Subscribe
Share
Share
Embed
A daily briefing on the AI systems, products, companies, and policy shifts that are just becoming possible.
Want a podcast for your own topics? Join early access: https://www.barelypossible.to/waitlist/?source_path=public_feed&feed_source=rss
Okay kiddos, I'm your boy Tony DeLuca, coming at you with a fresh menu of morsels for the first of September. Grab your coffee, pull up a chair, and let's have at it. We got a fake bidder scheme cooked up by one of the biggest companies on the planet, we got a Harvard dropout building a legal co-pilot for cops, we got a bunch of piracy chats coming back to bite an AI lab, and we got the Pentagon rolling out its own ChatGPT. Buckle up.
Let me start with the one that, if you're a founder who buys ads anywhere, should make you sit up straight. Because this is the kind of story that's easy to file under "big company gets sued again, whatever," and then move on. Don't do that today. Slow down for this one.
The Federal Trade Commission and twenty-two states filed a lawsuit against Amazon on Monday, accusing the company of running what they call a secret ad surcharge scheme. And when you actually read what they're alleging, it's not the usual vague antitrust hand-waving. It's specific, it's mechanical, and it strikes right at something every one of you who has ever run a Sponsored Products campaign has trusted without thinking twice.
Here's the setup. Amazon told more than five hundred thousand small and medium businesses that it ran what's called a second-price auction. Now, for those of you who don't live in ad tech, let me explain why that matters, because it's the whole ballgame. In a second-price auction, the winning advertiser doesn't pay what they bid. They pay just one cent more than the next-highest bid. That's the promise. And that promise changes your behavior completely. Because if you know you'll only ever pay a hair above the runner-up, you can bid high with confidence. You bid your true value, you trust the system to protect you, and you go about your day. That's the entire elegance of it. It's supposed to be honest by design.
Except, according to the complaint, starting in 2019, Amazon quietly changed the machinery. The FTC says they added a hidden surcharge that internally they called a "soft reserve price." And here's the part that made me put my coffee down. The complaint says Amazon used what one internal document called an "invented auction participant." A fake bidder. A ghost in the machine that doesn't exist, doesn't want your product, doesn't have a business, but whose job is to sit in the auction and push your price up higher than real competition ever would have.
The FTC calls that a shill bid. And in any old-school context, a shill is exactly that guy at the auction house who bids up the painting so the sucker in the front row pays more. Same move, just with servers and code instead of a paddle. The complaint alleges that as a result, Amazon charged Sponsored Products advertisers their own full winning bid close to eighty percent of the time. Read that again. The thing they sold you as a second-price auction, where you pay a penny over the next guy, turned into a first-price auction where you pay your whole bid, four times out of five, and nobody told you.
The FTC says Amazon did this because it wanted more advertising revenue, and kept it hidden because if advertisers knew, they'd lower their bids, which would cut into that revenue. And the money here is not small. The company generated more than sixty-eight billion dollars in advertising revenue last year. The lawsuit says the alleged practice affected more than a million brands and sellers and may have generated tens of billions of dollars in extra revenue.
Now, Amazon fired back. In a blog post they called the lawsuit "misguided," and argued that the complaint "fundamentally misunderstands how advertisers operate." They say their auctions evaluate billions of bids across different placements and formats, so prices naturally vary, and that advertisers are, quote, "properly" informed about the pricing system. That's their defense, and it'll get litigated, and I'm not the judge.
But here's why I'm spending real time on this instead of just reading you the headline. If you're a builder, if you're a founder buying ads to grow, this is a story about trust in a black box. You cannot see inside the auction. You get a number at the end and you pay it. The entire relationship depends on the platform describing the rules honestly. And the allegation here is that the described rules and the actual rules were two different things for seven years. That's the whole game for anybody whose growth depends on a channel they don't control.
And I want you to sit with the second-order lesson, because it's bigger than Amazon. Every ad platform, every marketplace, every place you spend money to acquire customers, is a system where you're a price-taker in somebody else's machine. You optimize your campaigns, you A/B test your creative, you get your cost-per-acquisition down, and you feel like you're in control. But the floor you're standing on is a floor the platform can quietly move. The lesson isn't "don't advertise." You gotta advertise. The lesson is: diversify your acquisition, don't let any single black-box channel become your only lifeline, and stay a little bit paranoid about metrics you can't independently verify. When the counterparty controls the auction, the audit, and the invoice, "trust me" is the entire product. And "trust me" just got a twenty-two-state lawsuit.
