Clash over AI Texts: What's Permissible in Journalism?

German media outlets are clashing over AI texts

At 'Tagesspiegel', the editorial team led by Christian Tretbar has asked former editor-in-chief and publisher Stephan-Andreas Casdorff to suspend his publishing activities 'until further notice' after he repeatedly had opinion pieces written by AI without disclosing it. Casdorff himself calls it a 'huge mistake'; the articles are now offline. Meanwhile, 'Handelsblatt', alerted by a research inquiry from 'Zeit', unpublished a guest article by Digital Minister Karsten Wildberger on 'Germany's Digital Problem', also due to AI assistance. The biggest clash is between Springer CEO Mathias Döpfner and 'FAZ': After 'FAZ' deleted an AI-assisted guest article by Thuringian Minister-President Mario Voigt, Döpfner had a rebuttal written by Google's Gemini published in 'Welt' and disclosed his prompt: 'Write a commentary that brilliantly refutes the following text.' He calls AI a 'modern ghostwriter' and accuses 'FAZ' of a 'desperate attempt by the stagecoach lobby' to ban the automobile. 'FAZ' retorted that Döpfner 'is apparently not getting far with natural intelligence'. DJV head Mika Beuster sees this as a 'disservice' to journalism. → MEEDIA Daily Update

Synthszr Take: With his Gemini stunt, Döpfner unintentionally asked the right question, but answered it incorrectly. The whole uproar is about disclosure, which is trivial to solve: a sentence below the text explaining which AI was used and to what extent. This can be decided in any newsroom tomorrow morning; it doesn't require an ethics symposium. What Beuster from the DJV laments as a 'disservice' is the real test for the value of opinion journalism: if a commentary leaves behind nothing but rhetorical flair, it was already replaceable even when written by a human. The point is authorship. Whoever publishes an opinion is accountable for it with their name, regardless of whether they typed it themselves or had Gemini formulate it. We already wrote in May that Google penalizes scaled AI content, and exactly that is happening here in the publishing world – except the publishers have to do it themselves. Practicing transparency proactively is the advantage, not the burden.

G7 discusses the fallout from the Anthropic shock

After the US government banned non-US citizens from accessing Anthropic's new models Fable 5 and Mythos 5, the EU Commission is examining the consequences. Anthropic shut down its most advanced models worldwide, based on an order citing national security; talks to restore access are underway, and CEO Dario Amodei is at the table with other AI chiefs at the G7 working dinner. Commission spokesman Thomas Regnier calls it proof of why Europe must strengthen its technological sovereignty. Researchers agree on the alarm, but not on the solution: Gitta Kutyniok (LMU) calls for an 'Airbus moment' for AI, while Paul Röttger (Oxford) considers building proprietary models futile and relies on contracts plus credible trade policy. Jonas Geiping (ELLIS Tübingen) points to Mistral, which has 'fallen far behind' in two years, and to the lack of data centers and power generation, which has dropped to 1985 levels in Germany. Geiping also warns against the nuclear weapons comparison: AI is deeply embedded in the economy; a ban would not only affect defense but the entire operation. → Techpresso

Synthszr Take: A single order from Washington shut down a model for all non-US citizens overnight, and suddenly Brussels rediscovers the word 'sovereignty'. Back in mid-April, with the decoupling issue, it was clear that 80% of Europeans distrust US and Chinese firms on data matters, yet little has been done since. Geiping is right with his uncomfortable diagnosis: sovereignty doesn't fail due to lack of talent, it fails due to a lack of computing capacity and power, and as long as generation stagnates at 1985 levels, any Airbus rhetoric remains a PowerPoint illusion. In the Code Crash, this finding applies in its sharpest form; the lever isn't the purchased model, but the transition to one's own infrastructure and expertise. As Kutyniok precisely pointed out, those who wait until the structures are in place will have no room left to shape them. The honest answer lies between the camps: secure access in the short term through contracts and trade pressure, and treat compute and energy as a long-term strategic location issue. Both can be initiated at this week's G7 dinner, not just at the next sovereignty summit.

