> daily_signal(2026_09_14)
Meta was sued for scraping a decade of photos, a study found an AI class ban backfired, activists deepfaked a lawmaker backing her rival, and Google put Gemini in your camera to read for blind users.
PickBits Daily Signal · Monday, September 14, 2026
// tl;dr
- A new Illinois class action accuses Meta of building its NameTag face-recognition system, and training its Emu and Muse image models, on a decade of Facebook and Instagram photos taken without consent. It reaches past account holders to anyone whose face appeared in an uploaded photo since September 2021, and leans on the same Illinois biometric law that already cost Meta $650 million. It's a complaint, not a finding, and Meta says NameTag never shipped.
- A two-year controlled experiment at Vrije Universiteit Amsterdam found that banning AI in a law course left students worst off, below even the group that used it with no guidance. The group given structured training scored highest, but that edge nearly closed by year two, so the durable advantage was being taught to use AI well, not access to it.
- Two activists working with Turning Point Action posted AI-generated images of Arizona Rep. Lorena Austin holding "Democrats for Biggs" signs, boosting her Republican rival. A detector scored the image 100% AI. Arizona's 2024 law requires deceptive candidate deepfakes to carry a disclosure the posts never had, and the activists deleted them after a newsroom asked questions.
- Google put Gemini inside your phone's camera to describe the world aloud for blind and low-vision users. Announced in its September Android Drop, "Guided Vision" reads food labels, helps with dim menus, names objects, and coaches you on how to aim. It's rolling out free on Android 9 and later, though still a "coming soon" tool that can misread.
A class action filed this month says Meta turned a decade of your Facebook and Instagram photos into faceprints and training data, and nobody in those photos was ever asked. It leans on the same "we trained on public data" line every AI company uses, and this time that line meets the strongest biometric-privacy law in the country. The same week, two activists pointed the technology the other way, fabricated an Arizona lawmaker endorsing her opponent, and ran straight into a state law written for exactly that. And a two-year classroom study landed on something I keep chewing on: the reflex answer to AI in the room, ban it, came in dead last of the three approaches tested.
The Google one I actually liked. It took the camera already in your pocket and used it to read a nutrition label out loud to a blind shopper who couldn't. That is the same capability Meta is being sued over, this time pointed at someone who wanted it. The tech barely changes across these four; what changes is whether anyone got asked, and that is the part I can't get past this week.
Meta stands accused of turning a decade of Facebook and Instagram faces into training data without asking a single person first.
1. Meta is being sued for turning a decade of your photos into face-recognition data.
A new Illinois class action says the faceprints powering its NameTag smart glasses, and the training data behind its Emu and Muse image models, were taken from Facebook and Instagram photos without consent.
Nobody was ever asked, and that's the whole case. On September 4, the firm Wexler Boley & Elgersma filed a class action in the Northern District of Illinois (case 1:26-cv-10773) accusing Meta of extracting biometric identifiers from a decade of Facebook and Instagram photos, without notice or consent, to do two things. It built a facial-recognition system called NameTag, the capability that would let someone wearing Meta's Ray-Ban or Oakley glasses point them at a stranger and pull up an identity, and it trained its generative image models, Emu and Muse Image. NameTag is not hypothetical: code for it was spotted in Meta's companion app in June 2026, then quietly pulled after press attention. The suit reaches beyond account holders to anyone whose face merely appeared in an uploaded photo since September 4, 2021, and proposes national, Illinois, and California classes.
The legal weapon is Illinois' Biometric Information Privacy Act, the strictest biometric-consent law in the country, alongside California's right of publicity and constitutional privacy protections. That's the same statute whose per-violation statutory damages turned Meta's old photo-tagging case into a $650 million settlement. What makes this the one to watch is that it aims that law straight at the training layer: not at a shipped product, but at the claim that a company may harvest your face to build one. Meta says NameTag has not shipped to consumers and no final decision has been made, so, for now, this is a fight over how the training happened rather than over a deployed feature. And the allegations are unproven; class certification is not guaranteed.
Why this matters: Back on August 15 we flagged Meta's patent for face-scanning smart glasses as the thing to watch; this is that same NameTag capability, now with a federal complaint attached and a dollar figure in reach. Past Meta, the whole industry leans on the same three-word defense: "we trained on public data." My own read is that copyright fights have let that line survive so far, but BIPA is a different animal: it treats your faceprint as yours, requires consent up front, and does not care whether the image was public. The twist this time is that the claim goes after the training itself, not just a photo-tagging feature.
