> daily_signal(2026_07_22)

A court just priced a pirated training book at $3,000. DHS is tripling its AI border towers by 2034. Apple started recording the Genius Bar. Samsung's new assistant reads your health data back to you.

PickBits Daily Signal · Wednesday, July 22, 2026

By Mark Pickering · 9 min read · July 22, 2026

// tl;dr

A federal judge approved Anthropic's $1.5 billion book-piracy settlement at $3,000 a book. Judge Araceli Martínez-Olguín signed off on July 20, covering more than 482,000 works, about 91% of them already claimed. The court did not rule that training on copyrighted books is illegal; Anthropic's fair-use position on training held. Liability attached to how the library was acquired, from Library Genesis and Pirate Library Mirror.
DHS plans to grow its network of AI surveillance towers from 830 to 2,300 by 2034, with $1 billion already appropriated. The EFF broke down a July 2026 GAO report on July 20. The newer towers combine radar, thermal infrared, and optical sensing with software that classifies and tracks targets on its own, so a human gets an alert and never sees the feed. Anduril builds the Sentry model; expansion is named for Arizona, Texas, and coastal California.
Apple has started recording and AI-summarizing conversations between Genius Bar staff and customers at a small number of stores. Bloomberg's Mark Gurman reported the Live Notes pilot on July 19. Both the worker and the customer must consent, audio is not retained, managers do not receive transcripts, and employees can edit before filing. Store workers still say the same pipeline could be turned on them.
Samsung switched on an opt-in AI Health Assistant for US users, and it is built to read your health data against itself instead of leaving it siloed. Live in beta on July 21 across Galaxy phones, watches, and the ring, it reads sleep, activity, nutrition, mindfulness, and vitals together. Samsung says physicians and certified health coaches validated the guidance. It will not diagnose, and it reads your calendar and smart home to do the correlating.

We covered this settlement on July 7 from the opposite end. Roughly a hundred authors had opted out of the class to sue on their own, chasing up to $150,000 a book, and at the time that was a bet placed against a number nobody had. Monday the number arrived. It is $3,000, and my read is that the authors who stayed in did worse than the headline suggests, because that figure now sits on every negotiating table as the cheapest a lab should ever expect to pay for a book it took. Elsewhere today, Homeland Security already has a billion dollars appropriated through 2034 for towers that decide what people are doing in their own towns, and Apple started recording what you say at a repair counter to save its staff some paperwork. Samsung is the story I would keep out of today's four. It reads years of your own health data back to you and refuses to diagnose, which is the opposite direction from everything else today.

A court decided Anthropic could keep everything it learned from 500,000 books, and owed $3,000 each for how it got hold of them.

1. The settlement we picked up on July 7 finally has a number, and it works out to $3,000 a book.

The court let the training stand and billed Anthropic for the download.

Judge Araceli Martínez-Olguín, of the U.S. District Court for the Northern District of California, gave final approval on Monday, July 20 to Anthropic's $1.5 billion class settlement with authors and publishers, finding it provides "meaningful relief." The payout works out to roughly $3,000 per work across an estimated 500,000 works, shared among whoever holds the rights to each. The settlement covers more than 482,000 books, and about 91% of them have already been claimed by authors or publishers now due payment. Plaintiffs' lead attorney Justin Nelson called it "the largest known copyright recovery in history." The case was filed in 2024 by thriller novelist Andrea Bartz and two other authors.

The court did not hold that training an AI on copyrighted text is illegal, and Anthropic's fair-use position on training survived intact. Acquisition is where it lost. Anthropic bought and scanned part of its library, which was fine, and downloaded the rest from Library Genesis and Pirate Library Mirror, admitting it "illegally downloaded and stored millions of copyrighted books." Anthropic won the bigger half of that argument and still has to write the check, because piracy-sourced acquisition was never an Anthropic-specific habit. Basically everyone built their corpus that way back then. On July 7 we wrote up the authors who opted out of this deal to chase statutory damages of up to $150,000 a work; they made that bet against a number nobody had yet, and now they have it.