One more thing that ties this together with something we've circled before. We talked back on the twenty-second about the FTC going after personalized pricing through disclosure rules rather than an outright ban. And here's the same regulatory instinct showing up again: the government's theory isn't that Amazon can't run an auction how it wants. The theory is that they told people one thing and did another. It's a disclosure case at its heart. The FTC keeps reaching for the same tool — not "you can't do this," but "you can't hide that you're doing this." Founders, take the hint about where the enforcement energy is pointed. It's pointed at the gap between what you say and what you do.
Alright, let's shift from a company allegedly hiding a fake bidder to a lab getting its own words read back to it in court.
Anthropic is back in the legal soup, and this time the ammunition is coming from inside the house. Ars Technica reported on a Sony suit — this is part of the broader music publisher litigation, Sony Music Publishing, Warner Chappell and others — where the plaintiffs are citing Anthropic's own internal staff chats. And the headline quote, which the lawsuit pulls out, is an employee writing "Zlibrary my beloved." Z-Library, for those who don't know, is one of the big shadow libraries, a pirate repository of books. So you've got staff chats where people are, in the lawsuit's framing, extolling piracy, and now that's Exhibit A in a case about whether Anthropic illegally torrented and scraped copyrighted works to train Claude.
Now let me be careful and precise here, because this connects to something we covered just a few days back. We talked about Judge Alsup's ruling in the Bartz case — the one where the takeaway was that training AI on copyrighted works can be legal, but acquiring that content through piracy is not. Anthropic got hit with a one-point-five billion dollar judgment in that case on exactly that distinction. Legal to learn from it, not legal to steal it to learn from it.
This new music-publisher suit is building right on top of that foundation. Some of the same lawyers are involved. And the reason those internal chats matter so much is precisely because of the Alsup distinction. When the legal line is "was the acquisition itself piracy," then a staff message saying "Zlibrary my beloved" isn't just embarrassing. It's evidence of state of mind. It's the difference between "we accidentally ingested something" and "we knew exactly what we were doing and we loved it." Anthropic, for its part, has said in the related Warner case that it disagrees with the publishers' claims and intends to defend itself robustly in court.
Here's the builder lesson, and it's a boring one, but boring lessons keep you out of court. Your Slack is discoverable. Your internal chat is discoverable. The joke your engineer types at eleven at night about where the training data came from can end up quoted in a federal complaint two years later with your company's name on it. I'm not telling you to be humorless. I'm telling you that "move fast and pirate stuff" as a company culture leaves fingerprints, and the fingerprints are timestamped and searchable. If your data provenance can't survive being read aloud by a plaintiff's attorney, you don't have a data strategy, you have a liability with a countdown timer.
Now let's go from a lab in trouble to a startup that's trying to do the careful version. Let's dig into a Harvard Law dropout who's building what he calls a Harvey for police officers.
This is my deep dive today, and I picked it not because it's the flashiest AI story on the menu, but because it's the most instructive one for anybody building a vertical AI product in a high-stakes, high-scrutiny field. And it's got everything: a dramatic origin, a real product-market claim, and a genuinely uncomfortable set of questions underneath.
The company is called Blue Voice. The founder is David Lawrence. Per the TechCrunch reporting, Lawrence was a Harvard Law student when an on-campus shooting involving a police officer sparked intense controversy over police conduct. And his read on that — the insight he dropped out to chase — was that most policing errors happen when officers don't have instant access to the department's own rules. Not malice, necessarily. Just a human being under pressure trying to recall the right procedure from a mountain of law, policy, and ordinance they're supposed to carry around in their head.
He teamed up with two co-founders. Amit Patankar, described in the piece as a Harvard MBA and former Google engineer. And Michael Gropman, a retired Boston police deputy chief. And that third guy matters a lot, because building tools for cops without anybody who's ever actually been a cop is how you build something that gets laughed out of every roll call in America.
Three years later, Blue Voice is coming out of stealth with six million dollars in funding, led by SignalFire and Las Olas VC. And here's the traction number that made me pay attention: officers at two hundred twenty-five county agencies across twenty-five states now rely on the tool daily. Lawrence likens it to Harvey, the AI tool widely used by lawyers, or OpenEvidence, which is the similar idea for doctors. A specialist co-pilot trained on the exact rulebook of your profession.
Let me read you how Lawrence frames the problem, because he says it well. Quote: "When an officer makes a decision, they do so from memory or best guess, based on thousands of pages of law, policy, municipal ordinances, and state laws that they've internalized and keep in their head." End quote. And before Blue Voice, per the piece, an officer who couldn't recall a specific protocol — like the exact seventh step required at a crime scene — had few good options. Flip through a fifteen-thousand-page manual. Wake up a supervisor in the middle of the night. Or search a generic tool like Google or ChatGPT that has no police-specific training. And Lawrence claims those consumer AI tools delivered incorrect answers up to thirty percent of the time.
