> daily_signal(2026_10_28)
A Russian deepfake hit a Texas Senate race, a report found most firms hide AI work from workers' pay, the people who train AI are watched too, and Meta glasses narrated the world for blind veterans.
PickBits Daily Signal · Wednesday, October 28, 2026
// tl;dr
- A Kremlin-linked network put two AI deepfake videos into the Texas U.S. Senate race, falsely accusing candidate James Talarico of abusing a student, and no federal agency has named who did it. The clips drew more than 500,000 views days before the November 3 midterms; three groups of researchers, the Gnida Project, Clemson's Darren Linvill, and @antibot4navalny, attributed the campaign to Storm-1516. Attribution came from volunteers, not the government. The midterms are November 3.
- A new Oyster report found 95% of companies have employees doing AI work their official job descriptions never mention, and 62% of firms that added AI requirements changed neither the title nor the pay. AI requirements rose from 5% of roles in 2022 to 15% now, nine in ten without naming AI in the title. Only 38% of HR pros said automating part of a job is rewarded; 24% said the worker just gets more work. Oyster CEO: Hadi Moussa.
- The privacy group noyb filed five GDPR complaints against Scale AI over Outlier, the platform where 900,000 freelancers train AI, alleging hidden trackers send workers' personal data to third parties every few seconds. Trackers named: TikTok, LinkedIn, Facebook, Sprig and Datadog. Workers are also pushed through repeated face-scan verification via Persona, a vendor noyb says leaked ~70,000 government ID photos in October 2025. Complaints filed in France, Italy, the Netherlands and Austria.
- Up to 100 pre-qualified blind and low-vision veterans received Meta's AI glasses at UT Austin this week, with individualized hands-on training to actually use them. The glasses pair a camera, audio and AI to describe surroundings and read text aloud on demand. East Texas Lighthouse for the Blind organized it with the Blinded Veterans Association and Meta; it is part of a national pledge to reach 130,000 veterans. Requests go through bva.org/glasses, and this week's 100 are one stop in that pledge.
Two of today's four stories are really the same story about work, seen from different sides. We've spent the year on AI pointed at workers: the layoffs that blamed it, the skills quietly demanded, Monday's Proton re-interview. Today it's a report saying 95% of companies have people doing AI work for no extra pay, next to a complaint saying the 900,000 people who train AI get watched every few seconds while they do it. The tools got better all year while the deal for the people around them quietly got worse.
The other two are AI aimed at the public. A Russian network dropped an AI deepfake into a Texas Senate race, and no federal agency will say who did it, which is the accountability gap we'd flagged back in June, arriving right on schedule with a seat in the balance. In Austin, meanwhile, the same camera-and-AI hardware we usually cover for its surveillance patents got pointed at a problem worth solving and handed, with training, to a hundred blind veterans. The difference between the two is whether anyone is accountable for how it lands, and today only one of them is.
A Russian network smeared a Senate candidate and no agency will name it, Oyster found 95% of firms hide AI work from workers' pay, noyb says Outlier watches the people who train AI, and Meta's glasses started reading the world aloud to blind veterans.
1. A Russian network smeared a Texas Senate candidate with an AI deepfake days before the vote.
No federal agency has named who did it; the attribution came from volunteer researchers and a professor.
Two videos went up on X from an account barely two months old. They claimed James Talarico, a Democrat running for the U.S. Senate in Texas against Attorney General Ken Paxton, sexually abused a student in 2012. None of it happened. The clips were almost certainly AI-generated, drew more than 500,000 views, and were amplified by an account with over 185,000 followers. They surfaced days before the November 3 vote, which isn't a coincidence of timing.
Now look at who did the attributing. Three groups traced the campaign to Storm-1516, a Kremlin-backed network: the Gnida Project, Clemson University's Darren Linvill, and the researcher known as @antibot4navalny. The same wave has pushed AI deepfakes of Robert De Niro, Julia Roberts and Cher urging Americans not to vote for Democrats, which U.S. intelligence has attributed to Russia. What's missing is an official voice. As of late October, no federal agency had publicly named the actor. The people doing the naming are volunteers and a professor.