Action this week: Watch for the class notice if you are an Illinois or California resident, and do not assume "I never turned on face recognition" excludes you, because the claim is about training on your images, not a setting you toggled. Keep BIPA's per-violation damages in mind as the reason this has teeth, since that structure is what made the last Meta case worth $650 million. And if you train or buy AI on "publicly available" images, check whether any faces are sitting in that training data now, because a fair-use-style defense may not survive a consent law that never cared whether the photo was public.
wired.com: Meta sued over training data for its AI and face-recognition systems (September 11, 2026)
biometricupdate.com: Meta sued over alleged facial-recognition training for smart glasses (September 2026)
2. A two-year study found that banning AI in the classroom left students worst off.
The banned group finished below even the students who used AI with no guidance, and the trained group's early edge faded once everyone else caught up.
As US and Canadian law schools line up to bar generative AI from core teaching, a controlled experiment cuts the other way. Thibault Schrepel of Vrije Universiteit Amsterdam ran a law course across two years, 66 students in 2024 and 164 in 2025, split into three tracks: AI banned, AI allowed with no guidance, and AI with structured instruction on how to use it well. Written up on September 13 and forthcoming in Law, Innovation and Technology, the result is blunt. Across both years the no-AI group finished last, describing "idea exhaustion" and running out of ideas fast, while the structured-training group scored well above the others in 2024, especially on the demanding take-home exam.
By year two, the picture shifted. The trained group's advantage had almost closed by 2025 as baseline chatbot familiarity spread, which is itself the finding: the durable edge is not access to AI but being taught to use it deliberately. This is one instructor's law course, the samples are modest, and it's a European study, so it informs rather than settles the US fight playing out in the NYC K-8 moratorium and Florida's new district rules. A separate UC Berkeley analysis it cites, finding A grades in writing- and programming-heavy courses jumped 13 points after ChatGPT, points the same direction without being a controlled test.
Why this matters: My own instinct, watching workplaces and schools react to AI all year, was that a ban at least protects the core skill while you figure the rest out. This study says the opposite, and it does it with a control group: the prohibited students did worse than the ones who used AI with no guidance at all, so the ban was not neutral, it was the worst of the three. The durability result is what stuck with me. The trained group's edge faded once everyone learned the basics, which means the transferable skill is judgment, prompting deliberately, knowing what to verify, not one-off access to a tool.
Action this week: Swap a blanket "no AI" rule for a structured-use policy that teaches when and how to use it, and put a way to measure the outcome next to it, because this is direct evidence that restriction did not protect skill. The advantage in this study came from instruction, not access, so build the how-to-use-it teaching in rather than just unlocking the tool. And weigh the caveats honestly if you are inside the US debate: it is one modest European course, so pilot a guided-use approach and measure it, do not read it as proof that every ban is always wrong.
the-decoder.com: Two-year university study finds banning AI from classrooms leaves students worse off (September 13, 2026)
vu.nl: Students need to know how to work with AI, not fear it (September 2026)
3. Activists deepfaked an Arizona lawmaker into endorsing her Republican rival.
A detector scored the image 100% AI, and Arizona's election-deepfake law requires a disclosure the posts never carried.
Over the weekend of September 7-8, two activists working with Turning Point Action, Mary Ann Mendoza and Matt Whitmire, posted images on X showing a figure resembling Arizona Rep. Lorena Austin (D-Mesa) holding "Democrats for Biggs" and "Demócratas por Biggs" signs, implying she backs Republican Andy Biggs in the governor's race. Austin says she was never there and the images are fabricated; a deepfake-detection tool scored the photo a "100% AI score," finding "no hallmarks of natural photographic origin." Both activists deleted their posts after the local newsroom KJZZ contacted the Biggs campaign, and Austin says legal action is on the table.
This is one of the first live tests of the state AI-deepfake election laws that took effect ahead of the 2026 midterms. Arizona's 2024 statute bars distributing a deceptive "synthetic media message" about a candidate within 90 days of an election unless it carries a clear, conspicuous disclosure that the content is AI-generated, which these posts did not include, and a companion law lets a candidate go to court for a judgment that they are not the person depicted. The stakes run past one Mesa district: a cheap, convincing fake aimed at splitting a coalition, here Latino Democrats, is exactly the abuse these laws were written for. No case has been filed yet, either, so the detector result and Austin's denial are the current evidence rather than a court finding, and the remedies, a per-day fine and a declaratory judgment, are modest against content that spreads in hours.
