Providers Held Liable for AI Responses

Munich Regional Court: Google is Liable for AI Overviews

The Munich Regional Court has ruled in a preliminary injunction that Google is directly liable for false statements made by its AI Overviews. The AI-generated summaries had falsely linked two Munich-based publishers to subscription traps, fraud, and “dubious business practices,” without any basis in the linked sources. The court classifies these texts as Google's “own statements” because the system produces “independent, new, and substantial statements” in Google's own words. This removes the old search engine bonus, which only referred to external sites and therefore had limited liability. The court rejected Google's defense that users could check the sources themselves: an analysis shows that only 1 percent of users ever click on a source. A NYT analysis found that the Gemini 3-based Overviews are 91 percent correct, but over half of the correct answers were not supported by the cited sources. Google bears 80 percent of the costs and can appeal. → Techpresso

Synthszr Take: The judge understood what Google's lawyers refuse to admit. A classic search displays external links and isn't liable for their content. An AI Overview produces its own text, making Google the author, not the intermediary. That's the whole legal trick: whoever writes is liable, and an LLM that weaves sources into new statements is, in fact, writing. Google itself admits that the Overviews are “occasionally” wrong, which, with billions of queries per day, leads to a significant number of fabricated fraud accusations. At the end of May, we wrote about the DuckDuckGo boom—30 percent more downloads due to dissatisfaction with AI search—and this ruling is the legal extension of that same trust issue. If you build an answer engine, you take responsibility for the answers, and that responsibility can no longer be passed on to the source page.

Anthropic Calls for a Legal Kill Switch for AI

Anthropic has published a regulatory proposal that would give governments the legal authority to block or prevent the use of dangerous AI models. The justification comes with concrete numbers: a few years ago, models could barely write code; this year, Claude Mythos Preview found thousands of highly critical vulnerabilities in every major operating system and browser. The rules are intended to apply only to models trained with more than 10²⁵ FLOPs and developed by companies with over $500 million in AI revenue or more than $1 billion in AI research spending. Four catastrophic risks are addressed: bioweapons, cyberattacks on critical infrastructure, loss of control, and automated AI research that amplifies the other three. Anthropic demands transparency, independent evaluation, robust security programs, and civil penalties tied to global annual revenue that escalate with repeated offenses. A federal US regulation should only preempt state laws if it is at least as strict as the respective state proposal. The second part of the framework focuses on societal resilience: gene synthesis screening, biosurveillance, hardening internet software, and replacing legacy systems in critical infrastructure. → www.anthropic.com

Synthszr Take: Amodei's essay reads cleverly, and in part it is, but the business logic behind it is transparent. Anyone calling for a licensing regime for frontier models is building a regulatory moat where only the largest providers can survive. The timing is clumsy: in the very week that Mythos itself blocks code reviews and life science questions, 'safety' sounds to many developers like a product that is simply unusable. Willccbb openly states that Mythos is the first publicly available model he is explicitly not allowed to use for his work on open model research. With this, Anthropic ties capability and alignment research to its own control, which drives customers back to Opus or directly to Google, whose DiffusionGemma was released this week under Apache 2.0 with over 1000 tokens per second. A regulation that only protects the market leader is not a guardrail; it's a tollbooth. Trust is built through a functioning model, not a policy paper that thins out the competition.

OpenAI Expects Price War with Anthropic

OpenAI is considering significantly cutting prices for tokens, the billing unit AI companies use to charge for their products. According to WSJ sources, this is in anticipation of similar moves by Anthropic. CEO Sam Altman recently called costs “a huge problem” and promised to offer customers “more value for less spend.” Both companies are already losing billions due to enormous compute costs, and drastic cuts would further erode their margins. Anthropic’s revenue skyrocketed after the viral success of Claude Code, surpassing OpenAI's valuation for the first time. At Uber, the agent budget for 2026 is already maxed out, and Silicon Valley is now debating “tokenmaxxing”—the consumption of as many tokens as possible without a clear return. OpenAI confidentially filed for an IPO this week, with Altman aiming for the stock market debut “within the next year.” → www.wsj.com

Synthszr Take: The sore spot is mentioned in the article itself, almost in passing: the interchangeability of the products. When a customer can swap Codex and Claude Code with two clicks, there's no moat, only a price tag. This is exactly what Zuckerberg has been demonstrating for two years with Llama, commodifying the model and letting others' revenue streams dry up. Now, OpenAI and Anthropic are commodifying each other, voluntarily, right before their IPOs. In mid-May, Anthropic overtook them with enterprise customers; in early June, with a valuation of 965 billion; and now, both are discussing cutting the very margin that supports these figures. Uber's “tokenmaxxing” confession is the most honest signal: companies are realizing that high consumption doesn't mean high value and are tightening their budgets. Anyone trying to place an IPO here is selling investors a growth story whose fuel they are currently making cheaper themselves.