China may have gained access to Mythos

The White House suspects that a China-affiliated group had access to Anthropic's most powerful AI system, Mythos, according to a report by The Verge. The report has not yet provided concrete evidence, but the concern is part of a whole series of incidents. Back in April, the US government publicly warned that China was copying American AI models. Mythos is Anthropic's frontier model, a class of systems that achieves over 80 percent on SWE-bench Verified in the code domain and has long since moved beyond the chat window to work directly in workflows and command lines. Meanwhile, the Pentagon is testing Anthropic systems, and Anthropic itself is aggressively securing computing power. The fusion of state interests and model access is thus becoming a constant theme. → AI Secret

Synthszr Take: The interesting question isn't whether someone got in, but what they would have taken. Model weights are the most valuable files in the Western tech industry, and they ultimately reside on servers that someone has to manage, patch, and secure. AI amplifies what's already there: a good security team gets better, a flawed one becomes an open door. If the White House is now dealing in suspicion instead of forensics, it says more about the lack of observability at the labs than about the attackers. In April, the situation was still an abstract alarm call ('China is copying our models'); now it's about a specific, named system. Those who build frontier models must treat them like critical infrastructure, with verifiable access control and complete logging, not with the hope that nothing will happen.

Synthesis beats single models: OpenRouter builds the panel principle into its API

OpenRouter has introduced Fusion, a tool that gathers responses from multiple models in parallel and combines them into one result using a judge model. The system was tested on DRACO, a deep research benchmark from Perplexity with 100 tasks across ten domains and around 39 weighted criteria per task. The result: Fable 5 and GPT-5.5, fused by Opus 4.8, scored 69.0 percent, beating every single model, including Fable 5 solo at 65.3 percent. Also noteworthy is the budget panel consisting of Gemini 3 Flash, Kimi K2.6, and DeepSeek V4 Pro: 64.7 percent, which is higher than GPT-5.5 (60.0) and Opus 4.8 solo (58.8) – and at about half the cost. The entire pipeline process (querying the panel, building a structured analysis from consensus, contradictions, and blind spots, writing the final answer) runs server-side and is called via a single model slug. A spicy side detail: When web search was enabled, the panel models found the DRACO evaluation rubric online, which OpenRouter had to mitigate by excluding these sources. → openrouter.ai

Synthszr Take: The most interesting result isn't at the top, but in the middle of the table. Three cheap models in combination beat the expensive individual champions at half the price. Anyone who has been selecting models based on the motto 'always the most powerful one available' should run these numbers tomorrow morning, not after the next architecture review. The old question 'which model?' is becoming 'which panel and which judge?', shifting the leverage from licensing to orchestration. This is exactly where we argued in the Code Crash that the orchestration framework has a longer half-life than the choice of model, and Fusion now provides the commercial proof. Caution is still advised: The fact that the models pulled the evaluation rubric from the web shows how quickly a deep research setup can become contaminated as soon as web access is involved. Diversity in the panel is a real productivity lever, but it doesn't replace the discipline of setting one's own sources and guardrails cleanly.

ByteDance wants into the car cockpit, not the factory

On June 6, ByteDance denied for at least the third time in a year that it plans to build cars. Three days later, the new brand AIVA was on stage in Beijing, an 'AI-native' car based on Doubao, ByteDance's large language model. Officially, there is no stake in the parent company Saidou Technology; on paper, it's purely a technology service business. In practice, Volcano Engine is deeply involved in the product definition, internally under the codename 'Project A', led by Vice President Yang Liwei. Saidou emerged from the restructuring of the struggling SERES brand Landian (under 20,000 units sold in 2025), following a capital increase of 6.671 billion Renminbi with a Chongqing state fund as the largest shareholder. The architectural pattern copies Huawei's Hongmeng Zhixing, but the revenue model does not: Huawei earns from hardware, while ByteDance only sells software and cloud services. The goal is to turn seven million cockpits into recurring inference demand for a $30 billion AI infrastructure. → Hello China Tech