Why this matters: There's no designated referee for this. When a foreign network drops a fabricated child-abuse claim into a Senate race, the confirmation that it is fake comes from unpaid researchers, not from the government that runs the election. We said this in June, when the synthetic political ad era arrived with no federal rulebook; a toss-up Senate seat is now the stress test. The gap is not that nobody knows; three teams figured it out fast. It is that nobody official will say so with a nameplate behind them, days before people vote.
Action this week: Slow down before you share a political clip that makes you angry, especially in the days right before a vote, and check who has actually confirmed it: a named newsroom or one of the research groups tracking these networks, not the account that posted it. If a fabricated clip about your own race is circulating, your Secretary of State's election office is the place to report it. My own habit, after a year of watching these spread, is to treat a brand-new account with a sudden enormous reach as the tell, because that is the shape every one of these has had. The clips are built to travel faster than any correction, so the pause is the whole defense.
san.com: A social media post accuses James Talarico of sexual abuse; researchers say it's Russian disinformation (October 21, 2026)
meduza.io: For weeks, deepfake celebrities have told Americans not to vote for Democrats; U.S. intelligence says Russia is behind it (October 23, 2026)
2. 95% of companies have people doing AI work that is nowhere in their job description.
And 62% of the firms that added AI requirements changed neither the title nor the pay.
A new report from Oyster, out October 27, read 15,000 job descriptions and surveyed more than 1,000 HR and people-management professionals. The headline finding: 95% of companies have employees doing AI work their official job descriptions never mention. Two-thirds of the HR pros said their firms added AI-skill requirements to non-technical roles, marketing, HR, operations, and 46% of those never formally told the employees it had happened.
The pay is where it bites. AI requirements rose from 5% of roles in 2022 to 15% today, and nine in ten of those roles still don't mention AI in the title. Among firms that added AI requirements, 62% changed neither the title nor the pay. Only 38% of HR pros said employees are rewarded for automating part of their own job; 26% said nothing changes and 24% said the worker simply gets more work. CEO Hadi Moussa frames it as a visibility problem. To the employee, it just looks like a raise that quietly went to the company.
Why this matters: This is the quiet version of a pay cut, landing on people who are good at their jobs. The work expanded, the skill bar rose, and the title and comp stayed frozen, the same arc we've tracked all year, from AI-skill job postings climbing to the Proton re-interview earlier this week. My read is that the real risk to a manager isn't that people resist AI; it's that your strongest people quietly notice the deal changed and start taking recruiter calls. A skill you require is a skill you are buying, whether or not the pay band admits it.
Action this week: Pull the roles on your team where AI has quietly become part of the job but never made it into the description, and get that written down before the next comp cycle rather than after. Put the plain question to HR on the record: if we now expect this skill, where does it show up in the title and the band? I've watched this go sideways exactly once, and it was the undefined version: nobody could point to where the new expectation was written, so nobody could argue it was being paid for. Make it legible while you still have the leverage of a review ahead of you.
itbrief.asia: Most firms hide AI work in roles, Oyster report finds (October 27, 2026)
3. A watchdog says the 900,000 workers who train Scale AI's models are surveilled while they do it.
noyb filed five GDPR complaints over hidden trackers and forced face scans on Outlier's gig workers.
On October 27, the privacy group noyb filed five GDPR complaints against Scale AI, one each in France, Italy and the Netherlands, two in Austria, over Outlier, its data-annotation platform. More than 900,000 freelancers train AI there by writing prompts and ranking model answers. The complaint is specific about the mechanism: noyb alleges that trackers from TikTok, LinkedIn, Facebook, Sprig and Datadog sit on workers' devices and send personal data, IP address, contact details, unique identifiers, the page being viewed, to outside firms every few seconds, with no valid legal basis or consent.
It goes further. To stay on the platform, workers are pushed through repeated biometric checks, a face video plus an ID upload, run by a vendor called Persona, which noyb says suffered an October 2025 breach that exposed roughly 70,000 government ID photos. Requests from workers to see their own data were ignored, the complaint says. Data labeling is real work, and somebody has to do it. The real question is procedural: who decided these workers should be watched this closely, and where was the step at which any of them could have refused? DiPLab's Thomas Le Bonniec puts it plainly: platform work is becoming fertile ground for a whole new kind of workplace surveillance.