TechCrunch July 20 2026 report that Judge Araceli Martinez-Olguin of the US District Court for the Northern District of California approved Anthropic's landmark 1.5 billion dollar copyright settlement delivering about 3000 dollars per work across an estimated 500000 works with Anthropic's fair use position on training upheld and liability attaching to books downloaded from Library Genesis and Pirate Library Mirror — techcrunch.com · July 20, 2026

Why this matters: If you buy or deploy AI models anywhere, your vendor questionnaire is asking the wrong thing. Stop asking whether they train on copyrighted material, that question now has a defensible yes. Ask how the corpus was acquired instead, because that's where the liability actually sits, and most model cards do not touch it. Action this week: Check whether your book sits in the class and file for it, because roughly 9% of the 482,000 works were still unclaimed at approval and each is worth about $3,000; start from the case coverage, follow through to the official claims site, and ignore any emailed link, since settlements this size reliably attract phishing. Vendor due diligence gets one new line in the template this renewal: can you attest that no portion of your training corpus was obtained from a shadow library? And if you're a rights-holder negotiating an AI licensing deal, $3,000 a work is your floor now, not your ask, because that's what a court just accepted for unlicensed past use.

techcrunch.com: Anthropic's landmark $1.5B copyright settlement is approved (July 20, 2026)
usnews.com (AP): Judge approves a $1.5B Anthropic settlement over pirated books used to train the Claude chatbot (July 21, 2026)
courthousenews.com: Anthropic to pay $1.5 billion copyright settlement to authors, publishers (July 2026)
fortune.com: Anthropic copyright settlement and what authors get (July 21, 2026)

2. DHS is going from 830 AI surveillance towers to 2,300, and the billion dollars is already appropriated.

The tower decides what your movement means before a person ever looks.

A July 2026 Government Accountability Office report, broken down by the Electronic Frontier Foundation on July 20, shows the Department of Homeland Security plans to grow its network of surveillance towers along US borders from 830 today to 2,300 by 2034. EFF's read puts the spend at $1 billion, drawn from the tax and spending law signed in 2025, which means the spending is already authorized and under way. The newer towers in the Integrated Surveillance Tower program combine radar, thermal infrared, and optical sensing with AI that autonomously detects, classifies, and tracks what it decides is a target over long distances, then hands a human the alert instead of the raw feed. Anduril Industries supplies the Sentry autonomous model, and DHS awarded the company more than $360 million for towers in December. Named expansion runs into Arizona, Texas, and coastal California communities.

Forget the tower count for a second. Each of these decides on its own what your movement means, and it does that permanently, in towns where the overwhelming majority of people inside the detection cone live there and are not crossing anything. Anduril's founder, Palmer Luckey, describes it as thousands of sensors pulled into one live picture of everything moving out there, every person and vehicle tagged automatically, so operators respond to problems instead of hunting for them. Luckey's not wrong that border security needed better eyes. What bothers me is what happens after the tower makes the call. The same GAO annual assessment, GAO-26-108118, records that the Integrated Surveillance Towers program did not achieve its major milestones in fiscal 2025, so they're building faster than they're checking. There's no published false-positive rate at all, for any of it. We covered the Supreme Court's ruling that phone location data is constitutionally protected on July 4, and DHS paying Thomson Reuters $125 million for ICE access to a commercial people-search database on July 20. This time it's bolted to poles instead of bought from a broker.