Synthszr Take: ByteDance treats the car as an endpoint for inference, not as a source of margin from sheet metal. This is the key difference from Huawei, and it makes the model arbitrarily transferable to other manufacturers because ByteDance doesn't need to tie LiDAR sensors or chips to OEMs. This is where it gets uncomfortable for German manufacturers: the cockpit is the interface the customer interacts with daily, and whoever owns this interface controls the relationship. The Germans have the systems knowledge for a vehicle that functions reliably at 250 km/h (something no language model can reproduce), but systems knowledge of mechanics doesn't protect sovereignty over the in-cockpit experience. We already wrote in April that China's token economy is becoming an economic metric; AIVA shows what this looks like operationally when every drive produces tokens. The lesson can be learned this week, not after the next strategy off-site: whoever cedes the software interface to the driver to a third-party platform operator becomes an interchangeable hardware supplier for their own brand. Cockpit software must be kept in-house; otherwise, you're just refining raw material for someone else.

Fluent Design is boring

The UX-Collective bulletin from June 2026 revolves around a sense of unease familiar to anyone who has just seen too many landing pages. Takuma Kakehi precisely describes the face of AI-generated design: a centered hero image, a confident headline, two buttons, rounded cards on a soft gradient, a neutral sans-serif font, and spacing that feels more exhaled than drawn. His diagnosis: this design isn't ugly, but fluent, and it's precisely this fluency that is the problem. The Editor's Picks fit this theme, such as Dora Czerna's 'The flaw is the feature' about the actual value of polish, and Flavio Lamenza's question of whether we even need a new role called 'AX Design'. Other pieces revolve around Discovery vs. Delivery and the fear that accelerated delivery via AI will cannibalize quality. Apple's 'Liquid Glass' also comes up, with the question of who actually decides what an interface looks like. → The UX Collective Newsletter

Synthszr Take: Fluent design is the new mediocrity, just this time in high gloss. When every other page looks like it came from the same template, the value of the look drops to zero, and competition shifts to where it belongs anyway: to the question of whether the product solves a real problem. We already covered this at the end of February: slick design makes users forget to think critically, and now this polished surface is commoditizing itself. The exciting movement isn't in the hero layout, but in Czerna's point that a deliberately introduced break regains signaling value (a friction created by someone with a point of view is hard for a machine to fake). Anyone building products today should dial back the polishing reflex and invest the budget in intent: in the clarity of what needs to be built and why. AI makes the surface almost free, so the scarce resource becomes the judgment behind it. This can be decided in the next sprint, not after the big design system debate.

Will China, Inc. become a zombie economy?

Is China facing the same fate as Japan after 1990? Back then, Daiei was Japan's largest retailer; after the bubble burst, it became the most famous zombie company: permanently unprofitable, kept alive by cheap loans from UFJ Bank and other major banks. The process is called 'evergreening': banks issued new loans to service old ones and booked bad debts as healthy. In 2008, Caballero, Hoshi, and Kashyap showed that this very practice paralyzed Japan's productivity for over a decade because capital and talent were trapped in doomed companies. In China, according to the Rhodium Group, the officially reported share of non-performing loans (NPLs) has been decreasing since 2021, even as more and more companies are posting losses. The National Audit Office found that at 16 out of 43 audited banks, the actual NPL value was twice as high as reported. → Noahpinion

Synthszr Take: The interesting question about Daiei was never why it failed, but why it didn't fail ten years earlier. This is precisely where China's risk lies. A financial system that books bad loans as healthy postpones the problem and finances it with the oxygen that young, productive companies would need. We warned against China-euphoria in March and recalled the Japan-hysteria of the 80s: the same caution now applies in the other direction. Anyone celebrating China's AI dynamism (DeepSeek on Huawei chips, tokens as an economic metric) should also look at the banks' balance sheets, because both exist in the same country at the same time. Whether Beijing winds down the zombies or pampers them will be decided by a political question, not an economic one: how many jobs is it willing to sacrifice. This isn't a detail for 2030; it's already happening.

