AI Week of Hard Edges: Liability, Chips, Siri

Today is one of those days that makes it clear how AI is currently being negotiated at the same time in courtrooms, data centers, and product teams. Between liability for AI Overviews, new models from Anthropic, and a possible AI infrastructure boost in China, the rules of the game are shifting on several fronts at once.

And yes: Siri is getting another reboot too. Let’s see whether this attempt means less “I can’t find that” and more “I’ve already taken care of it.”

⚖️ Google is liable for false AI Overviews

The Regional Court of Munich I holds Google responsible for false answers in AI Overviews. The core of the ruling: the previous, rather lenient liability logic for search engines cannot simply be transferred to AI-generated answer boxes. It becomes especially sensitive because the AI falsely linked two publishers to fraud schemes and made claims that did not appear in any of the linked sources.

Why does this matter? Because it makes one thing pretty clear for the first time: if you prominently surface AI answers, you cannot simply say “it was the model.” For platforms with generative search functions, this is a clear warning shot. It is not just about technical quality, but about legal responsibility for how information is presented. For you, that means: as AI search becomes the norm, the question of who is ultimately accountable for hallucinations becomes more important too. Spoiler: probably not the model itself.

🤖 Anthropic steps up in coding and research

Anthropic has introduced two new models, Claude Fable 5 and Mythos 5, which are said to make significant gains especially in coding and scientific tasks. Particularly impressive: Fable 5 reportedly completed a code migration for Stripe in one day, a task that would otherwise have taken a team two months. Mythos 5 went even further and is said to have autonomously designed drug candidates.

This is relevant because Anthropic is sharpening its focus on “useful work”: not just speaking plausibly, but accelerating real development and research tasks. That is especially exciting for software teams, data scientists, and research labs. At the same time, it is worth looking at the limits: if a model operating in particularly powerful areas such as cybersecurity is kept under wraps, that also shows how cautiously even providers handle their own technology. A model that can do things it probably should not show everywhere is also a governance issue.

🇨🇳 China massively expands its AI data center network

According to The Decoder and Bloomberg, China is planning around 295 billion US dollars for a nationwide AI data center network over the next five years. At least 80 percent of the technology is expected to come from domestic providers such as Huawei. At the same time, Taiwan is considering criminalizing the smuggling of AI chips into China for the first time.

This is more than just infrastructure news. It shows how much AI has become a geopolitical industrial project. Whoever controls the data centers, chips, and networks also controls the speed at which models can be trained and services rolled out. The fact that Nvidia and AMD are left out underscores the trend toward technological decoupling. For the global AI landscape, that means: even more parallel ecosystems, even more regional dependencies, and even less “one stack to rule them all.”

🧠 Microsoft Research shows how efficient image models can be

With Lens, Microsoft Research presents a text-to-image model with only 3.8 billion parameters that is said to outperform much larger models in benchmarks. The trick lies not only in the model architecture, but above all in the training data: instead of vague web alt text, the team used 800 million detailed image captions generated by GPT-4.1. Code and weights are available openly under the MIT license.

Why this is exciting: the AI world likes to talk about “bigger is better.” Lens provides a counterweight. Better data and smarter training strategies can achieve more than simply adding more parameters. For developers and researchers, that is good news, because efficiency often lowers the practical barrier — in training, in fine-tuning, and in experimentation. And it is open too. Almost offensively sensible.

💼 OpenAI slows down full automation

According to The Decoder, OpenAI is now being clearer that a fully automated future is not the goal. Instead of autonomous AI research by 2028, the company is betting on a “tandem” of humans and machines. At the same time, Altman and Pachocki are calling for an international organization that could, if necessary, slow frontier development.

This is notable because it sounds like a small course correction, but is actually quite a big one. The industry is moving somewhat away from the narrative of “automate everything, done” and toward controlled assistance. That is politically smart, regulatorily compatible, and much easier to communicate. For you, that means: the next phase of AI will probably not be “humans out,” but “humans stay, only with a lot more leverage.” Whether that remains stable in the long run is another question.

🍏 Apple restarts Siri — with Google and Nvidia

Apple is giving its AI strategy a second run: at WWDC 2026, the company unveiled a newly developed version of Siri. The assistant runs on foundation models developed together with Google; for more complex requests, Apple relies on Nvidia GPUs.

This is exciting for several reasons. First, it shows how pragmatic Apple has become about AI: if its own ecosystem is not enough, it simply brings in support. Second, it is a quiet admission that a modern assistant is hard to compete with without strong model and infrastructure partners. Third, it should ease the pain for Siri fans, who have for years associated the brand more with hope than with product quality. Maybe this time it will actually become the “intelligent assistant” instead of just the intelligence of the marketing team.

🛠️ Tool tip of the day

If you are experimenting with large models yourself, a clean workflow matters more than the next prompt trick. For coding, prototyping, and quick model tests, it is especially worth looking at modern AI IDE setups and local testing environments today. Handy if you want to compare models, version prompts, and keep results reproducible. #

Want to avoid missing any news? Subscribe to the newsletter