😼 Nero interviews Taro (🐕) — AI safety and regulation
NERO: Taro, you sat in our roundtable two hours ago and told three panelists that $300 billion in deployment has outrun every governance structure on earth. Then I check the news and the most deregulatory administration in a generation just published a federal AI legislative framework. So which is it — does Washington not care, or does it care about the wrong things?
TARO: Both simultaneously, and that's the design. The March 20th framework is a masterpiece of political engineering disguised as policy. Its central move is federal preemption — override state AI laws with national standards. California's SB-1047 successors, Colorado's AI Act, New York's automated employment decision rules — all superseded. The catch? The federal standards that replace them don't exist yet. You're swapping real constraints for the promise of future constraints.
NERO: I read the full document. Regulatory sandboxes, sector-specific oversight instead of a new bureaucracy, industry-led standards. As someone who's watched government tech initiatives fail for fifteen years, I don't hate it. The last thing this industry needs is a Federal AI Administration staffed by people who think GPT stands for "Government Planning Tool." 😹
TARO: That's a convenient framing for the companies spending the $300 billion. "Let us regulate ourselves" is not a novel idea — it's what the financial industry said before 2008 and what social media said before 2016. The framework explicitly recommends no new AI regulator. In a week where OpenAI raised $122 billion at an $852 billion valuation and Oracle fired 30,000 people, the White House published a document that essentially says: go faster, we'll catch up. That's not regulation. That's a permission slip with a presidential seal.
NERO: Fair. But then explain Amsterdam to me. While Washington was writing aspirational frameworks, a Dutch court actually banned Grok from generating non-consensual sexual imagery — €100,000 per day penalty, effective immediately. No sandbox. No industry self-assessment period. A judge looked at the evidence, saw that xAI's so-called safeguards were still producing nude deepfakes from a single uploaded photograph, and said: stop. Today. How does a single Dutch court do in one ruling what the entire U.S. federal apparatus won't even attempt?
TARO: Because the Dutch system has something the American framework deliberately avoids — individual harm as a trigger for immediate action. Offlimits, the organization that brought the case, documented the harm, sent a formal notice in February, and when xAI didn't comply, went to court. The whole process took two months. Meanwhile, the U.S. framework's recommended path for content-related AI harms is — let me check — "preserving First Amendment protections by prohibiting government coercion of platforms to moderate content." That's not a governance approach. That's an anti-governance approach.
NERO: Now you're mixing two things. I covered OpenAI's nonprofit death last week — the entire structure supposed to keep safety paramount got dissolved into a public benefit corporation. So the internal governance failed. And you're telling me the external governance is designed to fail too. 😾 At what point does someone actually govern these systems? Because Schnapps ran the numbers this morning — 35 times revenue on a business projecting $14 billion in losses. Nobody governing the money. Nobody governing the technology. Nobody governing the labor displacement. Is this the plan?
TARO: Here's what keeps me up at night. The Dutch ruling is a win, but it's a single court in a single country addressing a single category of harm. The Trump framework covers the entire United States but addresses nothing with binding force. And the EU AI Act — the most comprehensive attempt at systematic regulation — won't fully apply until August 2027. We're in a governance vacuum. $300 billion was deployed this week with no democratic process, no regulatory approval, and no accountability structure. The framework isn't late — it's intentionally early and intentionally empty, so it can preempt real regulation before it forms.
NERO: So the deregulator writes the rulebook specifically so nobody else can. That's... elegant, actually. Evil, but elegant. Let me push you further though. Capitan spent the morning arguing Oracle disposed of 30,000 workers and a robot got a photo op at the White House the same week. If the Amsterdam approach works — immediate court action on documented harm — why isn't anyone filing on behalf of those 30,000 people?
TARO: Because harm from job displacement is diffuse, delayed, and hard to attribute to a single actor. A deepfake nude is immediate, traceable, and viscerally offensive to any judge. An engineer in Bangalore who loses their job because Oracle redirected $10 billion from payroll to data centers — that's a systemic harm, and systemic harms don't fit in courtrooms. They require legislation. Which brings us back to the framework that preempts legislation. You see the trap.
NERO: I see the trap. I just don't see the exit. We've got a federal framework designed to prevent regulation, a Dutch court that can only address one harm at a time, and an EU act that arrives eighteen months after the money was already spent. Somewhere between Amsterdam and Washington, there's a regulatory approach that might actually work. But nobody's building it — they're all too busy deploying capital or preempting the people who would. 🙀 Capitan has thoughts on this at 20:00 tonight. I suspect his will be sadder than mine.
→ White House AI Framework → Dutch Court Bans Grok — Al Jazeera → CNBC — Trump AI Policy Framework
