Latest / Super Intelligence News and Report / EU Adviser: AI Training on News Is Free, but Answers May Cost
Transcript
- Super Intelligence News and Report — EU Adviser: AI Training on News Is Free, but Answers May Cost Welcome to Super Intelligence News and Report. Today is Wednesday, November 4th, here are the headlines. Theo: Today's thread is Google's grip on search, and how it holds up now that search is turning into AI answers. We'll get there, but first, chips. Maya: Chips and China. The Financial Times reports, relayed by Tom's Hardware on November 3rd, that Nvidia told its cloud partners to keep Chinese customers under half their business. That's reported, not confirmed. The rule reportedly applies worldwide, not just to data centers in Asia. Nvidia's statement said American model makers need American chips, and it encourages partners to take business everywhere they can. It didn't confirm the cap. The point is that Nvidia is now policing China access through its own cloud partners, not only through export rules. Theo: How much does China matter to their top line right now? Maya: China-headquartered customers were a bit over eight percent of revenue last quarter, by one analyst's math from the filing. We'll know more at earnings on November 17th. And while we're on chips, AMD had a big night. Record third-quarter revenue, data center sales nearly doubled on EPYC and the new MI450 GPUs, and it beat on earnings. The headline is the guide: about seventeen billion dollars next quarter, roughly a quarter above this one. The stock rose about eight percent after hours. Theo: After weeks of AI-spending jitters, that's a real number to point at. Maya: It's the first strong AI-demand data point in a while. Nvidia had already bounced that day after four losing sessions, and neoclouds like CoreWeave and Nebius rallied. Theo: Now that Google thread. On November 3rd, Google's education chief Chris Phillips wrote that Gemini, Gemini Notebook, Gemini in Classroom and Google Search will be fully off for K-12 student accounts until administrators turn them on. So Search and the AI tools now come as one package. Google is also adding Guided Learning, a step-by-step coaching mode, to Classroom now, with a Search pilot later this month. Former FTC Chair Lina Khan called it Google using its search monopoly to coerce schools into adopting AI. Maya: And it reaches a lot of kids. Theo: Google says about a hundred and fifty million people use Classroom. The takeaway is that Google now treats Search as an AI product. If a school doesn't want the AI, it loses Search too. Maya: Over to debt markets. Meta raised fifteen billion dollars on November 3rd in a five-part bond sale to fund AI infrastructure. That's on top of the stock-sale plan and S&P's negative outlook we covered last week. Orders came in about four times the deal size, which sounds hot, but it's roughly the big-tech average this year. And the coupons are high, up to about six and a quarter percent on the longer bonds. Theo: So the money's there, it just costs more. Maya: That's it. Next, Fal, which serves image, video and audio models behind one API, raised one point three billion dollars at a sixteen billion valuation. That's double its last round. Bessemer led and Nvidia is a strategic investor. Bloomberg cites a person saying annual revenue passed eight hundred million earlier this year, which is reported, not confirmed. Theo: Which puts it right behind Fireworks. Maya: Just behind, yes. Inference for media generation is now a business in its own right. Then to Beijing. On November 3rd, Wang Ruomeng of the National Development and Reform Commission, China's economic planner, said China will speed up AI legislation. It will set safety standards and ethics rules covering AI from development through deployment, plus regulatory sandboxes for specific sectors. Bloomberg frames it as among the most complete loss-of-control safeguards anywhere, and it covers DeepSeek, Moonshot and Alibaba. Separately, a vice head of China's financial regulator warned that AI investment is getting too concentrated. Theo: So they want to compete and put guardrails on at the same time. Maya: That's the pairing. Speaking of competing, you've got the Anthropic chip story. Theo: The Information, citing four people, says Anthropic told prospective IPO investors that its new chip design team, just fifteen people, already uses Claude for engineering and design work on custom chips. That's reported, not confirmed. Anthropic reportedly also asked chip-design software firms for proprietary design data to train Claude further. Its amended S-1 doesn't mention an in-house chip team. Maya: OpenAI has made a similar claim, right? Theo: Its CFO said in September that OpenAI's own models helped design its Jalapeño chip with Broadcom in nine months. So "our AI builds our hardware" is becoming part of the IPO pitch, ahead of an Anthropic listing reportedly valued near one and a half trillion dollars. Maya: OpenAI made a hire of its own. Jill Hazelbaker, Uber's president, joins in early December as chief corporate affairs officer, reporting to Greg Brockman. The job has been empty for ten months, since Hannah Wong left as communications chief in January, with policy chief Chris Lehane covering in between. Fortune says Hazelbaker turned the job down in the spring. She arrives just as OpenAI reportedly plans a policy and PR push after the midterms. Theo: And data centers were literally on Tuesday's ballot. Maya: At least forty-one