or certainty, check the outlet before you check your blood pressure: who published it, and is the claim carried by a named newsroom or only by the clip itself. If you run comms or a campaign, there is no federal safe harbor, so find out which of the 28 state laws applies to you and write your disclosure rule down before you need it, not after.

cnn.com: James Talarico and the AI deepfake Republicans are running in the midterms (2026)
usnews.com: AI deepfakes blur reality in 2026 US midterm campaigns (March 28, 2026)
stackcyber.com: The patchwork of state AI-deepfake election laws (2026)

2. A federal appeals court heard, for the first time, whether training AI on work you never paid for is fair use.

The question every AI company is quietly praying about finally reached an appellate bench.

On June 11, the Third Circuit became the first U.S. federal appeals court to take up whether copying copyrighted material to train an AI qualifies as fair use, in Thomson Reuters v. Ross Intelligence. A startup, Ross, trained a legal-research tool on Westlaw's copyrighted headnotes to compete with Westlaw; a Delaware district court had already found that it infringed more than 2,000 of them and rejected the fair-use defense outright. At the appeal, the judges did not sound convinced by the AI side. They kept circling the same fact: that Ross built a direct competitor from the other company's editorial work.

Ross is an older, non-generative tool, not a chatbot, and the court could cabin its holding to structured legal content. But it is the first appellate crack at the central question, and whichever way it lands, it moves the floor under every model trained on scraped data.

Why this matters: For two years, "we trained on it" has been a shrug. This is the first appellate signal that it might become a liability instead, and the direction of travel, narrow or broad, sets leverage for every content licensing deal that follows. Action this week: If you train or fine-tune on scraped data, watch whether the holding is confined to non-generative use or written broadly, because that single distinction decides how exposed your pipeline is. If you license content, a pro-Thomson Reuters ruling strengthens "pay-to-train" leverage, so track it before your next data deal and send this to whoever signs off on your training sources.

techjacksolutions.com: Third Circuit hears oral arguments in Thomson Reuters v. Ross (June 2026)
courtlistener.com: Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc. (docket)
reedsmith.com: The court, AI, and fair use — Thomson Reuters v. Ross (2026)

3. The tools schools and bosses use to catch "AI writing" just got caught calling humans machines.

The same human paragraph scored 0% and 76% depending on which detector you used.

The Authors Guild ran 10 articles written by real, named humans through five AI detectors. The good ones held up: Pangram flagged 0% AI across all 10, and Originality.ai was near zero. But ZeroGPT swung wildly, anywhere from 5% to 76%, and it tagged a Joan Didion obituary as 66% AI and a note on a Louise Erdrich Pulitzer at 76%. Same human words, verdicts ranging from "clearly human" to "clearly machine" depending on the tool.

The Guild's warning is blunt: professionally written human prose reads as "AI" to these systems, and they are increasingly the gatekeepers for students, writers, and submissions. A detector that flags a Didion essay as two-thirds machine is not a neutral check; it is a coin flip with consequences attached.

Why this matters: These tools are quietly becoming the infrastructure for grading, hiring, and editorial screening, while their accuracy on real human writing is little more than guesswork. The cost of a false "AI" flag does not fall on the vendor; it falls on the student or applicant who has no way to contest it. Action this week: If your work might get run through a detector, treat a false flag as a documented failure mode and keep your drafts and version history as provenance, since a revision trail beats a detector score. If you run admissions, hiring, or editorial, do not gate on a detector score alone; the same paragraph just scored 0% and 76%, so require human review and disclose the tool and threshold you use.

authorsguild.org: Can AI detectors be trusted? (2026)
the-decoder.com: Authors Guild test finds some AI detectors perfectly identify human writing while others fail on every text (2026)

4. An AI looked at a mammogram and cut the wait for a cancer biopsy from over two months to under ten days.

The boring, real version of this technology at work: not a robot doctor, but a smarter waiting line.

At UCSF, a tool called Mirai, developed with UC Berkeley and published in npj Digital Medicine, reads screening mammograms the moment they are taken, compares them against a library tuned to more than 114,000 archival scans, and flags the highest-risk ones for same-day workup. In a real run of more than 4,100 screening mammograms, it pulled about one in eight for immediate attention. For the women who turned out to have cancer, the wait from a suspicious scan to a biopsy fell from more than two months to under ten days.

It did not diagnose anyone. A human radiologist still reads every case. AI did not replace the expert; AI decided who the expert sees first: triage, not judgment.

Why this matters: Most AI-in-medicine hype promises to replace the doctor; this one is more useful precisely because it does not. The win is workflow: thresholds tuned on a huge archive, with a human still reading every scan. That integration pattern is copyable far beyond mammography. Action this week: If you or someone you love is in a high-risk screening program, ask whether your center does AI-assisted triage for suspicious findings, because here it is the difference between waiting weeks and waiting days. If you work in a health system, copy the pattern that worked: triage plus tuned thresholds plus a human in the loop, not the fantasy of an autonomous diagnostician.

ucsf.edu: How new AI cuts breast-cancer screening time for high-risk women (May 2026)
medicalxpress.com: AI speeds breast-cancer diagnosis for high-risk patients (June 2026)
dailycal.org: UC Berkeley/UCSF AI model reduces screening time for breast cancer in high-risk patients (2026)

» What to watch this week

Whether the FEC breaks its deadlock and issues a federal rule on synthetic political ads before November, or the 28-state patchwork remains the only referee. Every week without a federal standard hands the question to 28 statehouses deciding it 28 different ways, and the deepfake ad keeps running in the meantime.
How narrowly or broadly the Third Circuit writes the Thomson Reuters v. Ross holding. A ruling confined to older, non-generative tools barely moves the field; one written about copying-to-train in general resets the floor under every model built on scraped data.
Whether any school, employer, or platform that gates on an AI-detector score publishes its error rate, or quietly keeps using it. The Authors Guild test is now public; the open question is whether the institutions relying on these tools acknowledge the false-positive rate or continue flagging human writing as machine-generated.
Whether the UCSF/Mirai triage expands beyond the pilot, and whether other hospitals adopt the human-in-the-loop pattern or pursue autonomous diagnosis. The win here was workflow, not a robot doctor; the test is whether the next deployment keeps a radiologist reading every case.

Tomorrow's signal lands here.