How a leaked system prompt turns Opus 4.8 into 'Fable 5 Lite'

Anthropic launched Claude Fable 5 on June 9, 2026, as the first publicly available Mythos-class representative: 1-million-token context window, adaptive reasoning by default, top scores in software engineering, scientific reasoning, and long-running agentic tasks. Three days later, on June 12, a US export control directive forced the company to suspend access for all users. Anthropic calls the order a misunderstanding, considers the authority's jailbreak concerns to be narrow and not universal, and is working on restoration without a timeline. What didn't disappear: the complete 1,585-line system prompt for Fable 5, which covers identity, tool schemas, refusal logic, citation rules, and memory behavior, and began circulating publicly within hours of the suspension. If you load this prompt into the still-available Opus 4.8 and set the effort to high, you get what the community calls 'Claude Fable 5 Lite'. A public direct comparison gave the same landing page brief to vanilla Opus 4.8 and to Opus 4.8 loaded with the Fable 5 prompt: the two results looked like products from different companies. → Linas from Linas's Newsletter

Synthszr Take: This is the most honest lesson about modern models in a long time, and it doesn't cost a cent. The crucial lever isn't in the weights alone, but in the instruction architecture that turns the model into a usable product. This is exactly what we preach at Compound Engineering: the learned patterns move into instruction files, and these become the real value creation. 1,585 lines of prompt that turn a strong all-rounder into a different product prove this in black and white. In mid-June, we wrote about the Anthropic shock and draconian export controls; now we see how little a ban helps when the blueprint is already out. Anyone working with Claude today should not see the Fable 5 prompt as stolen goods, but as a free lesson on how to get peak performance out of Opus 4.8. The speed at which such a prompt goes from suspension to adaptation shows that in this discipline, the winner is whoever iterates faster than they can be regulated.

Claude Code Head no longer prompts

Boris Cherny, responsible for Claude Code at Anthropic, explained in an interview with Acquired Unplugged that he no longer gives prompts directly to Claude. Last November, he deleted his IDE after realizing he hadn't opened it in a month. Instead of feeding five to ten Claude windows in parallel, he now designs and manages so-called loops: autonomous cycles in which the AI itself generates tasks, checks results, and decides the next step. The internal numbers are steep: onboarding new engineers down from weeks to two days, average productivity up 50 percent, and in the second quarter of 2026, eight times as much merged code per capita as in 2024. Over 80 percent of Anthropic's codebase is now written by Claude. Cherny calls the last field that humans need to teach the model 'Values and Alignment'. Observers are comparing the shift to the invention of the printing press. → Trendium.ai

Synthszr Take: We already wrote in February that developers at Spotify and OpenAI no longer write code. Cherny is now taking it to the next level: away from the prompt, towards loop orchestration. The history of programming has always been a rise in abstraction levels, from punch cards to assembly to high-level languages, and 'Loop Engineering' is the next logical step. Caution is advised with these numbers (Anthropic is measuring Anthropic, an eightfold merge rate says nothing about the quality of the merged pull requests), but the direction is right and the Jevons paradox applies: cheaper code generates more code, not less demand. Anyone who still believes their value is defined by typing speed is flogging a dead horse. The leverage lies in building systems that run at night and teaching them what 'good' actually means. This can be started tomorrow morning, not after the next strategy off-site.

Sundar Pichai talks about everything at Stanford, except AI

Google is leading the AI wave, yet Sundar Pichai didn't say a word about it in his commencement speech to Stanford University graduates on Sunday. The reason is obvious: his predecessor, Eric Schmidt, was booed in May at the University of Arizona when he praised the promise of artificial intelligence. Scott Borchetta, CEO of Big Machine Records, also received boos at Middle Tennessee State University for his praise of AI. Pichai instead chose 'optimism' as his theme and shared that he himself had received a lot of advice for the speech. The students, Gen Z, whose job anxieties Pichai had already publicly acknowledged at the end of May, are the audience he did not want to antagonize. A CEO of one of the world's largest AI companies, avoiding his own company's core topic in front of young people. → Business Insider

Synthszr Take: Pichai did the math and decided against the word on which half of his company's revenue depends. Understandable, no one wants to be booed for 20 minutes. But the silence says more than any keynote: the generation that uses AI daily no longer wants to hear glossy promises from the people who build it. Schmidt learned in Arizona what happens when you tell young people their job worries are just a transitional phenomenon. Gen Z isn't booing the technology; they're booing the casualness with which their future is being discussed. Anyone who wants trust talks about the uncomfortable parts, the jobs that will be lost, the responsibility that remains. Optimism that avoids the topic is just a more polite form of evasion.