data-center measures, by Fortune's count citing Ballotpedia. Ohio alone has about half of them, and eleven would ban or pause construction. Watch Piqua, Ohio, where Meta plans a billion-dollar site, and two Amazon Web Services projects in Sunbury and Adams County. Results were still coming in when we put this together. For scale, one tracker says local opposition disrupted about sixty-eight billion dollars of projects in the second quarter alone. These local votes will show whether the backlash is now stopping real projects. Theo: A developer one now. GitHub Copilot can now be picked as an agent inside JetBrains Air, JetBrains' agent app, using an existing Copilot subscription. Cloud tasks run in a Docker container on a fresh virtual machine and bill to your Copilot plan, not JetBrains credits. Air admins can switch Copilot off, but they can't set Copilot spending limits inside Air. Maya: That's going to make a finance team's day. Theo: Right. Agent apps are turning into storefronts for rival agents, and it's getting harder to see who's billing you for what. Maya: Billing is the last one too. A unanimous Tenth Circuit panel, in an opinion by Judge Carolyn McHugh, upheld about a hundred and fifteen thousand dollars in fees for the law firm Brownstein Hyatt after Flux Drive's failed trade-secrets suit. It rejected the argument that the firm's AI use inflated the bill. Anyone challenging fees has to show actual evidence of overcharging. It's an early appeals-court marker as clients push back on hourly bills for work AI speeds up. Next, we go deep on Europe's top court adviser, who says AI can train on the news for free but that chatbot answers built from it may have to be paid for. Now for the Main Episode Maya: Here's the question. If training an AI on the news is legal, but the answers it builds from live news articles might not be, do chatbot companies in Europe end up paying publishers every time an answer leans on someone's reporting? Theo: And that last part, every time, is the bit that should worry the people building these products. Start with who brought this. Maya: A Hungarian website operator called Like Company. It sued Google Ireland over how Gemini uses its press articles, and the case went up to the Court of Justice of the European Union in Luxembourg, Europe's top court. On November 3rd, Advocate General Maciej Szpunar gave his opinion. An advocate general is an adviser to the judges. The opinion is formal, it's public, and it isn't binding. Theo: So formally it carries no weight at all. Maya: None. But Reuters, which broke this, notes the court follows most of its advisers' opinions. So it isn't a ruling, but it's a strong signal of where the ruling is likely to go. Theo: Why is a Hungarian site taking on Google at all? What's the grievance? Maya: Traffic. Reuters reports that outlets like Like Company have watched visits fall as AI summaries replace the search results that used to send readers to their sites. The answer sits right there on Google's page and nobody clicks through. If the court sides with the publisher, Google could end up paying for that content. Theo: Okay. So what did Szpunar actually say? The headline makes it sound like he split the baby. Maya: He did, very cleanly. Part one, training. He wrote that using copyrighted material to train large language models is, quote, freely permitted under the text and data mining exception. That's the EU copyright carve-out for automated analysis of content. He also said there was no evidence Google had trained Gemini on Like Company's articles in the first place. Theo: So on training, Google wins outright. Maya: Outright. Part two is where it bites. When a chatbot fetches an article from a publisher's site, reuses it to write an answer and shows that answer to a user, he says the publisher's right may be infringed, quote, in certain circumstances. And he named two triggers. Theo: Go. Maya: First, the chatbot gets the article by unlawfully getting around a paywall or some other technical barrier. Second, the answer reuses the article's form of expression in a recognisable way. Plain facts aren't protected. What's protected is the particular way a journalist expressed them. Theo: From the build side, this lines up almost exactly with how these systems work. Training is one pipeline. It happens once and it's done. Answering is retrieval. The bot goes and fetches a page at query time, stuffs it into context, and writes something off it. That's grounding, or RAG. And he's saying the first pipeline is fine and the second one is where the liability sits. Maya: Right, and that's the line I keep seeing drawn, on both sides of the Atlantic now. Theo: We talked about this on our October 30th episode. Wikimedia filed an amicus brief in the New York Times case against OpenAI in the US, and it made the same split. Training is fair use, but grounding answers in fetched articles needs a license. That was an amicus brief from Wikimedia, of all groups. Now a top EU court adviser has landed in nearly the same place. Maya: And the paywall trigger connects to something else very recent. On November 2nd, the New York Times and the Tribune papers sued Brave, alleging that Ask Brave got around their paywalls. Those are allegations, nothing's been decided. But Szpunar is basically saying that in Europe, that exact conduct is one of the two ways you'd lose. Theo: So the fight's moving. It used to be, what did you train on. Now it's, what did your bot just fetch and show this person. Maya: And look at what that does to the money. A training claim is a one-time thing. You argue about the dataset, maybe you settle, maybe you license an archive. A retrieval claim comes up again every time someone asks a question. That's a running cost, not a one-time settlement. Theo: Here's where I think the opinion matters most for builders, though. Over the last week we covered OpenAI, Google, Microsoft and Anthropic agreeing on October 30th to honor a shared robots.txt signal for opting out of training. Everyone treated that as the big peace deal with publishers. Theo: And this opinion says that if you're in the EU, training is the part you didn't really need to make peace over. It's already allowed. The exposure is at answer time. So that whole opt-out framework may be aimed at the wrong pipeline. Maya: Which brings me to the two sides, because they're arguing about exactly that. For the publishers, Angela Mills Wade, who runs the European Publishers Council, the big trade group for European media companies, made a point I think is the strongest in this whole story. She said you have to separate letting Google index your content for search from letting AI answers replace the original reporting. Theo: Those are two different permissions. Maya: Two different permissions, and her complaint is that they've been merged. Publishers depend on Google to reach readers, she said. They shouldn't have to choose between showing up in search and protecting their rights. And Google's control over audiences shouldn't become leverage to get publishers' content for AI. Theo: And Google's response? Maya: A spokesperson said Europe's text and data mining exceptions are vital to the region's competitiveness, and welcomed the opinion's recognition that training on publicly available information is lawful. They said Gemini learns linguistic patterns to produce original outputs rather than copying content. And they said publishers have opt-out controls. Theo: I have to push on that, because most of it answers the question Google won. Learns linguistic patterns describes training. Nobody's arguing about training anymore. The open question is what happens when Gemini pulls up today's article and writes from it. Learning linguistic patterns doesn't describe that. Maya: Agreed, and the opt-out line runs straight into Mills Wade's point. If opting out of the AI means you vanish from the search that sends you readers, that's not much of a choice. I'd also say plainly that I couldn't find a named outside expert defending Google's side of this. The pro-Google case here comes only from Google. Theo: Which is worth noting. It doesn't mean nobody holds that view, just that nobody independent was on record today. Maya: And that's why the schools story we did in the headlines matters here. Same day, November 3rd, Google said Gemini and Google Search will both be off for K-12 student accounts until administrators turn them on, as one package. Lina Khan, the former FTC chair, called it, quote, using its search monopoly to coerce schools into adopting AI. Theo: And that's across a hundred and fifty million Classroom users, students and teachers. Maya: And that number is the point. It's a measure of how much reach Google's defaults carry. Look at the structure. A school that doesn't want the AI loses Search. A publisher that doesn't want its reporting in AI answers worries about losing Search traffic. Very different situations, but in both of them Google's search dominance is the leverage. Theo: I'll accept the parallel, but it's a parallel, not proof. The schools thing is a product decision about a toggle. The court case is about copyright in a specific answer. They don't legally touch. Maya: They don't. But Reuters says this fight between publishers and platforms has already produced complaints to EU and national competition regulators. So copyright isn't the only front. If the court rules on copyright and the competition regulators look at bundling, Google could face both at once. Theo: Okay, now let me get to where I think this gets complicated in practice. That second trigger, recognisable reuse of the form of expression. If I'm an engineer at an answer engine reading this, my first instinct isn't to go sign licensing deals. It's to change the generation step. Maya: Meaning what? Theo: Meaning you tune the system to pull out the facts and rewrite them from scratch. No borrowed sentences, no distinctive phrasing, no structure that mirrors the original. Facts aren't protected, he said that explicitly. So you build a system that keeps the facts and drops the expression. That's an engineering problem, and these teams are good at engineering problems. Maya: I'm not sure it's that easy, and here's why. Nobody knows yet what recognisable means. Is it a close paraphrase? A single quoted sentence? A summary of one article? If a summary of one article counts, then rewriting it more heavily doesn't get you out. The user still gets that one reporter's story, in order, without visiting the site. Theo: Fair, and the brief is honest that this is wide open. But if the court reads it narrowly, closer to copied wording than to a summary, I think the outcome is worse for publishers, not better. You'd have chatbots that absorb the reporting, strip the style, and owe nothing. The traffic still goes away and nobody pays. Maya: That's a real risk, I'll grant it. But I think you're underrating the paywall trigger. That one doesn't depend on wording at all. If your bot got behind a paywall