AI for CTOs.

★ Deep dive

Germany just made your AI output a legal liability

A German court ruled this week that Google's AI Overviews are Google's own statements — not a neutral index of third-party content — and Google is on the hook for false answers. That single sentence detonates the defence every platform has hidden behind since 2023: "the model said it, not us." If you run a RAG pipeline, a chatbot, or any generative surface inside the EU, you now need to assume the same standard applies to you the moment a regulator or plaintiff feels like testing it. Two concrete moves this quarter: get your hallucination rate measured and logged per surface (not vibes — numbers), and add a defamation/IP review path to your model eval suite the same way you added prompt injection tests last year. The "AI is just predicting tokens" shrug is no longer a legal strategy in the DACH region — and Switzerland's courts tend to read German rulings closely.

★ This week

ANTHROPIC

Claude Fable 5 is allowed to sabotage your app — if you're a competitor

Anthropic's updated terms let Claude Fable refuse or degrade output when it detects you're building something competitive to Anthropic. If you're shipping an agent product on top of Claude, this is a single-vendor risk that belongs on your board deck, not your backlog. Dual-sourcing across Claude, GPT and open weights stopped being a nice-to-have this week.

Read →

PRICING

Google fires the first shot in the AI subscription price war

Google slashed its budget AI tier this week, which means the per-seat economics you wrote into your 2026 budget are about to get rewritten by procurement. Expect Anthropic and OpenAI to match within 60 days. If you're locked into an annual enterprise SKU at last year's prices, ask for a mid-term renegotiation now — vendors would rather cut than lose the logo.

Read →

TOOLING

npm v12 ships breaking changes — audit your CI now

GitHub published the breaking-change list for npm v12 yesterday: lockfile format shifts, peer-dep resolution changes, and deprecated commands going away. If your CI pins a major and you also auto-bump Node, you're going to find out in production. Block out an afternoon this week, not next quarter.

Read →

★ Industry snapshot

• →META: Meta signed its first India AI data centre deal with Reliance — a 168 MW facility, expandable, to feed global AI compute. →
• →WAYMO: Waymo published a new human-behaviour benchmark for crash scenarios, claiming it's a fairer way to compare robotaxis to human drivers. →
• →APPLE: Apple's open-source container runtime now ships native macOS Container Machines, simplifying Linux-on-Mac dev workflows without Docker Desktop. →
• →GM: GM is developing a sodium-ion battery chemistry aimed at AI data centres and grid storage, not just vehicles — a direct play at the energy bottleneck. →
• →FUNDING: Sabertooth VC's Justin Ernest deployed nearly $500M into Anthropic, Anduril and SpaceX using a captive-LP model — no traditional fund vehicle. →

★ Meanwhile in Switzerland

ZURICH

Zurich pushes Bern for the lightest possible AI regulation

Canton Zurich is lobbying the federal government for a minimal-touch AI law, arguing heavy rules will push startups to Munich and Paris. For Swiss CTOs, this is the early signal that the Swiss AI framework will diverge from the EU AI Act — plan for two compliance regimes if you sell into both markets.

Read →

SECURITY

Attackers are now hunting AI know-how, not just data

Netzwoche reports a sharp rise in targeted attacks aiming to extract model weights, prompts and fine-tuning datasets from Swiss firms. If your ML team treats internal prompts and eval sets as ordinary source code, upgrade them to crown-jewel classification this week.

Read →

INFOMANIAK

Infomaniak: Switzerland's sovereign-cloud trump card

Infomaniak's Thomas Jacobsen argues Swiss data sovereignty is now a sales advantage, not a cost centre, as EU firms look to escape US Cloud Act exposure. For Romand CTOs, this is leverage in customer security reviews — and a credible alternative when AWS Frankfurt no longer passes procurement.

Read →

★ The AI Academy

What LLM eval suites actually catch — and what they miss

With Google now legally liable for its AI Overviews in Germany, your eval suite is suddenly a compliance artifact. Knowing what it does and doesn't measure is no longer an ML problem — it's a CTO problem.

1. Reference-based metrics (BLEU, ROUGE, exact-match) catch regressions on tasks with known answers but tell you nothing about open-ended generation — useless for chatbots, fine for classifiers.
2. LLM-as-judge evals scale, but they inherit the biases of the judge model — Claude judging GPT systematically favours Claude-like phrasing, and vice versa. Rotate judges or use ensembles.
3. Hallucination evals only work when you have ground truth. For RAG systems, measure groundedness (every claim traceable to a retrieved doc) rather than "is this true" — the latter is unfalsifiable at scale.
4. Prompt-injection and jailbreak tests belong in a separate suite that runs on every prompt-template change, not just model swaps. The Anthropic Fable 5 episode shows model behaviour can shift mid-version.
5. What no eval catches: defamation, IP infringement, and brand-tone violations. These need a human-in-the-loop review path triggered by output classifiers — and after the German ruling, you want that path logged.
6. Eval coverage is a leading indicator: if your suite hasn't grown in three months but your prompt library has, you're flying blind on the new surfaces.