Why this matters: A detector put the image at 100% AI, with no trace of a real camera, and the people who posted it deleted it the moment a reporter started asking. That's the whole threat in one example, a fabricated endorsement built to peel one bloc of voters away from their own party. Arizona is one of the states that saw this coming and wrote a disclosure law for it, so the real question is whether that law has any bite the first time it's tested. My read is that the fine is too small to deter and the fake moved faster than any remedy, which is the real gap these statutes still have.
Action this week: Treat an undisclosed, emotionally perfect candidate image as suspect if you vote in a state with one of these deepfake-disclosure laws, verify it against the candidate's official channels, and report a violation to the state, because within 90 days of an election that disclosure is your legal protection. If you run communications or trust-and-safety for a campaign or platform, keep a record of where every candidate image you publish came from, and have a way to pull and correct a fake fast, since the damage lands in hours while the legal remedy crawls. And if you are the public figure, keep dated originals of your real appearances and know the court mechanism that lets you formally declare a fake is not you.
kjzz.org: Turning Point activists use fake posts to claim Democrats are backing Biggs for Arizona governor (September 8, 2026)
azcapitoltimes.com: State AI-deepfake laws face first big test in 2026 midterm elections (July 21, 2026)
PickBits Daily Signal is free. If it lands in your inbox every day and it is worth something to you, the best way to support it is to forward it to someone who would read it. Subscribe today!
4. Google put Gemini in your phone's camera to read the world aloud for blind users.
Guided Vision reads food labels, dim menus, and household objects, and coaches you on how to aim the camera.
On September 1, in its monthly Android Drop, Google announced Guided Vision, a feature that pairs the phone camera with Gemini to describe the world aloud for blind and low-vision users. It reads the fine print on a food label, helps you order from a menu in a dimly lit restaurant, and identifies household objects, and when the camera is not framed right it gives spoken feedback, telling you to pan, reframe, or center the object so it can answer. Built into Gemini Live and reachable as an Accessibility shortcut or from the TalkBack menu, it's rolling out to phones on Android 9 or later in countries where Gemini is available. Google says it was designed "for and with" the blind and low-vision communities.
This goes after something people who can't read a label or make out a dim menu have needed for years: a way to handle it on their own. And it ships free, on hardware people already own, to a potential audience in the hundreds of millions. The caveats matter, though. It's a "coming soon" rollout, not a finished, everywhere-available product; a camera-and-model description can be confidently wrong, so a misread label or dosage is a real risk this does not erase; and it needs a Gemini-capable device and connectivity, which leaves out the oldest and lowest-end phones. But it is aimed squarely at helping: AI so a blind shopper can read a nutrition label without having to ask a stranger.
Why this matters: A blind shopper holding a can they cannot read, or a diner stuck with a menu in a dark room, can now read it without waiting on anyone. It uses the same camera-and-model capability Meta is being sued over, except here the person chose to point the camera themselves. In the Meta suit, nobody was asked at all.
Action this week: Turn Guided Vision on when it reaches your device if you or someone you support is blind or low-vision, set it as an Accessibility shortcut or open it from TalkBack, and try it first on low-stakes reads, a cereal box, a light switch, before trusting it with anything critical. Keep a human or tactile backup for high-stakes reads like dosages, allergens, and expiry dates, because camera-plus-model description can be confidently wrong even when it is usually right. And watch how the rollout handles the equity gap, since the people who would benefit most may be on the older, lower-end phones this leaves out.
techcrunch.com: Google's Android update tackles motion sickness, accessibility and more (September 1, 2026)
blog.google: The September 2026 Android Drop (September 1, 2026)
» What to watch this week
- The Meta case turns on a question no court has answered: can a biometric claim reach the training layer at all? The fight is not over a shipped product, it is over whether "we trained on public data" survives a consent law, so the first rulings on the motion to dismiss and class certification are the tell.
- Districts are writing AI rules right now. The VU study argues structured use beats both a ban and a free-for-all, so watch whether any of them pilot a guided-use policy and actually measure it instead of reaching for the ban.
- Arizona wrote its deepfake-disclosure law for exactly the Turning Point posts, and no case is filed yet. A per-day fine or a declaratory judgment actually landing is what tells you these state laws have teeth and not just text.
- Guided Vision demos well on labels and menus. The real test is accuracy on dosages and allergens once it ships wide, and whether the rollout reaches the older phones the people who need it most actually own.
Tomorrow's signal lands here.