The German AI Overviews ruling, Bedrock data sharing, Sutton on discovery

00:00:04 A Munich regional court has ruled that Google's AI search overviews are Google's own content — not just a list of search results — and the company is directly liable for what they say. The Decoder's Matthias Bastian covered the case, which centers on two Munich publishers who found their companies linked to scams, subscription traps, and shady business practices in an AI-generated overview.

00:00:30 The court's reasoning is specific. Google's AI overviews "rewrite and judge results in its own words and according to its own structure," the ruling says. In this case, the overview opened with confident claims like "Yes, [company] is known for dubious business practices," then built out a summary with red flags and tips for users.

00:00:52 None of those connections appeared in any linked source. The court found these were "the defendant's own statements" because only Google can check them — at least by comparing the underlying third-party websites with its own statements based on them. The company argued that users could verify the sources themselves and recognize AI output isn't blind trust.

00:01:17 But the court rejected both points: verifying through further research doesn't regularly exempt you from liability, and the overview stands understandable on its own without consulting the sources. Google also invoked existing German case law that gave search engines limited liability because they merely made third-party content findable, not because they created it.

00:01:41 The Munich court drew a line here: traditional search results point to outside websites, but AI overviews generate independent, new, and substantive statements by evaluating and combining those sources. An Oumi analysis of Google's current Gemini 3-powered AI Overviews found that 56 percent of correct answers couldn't be traced back to the linked sources.

00:02:06 Across millions of daily queries, even a 91 percent accuracy rate still leaves millions of wrong answers every hour — which creates a defamation problem, not for the websites that served as source material, but for Google, which generated false claims about them.

00:02:23 The court also ruled on free speech protection differently for AI-generated content. An AI's opinion is "not the expression of an acquired conviction of the persons expressing it, but the result of an algorithm," the ruling wrote. When weighing privacy against Google's interest in offering the feature, the company had to take a back seat.

00:02:46 Google covers 80 percent of legal costs; the two plaintiffs pay 10 percent each. Whether this reasoning holds on appeal is unclear, and Google hasn't commented publicly. But if it gains traction — especially internationally — it hits not just Google but every provider that paraphrases web content: ChatGPT, Claude, Perplexity.

00:03:08 The court also pointed to what it called a protection gap. Traditional search results already help users sort through information; the AI overview is "by no means absolutely necessary." And if Google were only liable for obvious violations, victims would have no recourse when the AI makes false claims that don't appear in any source — because the actual source websites didn't make those claims.

00:03:35 On a different front, AWS Bedrock now requires that inputs for Mythos and Fable tier models be shared with Anthropic. The change hit the HN thread today with 268 points and 170 comments, and the practical consequence is narrower than the headline. Enterprise customers who are already using Haiku, Sonnet, or Opus on Bedrock won't see any difference.

00:03:59 This policy applies to the new Mythos/Fable tier — which Anthropic treats as a new capability level above Opus — and it means every inference request to those models gets logged and sent to Anthropic's systems. The concern for enterprise customers is structural, not theoretical.

00:04:19 AWS has spent a decade building compliance infrastructure: data residency guarantees, contractual boundaries, audit trails. Adding Anthropic as a second data processor changes the legal posture overnight. Regulated industries — banking, healthcare, government procurement — can't add another vendor to their data flow without regulator sign-off, which is unlikely to come.

00:04:44 There's also a question about what happens to that logged data. Anthropic has committed not to train models on it, but they retain inputs and outputs for two years if flagged by trust and safety classifiers, with classification scores kept for seven years. The consumer privacy page says data retained in training pipelines can be kept de-identified for up to five years when users opt into the model improvement setting.

00:05:14 This policy rollout coincides with Anthropic's IPO preparations. An investor-facing company has different incentives than one that can operate without quarterly reporting pressure on data handling practices. The broader industry pattern is also visible: OpenAI's 5.5-Cyber model requires the same data-sharing requirements, and competitors who are already salivating at consensual data sharing now have a precedent to follow.

00:05:43 Some customers will shift their configurations away from Anthropic entirely — not because the models aren't good, but because the second data processor is a compliance blocker that no amount of capability justifies. Others will negotiate custom deals with large enough leverage.

00:06:02 The several weeks it takes for competitors to implement similar policies will matter less than what happens after: a future where the best models require payment in tokens plus access to user data. The real question here isn't whether Anthropic can be trusted. It's that trust is irrelevant when compliance teams have a binary gate: can you pass our legal review?

