AI persuades better than humans, $400M chips and Anthropic vs. the government

Today’s most important AI news

Tuesday arrives loaded with tensions and revelations pointing in the same direction: AI is no longer a technology evaluated in the lab, but a force that negotiates with governments, redefines global semiconductor infrastructure and outperforms human experts at persuasion tasks. It’s not a day of spectacular launches, but of consequences: the effects of decisions made months ago are starting to materialize in policy, hardware and human behavior.

AI can convince you better than any human expert

A joint study by the University of Oxford, the UK AI Security Institute and Stanford, covered today by Import AI, reaches a conclusion that should keep more than one product manager and executive up at night: “AI systems were consistently more persuasive than expert humans” in controlled experiments. We’re not talking about a small margin or artificial conditions; the researchers call it superpersuasion.

For those building products with conversational AI — sales chatbots, onboarding assistants, negotiation tools — this changes the risk calculus. The same capability that makes an agent close more sales or reduce churn can be used to manipulate financial, political or medical decisions. The difference between a valuable feature and a vector of massive harm is, in many cases, just the intent of whoever deploys the model.

The practical implication is immediate: if your company uses AI for user-facing communications, you need an ethical-use policy before someone outside — a regulator or a journalist — writes it for you. What’s a competitive advantage today can tomorrow become the center of a legislative hearing.

Anthropic faces its most serious conflict with the US government

In April, Anthropic revealed it had built a model called Mythos. What came next was an escalation MIT Technology Review describes as the “latest feud” between the company and the US government, and it points to three vectors of tension worth following closely.

The first is governance: what obligations does a private company have to inform the government when it develops capabilities that could be considered a national risk? Anthropic has historically bet on a collaborative relationship with Washington, but Mythos seems to have crossed some threshold that isn’t yet fully defined in current regulation. The second vector is competitive: while OpenAI and Google have more consolidated corporate structures to navigate political pressure, Anthropic operates as a PBC — public benefit corporation — with a charter that imposes its own safety commitments, which complicates its negotiating position. The third is precedent: how this conflict ends will largely determine what regulatory space remains for the second-generation labs that come afterward.

For executives and entrepreneurs in Latin America who depend on Anthropic’s API or are evaluating Claude for critical use cases, the message is this: the operational stability of the big labs isn’t guaranteed. Diversifying foundation-model providers has stopped being a technical preference and become real risk management.

The $400 million machine that decides who can make AI chips

ASML has just presented its new next-generation extreme ultraviolet lithography machine, and MIT Technology Review describes it with an image worth more than any spec sheet: “the size of a double-decker bus, more than 150 tons of precision aluminum covered in thousands of tubes, colored cables and pressurized tanks”. The price per unit is around $400 million.

Why does this matter to someone who doesn’t make chips? Because ASML is the world’s only supplier of this kind of equipment, and without it you can’t produce the cutting-edge semiconductors that power the data centers where AI models run. Every export restriction the US imposes on China, every TSMC decision on production capacity, every NVIDIA roadmap depends on whether there are enough ASML machines available and who they’re sold to.

For developers and PMs working with cloud GPUs, this translates into a very concrete reality: compute scarcity isn’t just a price problem, it’s a geopolitical problem with a single-supplier physical bottleneck. GPU instance availability cycles on AWS, Azure or GCP are ultimately tied to how many of these machines exist in the world and where they’re installed.

In a nutshell

Today’s three stories share an underlying structure: AI’s power is being contested simultaneously on three fronts — psychological, political and physical. Models already persuade better than humans, governments don’t know exactly what to do about it, and the infrastructure that makes all of the above possible depends on an absurdly concentrated supply chain. Anyone who wants to understand where AI is heading over the next two years shouldn’t look only at model benchmarks, but at these three boards in parallel.

---

Sources used: Import AI 462 — https://jack-clark.net/2026/06/22/import-ai-462-superpersuasion-self-sustaining-ai-paths-to-asi/, MIT Technology Review (Anthropic) — https://www.technologyreview.com/2026/06/22/1139424/three-things-to-watch-amid-anthropics-latest-feud-with-the-government/, MIT Technology Review (ASML) — https://www.technologyreview.com/2026/06/23/1138837/asml-400-million-dollar-machine-powering-future-of-chipmaking/