That thirty-percent number is the whole reason this is a company and not a ChatGPT prompt. This is a theme we keep coming back to on this show — the vertical specialist beating the general-purpose giant not because it's smarter, but because it's grounded in the right data. Blue Voice says it's trained on department-specific laws, local ordinances, protocols, and guidelines that general tools can't even access on the public internet. Your municipal code isn't sitting in a crawl of the open web. And here's the design choice I actually respect: Lawrence says officers trust it because it always points them directly to the original regulation rather than just generating an answer on its own the way ChatGPT would. It doesn't say "here's what to do." It says "here's the law, here's the citation, you decide." The tool provides the rule. The officer, combined with field experience, makes the call.
That's a really important architectural decision for a high-stakes domain, and I want builders to hear it. When the cost of a hallucination is somebody's civil rights or somebody's life, you don't build an oracle that hands down verdicts. You build a librarian that hands you the source and gets out of the way. Retrieval with citation, human in the loop, decision stays with the person accountable for it. That's not just good ethics, it's good product. It's how you get an officer to actually trust the thing enough to pull it out on a Tuesday night.
And Lawrence brings receipts. Over the last year, he says Blue Voice grew its customer base elevenfold, because departments see concrete results. He tells a story of a rookie officer who spotted a man pressuring a young girl to get into his vehicle, and the officer wasn't sure he had legal grounds to intervene. He consulted Blue Voice on his phone, and the app confirmed the situation met the legal criteria for "child enticement," which gave him the authority to act. Another story: the tool reminded a department chief he couldn't return an officer to active duty after a shooting without first getting a third-party mental health evaluation. They've even built functionality to help detectives work cold cases. The company competes with a PE-backed incumbent called Lexipol, so this isn't a greenfield — it's a modernization play against an established player.
Now. Here's where I put on my skeptical hat, because that's my job, and because you're a builder and you need to think about this stuff before you build in a space like this. This is AI for policing. And the piece itself is honest enough to note the elephant in the room: at a time when police use of AI, like Flock Safety's license plate surveillance systems, is drawing growing criticism, Lawrence is positioning Blue Voice as the good version. He says he hopes it proves that properly deployed AI can deliver public safety benefits without sacrificing civil rights.
And look, on the merits, a tool that tells a cop "the law does not permit you to do this" is genuinely different from a tool that scans every license plate in your town and never forgets. We've spent a lot of time on this show on the Flock backlash — cities canceling contracts, surveys showing more Americans oppose those cameras than support them. A compliance co-pilot that keeps officers inside the rules is, at least in theory, the opposite of dragnet surveillance. It's friction against overreach, not fuel for it.
But here's the question I'd want answered if I were an investor or a citizen. Who audits the answers? Who decides what "the law" is when the situation is ambiguous, which is exactly when officers reach for it? If the model is confidently wrong in that thirty-percent gray zone, an officer with a citation on his phone feels more certain, not less, and false confidence in a squad car is its own kind of danger. And what happens to the queries? Every time an officer asks "do I have grounds to intervene," that's a log. That's a record of hesitation, of judgment, of what was happening at a scene. That data has enormous value and enormous sensitivity, and the piece doesn't get into who holds it or how it's protected.
I'm not saying Blue Voice is bad. Honestly, of all the ways AI is showing up in policing, "give the officer the actual rule and let them decide" is one of the more defensible ones. What I'm saying is that if you build in a domain like this, the product is only half the job. The governance, the audit trail, the answer to "who's accountable when the tool is wrong" — that's the other half, and it's the half that determines whether you're building the good version or just a better-marketed version of the thing everybody's mad about. The builders who win in high-stakes verticals aren't the ones with the best model. They're the ones who took the accountability questions as seriously as the accuracy questions.
Alright. From cops with AI co-pilots to the biggest customer of all rolling out its own chatbots. Let's talk about the Pentagon.
The Department of Defense has launched custom versions of OpenAI's ChatGPT and xAI's Grok, giving three million civilian and military personnel access to generative AI tools tailored to, in their words, "warfighter needs." They're called ChatGPT Mil and Grok for Government, and they're part of a secure portal called GenAI-dot-mil that launched last year. That portal, by the way, started out offering Google Gemini. The whole point of it is to let Defense Department employees use commercial frontier models without shipping sensitive government data through ordinary consumer channels — the military version is exempt from the data collection you can't avoid in normal consumer products.
And here's a number that tells you the demand is real, not a press release: the portal has already onboarded more than one-point-seven million unique users out of three million personnel. That's more than half the department using it. That is not a pilot. That is adoption.