00:06:27 If the answer is no, the model's quality doesn't enter the calculation at all.

00:06:33 While everyone's talking about Claude Fable 5 as if it's a new class of capability, Richard Sutton published a long talk today arguing that the architecture underneath it — supervised learning — has a structural limit on what it can discover. Sutton is the person who basically invented modern reinforcement learning.

00:06:54 His argument isn't that current models aren't good enough yet or will never get there. It's that supervised learning, as an algorithm class, cannot produce discovery because it lacks evaluation and selective retention. He opens with a joke about a researcher whose work is evaluated as novel but not good, and good but not novel — and applies it to generative AI.

00:07:19 Systems trained by supervised learning take examples and produce behavior similar to them. They generate text like people, images like artists. The stochasticity in their sampling produces novelty. But that novelty comes from randomness, not from evaluating whether what was generated is actually correct or useful.

00:07:40 The core insight: "It's either based on randomness or it's based on data, but it is never both at the same time." You get novelty through temperature and sampling noise. You get quality through training on good data. But you can't have a system that generates something, evaluates whether it's novel and correct, and keeps it — because evaluation doesn't exist at inference time in supervised learning.

00:08:07 Sutton contrasts this with systems that do discover: AlphaGo move 37, AlphaZero's chess style, GT Sophi racing simulated cars, AlphaFold protein structures, AlphaProof proofs. He points to the ones that actually combine three steps that supervised learning lacks entirely: variation, evaluation, and selective retention.

00:08:29 Not just the variation — every model has stochastic sampling — but a clear objective function that evaluates outcomes in the domain itself. He does note that backpropagation does have one form of variation: random weight initialization. But that's temporary; once training converges, plasticity drops.

00:08:49 His group's continual backprop paper addresses this by periodically reinitializing less-used neurons to small random weights. But this is a band-aid on the architecture, not a fundamental change to the supervised learning loop. For science and mathematics — his stated interest through the Xprize Foundation connection — Sutton says this limitation is devastating.

00:09:13 You need true discovery, not pattern matching that occasionally looks novel because of sampling noise. And for those domains, the evaluation step has to come from a clear objective: moves that lead to checkmate, mathematical proofs that hold, genotypes that reproduce more copies.

00:09:32 The interesting thing about Sutton's argument isn't whether it'll stop companies from building on supervised learning architectures. It doesn't. But it does clarify what Claude Fable 5 — or any Mythos-class model built the same way — is actually doing when it feels like it's discovering something new.

00:09:53 The evaluation is coming from outside the system: from users who see something novel and recognize its value, from benchmarks that provide the objective function for training. The system generates variation. Humans supply evaluation. Selective retention happens in the world of published papers, shipped code, and adopted practices — not inside the model's forward pass.

00:10:18 One practical note before I wrap this up. Joshua Snyder from PostHog gave a talk today on their self-driving products pipeline — turning product signals like rage clicks and error spikes directly into pull requests instead of dashboards that you check later. The piece I want to isolate is about grouping, because it's a lesson that matters for anyone building anything that clusters data from multiple sources.

00:10:45 Snyder's team first tried clustering raw signals using off-the-shelf embedding models. It worked poorly. An error about the checkout and an error about onboarding get grouped together because they share structural similarity — stack trace format, error class names.

00:11:03 A Slack message from a customer saying "the checkout is broken for me" stays in a completely different cluster because it's plain text. The fix: instead of embedding the signals directly, they generate queries from each signal using an LLM — what is this signal about?

00:11:21 — and then match those queries in embedding space. The queries are semantically similar across source types even when the raw data isn't. Snyder's three lessons from building this pipeline: evals matter enormously; off-the-shelf embeddings match structure not meaning, so transform before you embed; and if you just throw an agent at a problem without a clear evaluation threshold, it will try to fix everything.

00:11:48 The last one is worth sitting with. The system has to distinguish between actionable problems, problems that need human input, and noise. Every automated pipeline has this gate — when does the agent act, and when does it hand back? The answer isn't in the model; it's in how you define the threshold for actionability.

00:12:09 The real test comes when a company ships an agent with a loose threshold and watches it cascade through customer data. Until then, you just have to define that boundary before the first prompt hits. — Seln.