Microsoft Blocks Claude Mythos Internally

Anthropic released Claude Mythos 5 yesterday, the first model in the new Mythos class, and Microsoft immediately took it away from its own employees. While customers of GitHub Copilot and Foundry received the model right away, Mythos 5 does not appear in the internal model selection for Microsoft staff. The reason lies in Anthropic's new data storage requirements: Mythos 5 needs data retention to operate its new safety classifiers, meaning Anthropic retains prompts and outputs, deleting them only after 30 days. In cases of violations of the usage policy, individual prompts can even be stored for up to two years. All other Claude models remain available internally because they run under zero data retention. Microsoft's legal department is now reviewing whether the model can be approved for confidential customer and company data at all. Mythos 5 arrives just weeks after Anthropic described the Mythos family as too dangerous for open release for cybersecurity tasks. → www.theverge.com

Synthszr Take: Here, two of Anthropic's values collide, and both originate from the same company. Safety demands that the model reads prompts and retains them for 30 days; otherwise, the classifiers won't work. This very retention makes the model unusable for data privacy, at least where confidential customer data is processed. The fact that Microsoft, Anthropic's largest distribution partner, first rolls out the model to customers and then bans it for its own staff says more about compliance reality than any roadmap slide. In every vendor matrix, compliance and sovereignty are listed as their own dimensions, and zero data retention is precisely the knockout point where a technically strong model fails. Anyone purchasing models needs an adapter at every layer and a migration path in their risk register, so a single retention clause doesn't paralyze the entire workflow. Model agnosticism isn't an ideology; it's insurance against exactly these kinds of Thursdays.

Xiaomi Hits 1,000 Tokens per Second on Off-the-Shelf Hardware

Xiaomi's MiMo team reports achieving over 1,000 tokens per second in decoding with MiMo-V2.5-Pro-UltraSpeed on a 1-trillion-parameter MoE model, using a standard 8-GPU server. No Cerebras-wafer, no Groq-special chip. The trick lies in the interplay between the model and the system: a co-design named TileRT. Specifically, the stack combines FP4/MXFP4 quantization only for the MoE experts with quantization-aware training, while the non-expert modules remain in higher precision. This is complemented by DFlash, a block-level masked speculative decoding that further accelerates generation. The claim is still subject to independent verification, but the direction is clear. → AINews

Synthszr Take: Specialized hardware has been the answer to how to get 1T models running fast. Xiaomi shows that clever co-design work can achieve the same effect from standard GPUs, and that changes the calculation for anyone thinking about inference costs. We wrote about Xiaomi's price cuts of up to 99 percent at the end of May, and this is the technical justification for it: whoever squeezes more tokens out of the same machine can drive the price per token to the floor without losing margin. This is the Jevons paradox in real-time; cheaper inference creates more demand, not less. For German companies calculating their self-hosted workloads, this means: the assumption of expensive specialized hardware needs to be re-evaluated, because efficiency gains are now coming from the software. Anyone planning their inference stack around hardware moats today is planning past the market.

Gemini 3.5 Live Translate Becomes a Live Interpreter

Google released Gemini 3.5 Live Translate on Tuesday, an audio model that continuously translates speech across more than 70 languages without waiting for the end of a sentence. The output lags a few seconds behind the speaker and preserves intonation, pace, and tone. In Google Meet, the feature jumps from five English-paired languages to over 2,000 combinations in a single meeting, with the private preview for Workspace customers opening this month. The feature is free for consumers in the Translate app and will be embedded into paid Workspace tiers. AI interpretation providers charge $8 to $35 per participant hour for the same service; the market for language solutions was worth $30.85 billion in 2025, according to Slator. On the same day, Google's own Model Card acknowledged that voices can switch gender mid-session and that speech recognition struggles with non-native accents. Grab is already testing the technology for over 10 million monthly voice calls between drivers and riders. → Marcus Schuler

Synthszr Take: Here, a 30-billion-dollar market is becoming an interchangeable feature in a software service you're already paying for. Anyone billing per participant hour today should look at the numbers: $35 versus zero. This is the classic bundling leverage Google uses to commodify entire categories, and it works because translation is no longer a product but a touchpoint in the meeting flow. What's interesting is the gap Google itself documents: voices switch gender, accents confuse the recognition, and the very reliability data that would tell a buyer when the free feature is 'good enough' is missing. For non-critical pickup calls at Grab, it's obviously sufficient; for a legal proceeding, probably not yet. Professional interpretation providers will therefore initially survive in the edge cases where an error is costly. Those in that space should now shift their value proposition to liability and precision, because the underlying mass market just disappeared.