Electronic Frontier Foundation Deeplinks July 20 2026 explosion of surveillance towers coming to US borders costing over 1 billion dollars reading a July 2026 GAO report showing DHS plans to grow towers from 830 to 2300 by 2034 with autonomous towers using radar thermal infrared and optical systems and AI target tracking from Anduril Industries Sentry model expanding into Arizona Texas and California coastal communities — eff.org · July 20, 2026

Why this matters: If you live in a border community in Arizona, Texas, or coastal California, an AI is already classifying what your movement means, and there are about to be nearly three times as many of them doing it. Action this week: Go find out what's already near you before the next tranche lands, because tower siting is decided locally and the existing locations are the input to any council comment or public-records request; EFF maintains a public map and dataset of border surveillance towers as part of its Atlas of Surveillance work, linked from its July 20 breakdown. Policy, journalism, and oversight folks: pull GAO-26-109195 and GAO-26-108118 yourselves, because the missed fiscal-2025 milestones only show up in the annual assessment and never make the coverage. Nobody building or buying detection systems like this has had to answer what happens to the person the tower flags by mistake. There's no published false-positive rate for a $1 billion, 2,300-node deployment, and no answer for who tells that person, or how they argue back. Send this to whoever sits on your town council.

eff.org: An explosion of surveillance towers is coming to US borders, costing over $1 billion (July 20, 2026)
gao.gov: GAO-26-109195, northern border security (July 2026)
gao.gov: GAO-26-108118, DHS annual assessment, Integrated Surveillance Towers milestones
fedscoop.com: DHS budget, border wall and surveillance spending (2026)

3. Apple is recording and AI-summarizing what you say at the Genius Bar, and every safeguard on it is a policy.

Apple wrote three good rules and can rewrite all three.

Bloomberg's Mark Gurman reported on July 19 that Apple has begun a limited pilot called Live Notes at a small number of Apple Store locations. It auto-transcribes the conversation between a Genius Bar employee and a customer, generates an AI summary, and stores that in the repair session record on the employee's iPad. Both the employee and the customer must consent or Live Notes does not activate. The audio recordings themselves are not saved. Managers do not receive the transcripts. And the employee can edit both the transcript and the summary before either is written into Apple's internal system. The stated purpose is cutting the documentation time Genius Bar staff spend writing up interactions, which is a real burden on that job.

This same pipeline is already being sold into call centers and clinic intake elsewhere. Apple just happens to be the biggest name running it first, so anyone configuring this can just copy Apple's three choices. The software does not enforce any of them. Apple wrote those rules and Apple can rewrite them, and once the transcripts are sitting in the work record, that rewrite is the only thing between them and a performance review. Apple Store workers quoted in the coverage put it exactly that way: the concern is not what Live Notes does today, it is that "something almost exactly like this could be used for monitoring and evaluating employees" tomorrow. Workplace monitoring software already went further than this. It has tracked screen time, keystrokes, break duration, and idle time, rolled it into a performance record the worker never saw, and in Amazon's warehouse case ended jobs when the numbers came up short, which is what produced the 2018 class action. Back in May we called the leaked Meta audio the template for what your employer is allowed to do with a company laptop. Live Notes just adds the microphone.

Why this matters: If you manage anyone who talks to customers for a living, an AI note-taker is coming to your roadmap, and Apple has now worked through the three decisions you will be handed. Action this week: Put those into your requirements before procurement, not after: consent from both sides with a refusal path that penalizes neither, and nothing retained that a manager can later pull into a review. Vendors default to the opposite of all three, because retention and manager dashboards are what they sell. Being recorded at work already? Get a written answer to one question: is the transcript or summary admissible in a performance review or a disciplinary process? Apple's pilot says managers do not receive transcripts, so get your own employer's version of that sentence into the policy, because a verbal assurance at rollout is not enforceable at review time. Whoever sets that policy has a legal reason to want dual consent too, not just an ethical one: it's what keeps the tool lawful in the eleven US all-party-consent recording states, so audit which of your sites sit in those jurisdictions before approving a single-consent configuration. Send this to whoever owns that policy.

gizmodo.com: Apple has reportedly started recording and AI-summarizing conversations at the Genius Bar (July 19, 2026)
petapixel.com: Conversations between Apple Store workers and customers can now be recorded (July 21, 2026)
bloomberg.com: Mark Gurman, Power On newsletter (July 19, 2026)

4. Samsung switched on a health assistant that reads your sleep, activity, and vitals against each other, and it will not diagnose you.

Years of your own data, finally read together.