But here's the part that matters for you as a builder watching the AI market, and it connects directly to a story we've been tracking. Notice who's absent. Anthropic's Claude is not on that list. And the piece is explicit about why: Anthropic was labeled a supply-chain risk by the Trump administration after it refused to give the Pentagon unrestricted use of its tools and instead insisted on certain safety guardrails. We covered this — the judge ruling that label illegal, First Amendment retaliation, the whole thing. Anthropic is fighting that designation in court, and got an early win.
So here's the picture that emerges. The Pentagon wanted no-strings access. Anthropic said no, we have guardrails, we won't do autonomous weapons or mass surveillance of Americans. And now, while Anthropic is winning in court, its competitors are winning in the building. OpenAI's in. xAI's in — and notice the tonal difference in how they're pitched. ChatGPT Mil is described in careful terms: document-heavy, routine unclassified work, admin tasks, logistics, planning, policy. Grok for Government, running on SpaceX's Starshield satellite network, gets described in much more militaristic language — helping the military "execute missions faster and with greater precision across numerous operational contexts."
The builder lesson here is a genuinely hard one, and I'm not going to hand you a tidy moral. There's a real business cost to drawing ethical lines. Anthropic drew a line on how its tools could be used, and the immediate consequence is that its competitors are onboarding a customer with three million seats while Anthropic is in a courtroom. That's the trade. Sometimes the principled position and the market position point in opposite directions, at least in the short run. Whether the line pays off long-term — in trust, in enterprise credibility, in not being the company whose model steered a weapon — that's a bet Anthropic is making with real money on the table. And if you're building in a regulated or high-stakes space, you're going to face a smaller version of that same fork in the road. Where's your line, and what does it cost you to hold it? Amazon Web Services, Microsoft, Nvidia, and Reflection AI are all in the Pentagon tent too, by the way. The building is filling up. The question is just who's willing to sign the terms.
Okay, let's do a lightning round of things you should know about, and then I'll get out of your hair.
First, a data breach you need to hear about if you're anywhere near healthcare or you just, you know, have a body that sees doctors. A hacking group called ShinyHunters — one of the most active data-extortion crews of the last couple years — has taken credit for a cyberattack on McKesson, the pharmaceutical distribution giant. McKesson confirmed hackers broke into several of its cloud-hosted accounts and exfiltrated data. And the how here is the part that should make every founder's stomach turn: ShinyHunters told TechCrunch they got in by tricking several employees into granting access through phishing and social engineering. Not some zero-day exploit. Not a genius hack. They talked their way in. They pulled millions of rows of patient data out of the company's Snowflake and Salesforce environments — names, addresses, Social Security numbers, diagnoses, medications, allergies. Bleeping Computer reported a fifty-five million dollar ransom demand. And this is part of a pattern — Boston Scientific, Stryker, and others all hit recently. The lesson, said plainly: your fanciest cloud security stack does not matter if Dave in accounting clicks the wrong link and hands over the keys. The perimeter is your people, and your people are tired and busy. Phishing training isn't a checkbox, it's the whole ballgame.
Second, on the OpenAI side of the ledger, the company announced that ChatGPT Ads has reached one billion dollars in annualized revenue run rate and is expanding globally. Now that's a short post with not a lot of meat, so I'll keep it short too, but the direction is worth marking. The company that spent years insisting it was about the mission is now running a billion-dollar ad business inside the chatbot. If you're building a consumer product on top of ChatGPT, understand that your platform provider now has an ad model with its own incentives, and ad models and clean user experiences have a long history of not being friends. Watch that space.
And third, from across the pond, the EU's toughest online safety rules — this is the Digital Services Act framework — now apply to ChatGPT and to Reddit. The Financial Times reporting, carried by Ars, frames it simply: explosive growth comes with a new regulatory burden. When your platform gets big enough in Europe, you graduate into the strictest tier of content and safety obligations. For founders, the meta-lesson is one we keep relearning: scale is not just a growth metric, it's a regulatory trigger. The bigger you get, the more governments treat you like infrastructure. Plan for it before it arrives, not after the letter shows up.
And I'll leave you with one palate cleanser, because it's the first of September and you deserve a little something. Google Maps went ahead and renamed Lake Ontario to "Lake America" — and did it faster than the actual US government, whose official maps are, per the reporting, still waiting. Make of that what you will. When the map company moves quicker than the mapmaker, you learn something about who's actually running the joint.
That's the menu, kiddos. Watch your ad spend, watch your Slack logs, and if a device promises you free movies in exchange for hooking into a mystery network — and yeah, that's a real thing on today's list — just walk away. This is Tony DeLuca, and I'll be right back here in your ears tomorrow. Be good to each other out there.