Why this matters: The people who make AI usable are turning out to be among the most surveilled workers there are, and the surveillance arrived the way these things usually do, bundled into a sign-up flow nobody voted on. This sits on the data-privacy arc we've tracked all year, with the roles inverted: here the humans training the model are the data being harvested. To be fair, fraud and quality control are real problems on a platform of 900,000 strangers, and some verification is defensible. Running it through a vendor with a fresh breach, and then ignoring the workers who ask what was collected, is not.
Action this week: File a data subject access request if you do this kind of annotation work, because it's your right under GDPR and it forces the platform to show you exactly what it collected and where it was sent. Read noyb's complaint first; it names the specific trackers and the legal basis it says is missing, which is the template for what to ask. My own read is that the access request is the sharpest tool here precisely because the complaint says those requests were ignored; a refusal on the record becomes evidence. If you are in the EU, the data protection authority in your country is the body now holding these complaints.
PickBits Daily Signal is free, and the best way to support it is to forward it to one person who would read it. Subscribe today!
4. Up to 100 blind Texas veterans got Meta's AI glasses this week, with the training to actually use them.
The glasses narrate the world on demand; it is part of a program aiming to reach 130,000 veterans.
Start with one moment. A blind person is walking, their guide dog stops, and they have no way to see why. On October 27, at UT Austin's Dell Campus for Advanced Research, up to 100 pre-qualified blind and low-vision veterans got a tool for exactly that gap: Meta's AI glasses, which pair a camera, audio and AI to answer hands-free questions about what is in front of you: an object, a sign, a page of text. East Texas Lighthouse for the Blind organized the event, with the Blinded Veterans Association, the Texas Technology Access Program, Austin Lighthouse and Meta.
What separates this from a press release is what came with the hardware: individualized, hands-on training, so a veteran walked out knowing how to use the thing rather than holding a box. BVA's national executive director, Lea Rowe, calls it blinded veterans helping blinded veterans. It's one event inside a Meta program aiming to reach 130,000 blind and visually impaired U.S. veterans nationwide. The caveat is real and worth stating once: this is one event, and the 130,000 figure is a national pledge, not people already reached. This is a rollout under way, not a finished one.
Why this matters: The move here wasn't the hardware. It was the person sitting beside each veteran until the device actually worked for them. We've covered Meta's glasses plenty, usually from the other direction: the always-on camera, the facial-recognition patent, and it's the same hardware here, pointed at a problem the market was content to skip. Lea Rowe's line about veterans helping veterans is the mechanism, not a slogan: the training is being delivered by people who navigate the same loss.
Action this week: Watch whether the training stays attached as the program scales toward 130,000, because that's the difference between independence and an expensive paperweight in a drawer. If you are a blind or low-vision veteran, or you know one, the request path is bva.org/glasses through the Blinded Veterans Association, with the honest note that this week's 100 are one stop in a 130,000-person pledge, not an instant supply. The thing I'll be checking next is not the hardware count but the training count: handing out 130,000 pairs is easy, and teaching 130,000 people to use them is the part that decides whether any of this reaches a daily life.
gov.texas.gov (Texas Governor's Committee on People with Disabilities): Meta AI glasses distribution and training for blind veterans at UT Austin (October 27, 2026)
kltv.com: East Texas veterans receive free Meta AI glasses (August 21, 2026)
» What to watch this week
- Whether any federal agency names the actor behind the Talarico deepfakes before November 3. Right now the only attribution is from volunteer researchers; if that is still true on election day, the accountability gap is the story, not the clip.
- Whether Oyster's finding changes a single comp cycle. The easy response is to treat "95% have hidden AI work" as a headline stat; the real test is whether any employer actually moves titles and pay to match the work people are already doing.
- Which of noyb's five complaints a data protection authority takes up first, and whether Scale AI changes Outlier's trackers or the Persona step in response. A complaint filed is not a ruling; the trackers either come off the workers' devices or they do not.
- Whether Meta's veteran-glasses program keeps the hands-on training attached as it scales, and how fast it reaches veterans beyond this first 100. The hardware number is easy to grow; the training number is the one that decides whether it reaches anyone's daily life.
Tomorrow's signal lands here.