Samsung launched Health Assistant in beta on July 21, 2026 for eligible US users, opt-in, inside the Samsung Health app and working across Galaxy phones, Galaxy Watches, and the Galaxy Ring. It monitors five wellness pillars, sleep, activity, nutrition, mindfulness, and vitals, and its actual function is "showing how each category influences the others," which no single-metric dashboard does. Samsung says the recommendations "have been validated by a team of physicians and certified health coaches." The product also draws a hard boundary: "It won't offer medical advice or suggest treatments, and it can't diagnose health conditions." The rollout landed the day before Samsung's July 22 Galaxy Unpacked event but is a separate release.

Samsung's version finally does that reading for you. Most people wearing a watch or a ring are sitting on years of health data they've never actually looked at together, because sleep shows up on one screen and heart rate on another, and the useful part is almost always how one drags on the other. It won't diagnose, and that's the line between a wellness tool and an unapproved medical device. It's also the line the FDA-cleared systems we covered this month had to go through clearance to cross: UpDoc on July 6 for insulin management, iHealthScreen on July 14 for diabetic-retinopathy screening. Cardiologists using consumer wearables in practice describe the same discipline from the clinical side, confirming a watch reading on medical-grade equipment before anyone touches a treatment. It reads more than your health data to do that, including your calendar and your smart-home devices. Engadget's worked example is the assistant flagging late-night smart-light usage as a factor in poor sleep. That's a lot more of your life than a health app usually sees, which is why the opt-in actually matters here. And a wellness app isn't a HIPAA-covered entity, so the protections you assume sit around medical records don't automatically sit around this one.

Why this matters: You've got years of health data on your wrist, and reading it across sleep, activity, nutrition, mindfulness and vitals at once has meant doing it yourself. Samsung's free, opt-in version went live yesterday for US users. Action this week: Galaxy phone, watch or ring owners: open Samsung Health, enable Health Assistant, and ask it a cross-pillar question such as why your sleep quality dropped this month. Single-metric questions waste it, because connecting the dots across categories is the one thing it does that your existing screens don't. Before you enable it, decide separately whether to grant the calendar and smart-home permissions, since those are what let it explain your sleep using your schedule and your lights, and they are also the largest expansion of what a health app can see about your life; the health features still work without them, so decide it deliberately instead of clicking past it at install. And if you manage a condition or take a prescribed medication, use this as conversation prep and nothing more, because Samsung states plainly that the assistant will not advise, treat, or diagnose. Bring its pattern findings to a clinician who can act on them.

engadget.com: Samsung aims to help you make more sense of health data with a new AI-powered assistant (July 21, 2026)
news.samsung.com: Samsung launches Health Assistant beta, first fully integrated AI-powered assistant (July 2026)
thenextweb.com: Samsung Health Assistant arrives in beta for US users (July 2026)
pymnts.com: Samsung adds AI-powered assistant to health offerings (July 2026)

» What to watch this week

Whether the $3,000 figure starts showing up as an anchor in the other training-data cases, starting with the publishers and authors now suing Google over Gemini. This deal settled with money instead of a precedent, so the next filings tell you whether plaintiffs treat $3,000 a work as a floor. My guess is they will, and that Google argues it down hard.
The appeal we flagged on June 21, the first to test whether training on copyrighted works is fair use. That one binds every model builder in a way a negotiated settlement never will.
Whether GAO's fiscal-2026 milestones for the Integrated Surveillance Tower program are met, and whether a false-positive rate is ever published. The program missed fiscal 2025 while the installed base kept growing, and the annual assessment is the only public place that gap shows up. I do not expect a false-positive rate. I would like to be wrong.
Whether Apple expands Live Notes past the pilot stores, and whether the no-manager-transcripts rule survives it. The workers are worried about what changes once this scales past a handful of pilot stores. The policy text is where you'd see that first.
What Samsung retains from the calendar and smart-home access it now has. The correlation is the whole product, so check the retention terms on the non-health inputs before this ships beyond the US beta.

Tomorrow's signal lands here.