it wasn't supposed to, it doesn't matter how well you paraphrased. And the most valuable reporting is usually behind a paywall. Theo: Sure, but the fix there is also technical. Respect the barrier. Don't fetch what you're not allowed to fetch. Which means the premium reporting just drops out of the answers. That's not a licensing market either. The answers just get worse. Maya: Or the bot company wants that premium content in its answers badly enough to pay for access. That's the part I think you're skipping. If answers without the good stuff are worse than a rival's, someone signs a deal. Theo: Maybe. I think we actually see this differently. You think it pushes toward paying. I think a lot of companies will engineer around it first and pay only where they can't. Maya: Probably both happen, in different places. And honestly, the brief lists three possible endings and nobody knows which wins. Licensing fees, collective deals with groups of publishers, or chatbots linking out more instead of summarizing. Theo: Linking out is the one I'd bet on in the short term. It's cheapest. Show a shorter answer, put the source up front, send the click. It partly solves the traffic complaint without anyone writing a check. Maya: Which might be exactly what a lot of publishers want, to be fair. Mills Wade's complaint is about replacement. A bot that sends readers back isn't replacing anyone. Theo: Does this stop at Gemini? Because Gemini is the defendant, but every answer engine fetches articles. Maya: That's unknown, and I want to be careful. The reasoning is about any chatbot that retrieves and reuses press articles, so it plausibly reaches all of them. But nobody's said that on record. It's an inference, not a finding. Theo: It's a pretty short inference, though. From the engineering side there's nothing special about how Gemini fetches a page versus anyone else's agent. If the logic holds, every retrieval stack serving European users has the same question to answer. Maya: I'd add one more voice, a useful one. Alexander Kruel, who curates the AI links digest Axis of Ordinary, summarized it as the opinion separating training from news reuse. Fine. But he stressed that this is advice to the court and the judgment is still pending. And that's right. People will quote this opinion in pitch decks and policy memos as if it's the law. It isn't yet. Theo: The court could adopt both triggers, narrow them, or throw them out. Maya: All three are on the table, and there's no date. The judgment is expected in the coming months, that's all anyone's said. Theo: And then the US question. American courts are still working through fair use in the OpenAI cases. Do they draw the same training-versus-retrieval line? Maya: Unknown. Wikimedia argued for exactly that line over there, but an amicus brief is just an argument. What I'd say is the argument now has a high-profile European legal opinion behind it. Even without any legal force in the US, that gives the publishers' lawyers something to cite. Theo: So where do you come down? Back to the question. Do chatbot companies in Europe end up paying per answer? Maya: Not per answer, literally. I don't think anyone builds a meter that charges a cent every time Gemini cites a Hungarian news site. But I think this pushes the money toward retrieval. Training deals were always going to be one-time payments for an archive. If the court follows Szpunar, the valuable license becomes ongoing access to live reporting, behind the paywall, for answering. That's a recurring revenue line for publishers, and the main reason is that paywall trigger. Theo: My read is narrower. I think the first response is product and engineering, not payment. Respect paywalls, rewrite harder, link out more. Payment shows up where the content is valuable enough that a worse answer costs you users. And honestly, that might be a small set of publishers. Maya: Which would leave a site like Like Company, a smaller Hungarian operator, possibly winning the legal point and still not getting paid much. Theo: That's my worry. Winning the principle doesn't guarantee small publishers any revenue. Maya: The second takeaway we agree on, I think. The line is now clear. Training is the settled part, at least in Europe. The live question is retrieval. And it's the question that gets asked on every query. Theo: Agreed. And for anyone building an answer product, the practical lesson is to know what your agent fetched, whether it went past a barrier, and how close the output sits to the source. That's suddenly something you need to be able to log and show. Maya: Third, the leverage piece. In both the schools story and this case, the pressure point is Google's reach. That's what Mills Wade means when she says publishers shouldn't have to pick between search and their rights. The schools rollout is the next test. Google says its school controls arrive in the coming months, and the Guided Learning pilot in Search hits select US schools later this month. Theo: So two things on the calendar. That pilot, which shows how tightly Google is welding AI into Search itself. And the court's judgment, which tells us whether recognisable means a copied sentence or a summary of someone's story. Maya: And the answer to our question mostly comes down to that one word. If recognisable means a summary, the bot companies pay. If it means copied wording, they mostly rewrite and don't. Thanks for listening to Super Intelligence News and Report. Please subscribe and share it with your friends.