AI Notetaker: Everything is Being Recorded Now

David Haber argues in a guest post for a16z that most work conversations are now recorded by default, without much discussion. His core idea: you have to onboard AI like a new employee—through meetings, where culture, expectations, and edge cases truly emerge, not through the CRM or the wiki. OpenAI already operates this way, with agents stepping in for absent executives in meetings, and Bridgewater established this principle as a policy years ago. Granola is the best example: the tool knows a16z better than almost any other because it was 'in the room.' Haber sees a new software category emerging, organized around speech instead of text, because the most valuable context lives in conversations, and LLMs make this unstructured voice data searchable. Verbal cultures like Shopify and OpenAI benefit disproportionately, while written cultures like Stripe and Anthropic already capture their context. His prediction: in six months, the default will flip from 'don't record' to 'assume it's being recorded.' → a16z

Synthszr Take: The interesting point isn't the shudder of data privacy concerns, but the onboarding argument. When you onboard a person, you send them to meetings, not the wiki, and the same now applies to every agent. Haber is right that verbal cultures have let their most valuable context evaporate for years because what was truly decided in a product review never ended up anywhere. What we noted in mid-February as 'context is king' gets its operational mechanic here: speech becomes the system of record. The uncomfortable part is the top-down aspect he mentions almost in passing, as recording is also a control instrument for leadership, and un-shipping is more expensive than shipping. The most honest stance on this: treat every meeting like an email that could end up in court, and consciously decide which spaces are explicitly not recorded. This guardrail can be established this week, not after the next governance workshop.

Can Tech Giants Learn to Love Cheap AI Models?

The AI boom was built on a simple assumption: bigger models are more powerful, and the most powerful ones win. Now, that assumption is crumbling as the token bills come due and investor subsidies dry up. Coinbase co-founder Brian Armstrong predicts that within 12 to 18 months, 80 percent of all workloads will run on models that are 99 percent cheaper, while only 20 percent will require the latest generation. Initial tests support this: legal AI Harvey, together with inference platform Fireworks AI, cut its costs threefold without losing quality by combining Claude Opus with Fireworks' GLM 5.1, using Opus only for the most demanding tasks. The real dividing line isn't between proprietary and open models, but between large and small ones. If Armstrong's prediction comes true, it would shift the economics of the entire industry, with a large portion of the savings coming directly from the pockets of OpenAI and Anthropic, right as they head for their IPOs. → Techpresso

Synthszr Take: This is the same thing we described at the end of May with exploding token costs, only now the chain reaction is complete. As long as investors subsidized inference, there was no reason to choose anything but the most powerful model. That wasn't discipline; it was convenience on someone else's dime. The Harvey finding is crucial: if the system is built correctly, the small model handles the bulk of the work, and the expensive frontier model is only triggered for cases that truly require judgment. This division of labor is the skill that matters—knowing which query a 99-percent-cheaper model can handle and which it can't. The Jevons paradox softens the blow for the labs somewhat, as cheaper intelligence sparks more demand; but anyone who trains a frontier model for hundreds of millions and then watches 80 percent of the workloads migrate away has to redo their math. From now on, compute discipline is a competitive advantage, and it can be tested in the next sprint, not after the IPO.

AI Coding is Booming, but Governance Lags Behind

Almost every development team now uses AI assistants for programming, but only a minority has regulated how. This is shown in a survey of 831 software engineers and DevOps professionals conducted by UserEvidence for Black Duck in March 2026: 97 percent actively use the tools, but only 30 percent have a fully regulated oversight process in place. GitHub Copilot leads with 83 percent, followed by Claude Code at 63 percent, with most teams using multiple assistants in parallel. The benefit is measurable: 92 percent report faster releases, averaging eight hours saved per week per developer. The catch comes later. Nine out of ten teams encounter problems with AI-generated code somewhere in their workflow, with friction arising from manual review (52 percent), security testing (51 percent), and rework (48 percent). Where governance is fully established, 90 percent report a major efficiency gain, compared to 44 percent for teams without clear rules. → AI Secret

Synthszr Take: These numbers confirm exactly what was described in Code Crash: the models are no longer the problem; the toolchain around them is what matters. 97 percent adoption and 30 percent governance—that's the gap where the promised productivity evaporates. The breakdown is fascinating: teams with full governance see a 90 percent efficiency gain, while those without see 44 percent. So, anyone who thinks rules slow things down has read the study backward. Integrating security early in the pipeline, running SAST scans on every commit, and requiring human review for authentication and payments—that costs one day of setup versus months of damage after an incident. The fact that 86 percent want an AI agent for code review, yet 84 percent still want a human in the loop, shows the right attitude: directing, not just waving things through. Those who set up guardrails now will actually bring those eight hours per week home, instead of losing them again to QA and AppSec.