AI regulation will be some of the most consequential work the government does over the next generation. It's critical this work be consistent across industry and apply a clear, rules-based process. Our statement on new export controls on Mythos/Fable: ari.us/trump-admin-blocks-an…

Unions, kids-safety advocates, and civil liberties groups are raising the alarm about @RepLoriTrahan's proposal to preempt state AI development laws. In the Boston Globe today, @bradrcarson explains why a 3 year freeze of state laws poses a major threat: bostonglobe.com/2026/06/12/o…

Without clear guardrails for government data acquisition, AI will greatly increase the scope and scale of surveillance. Amodei’s policy paper is clear. Close the data broker loophole. darioamodei.com/post/policy-…

My new Op-Ed in The Well News: The opioid epidemic has important lessons for AI regulation, that American society has not fully grasped. A thread 🧵 thewellnews.com/opinions/wha…

Today, Anthropic released its Advanced AI Framework (link in next post). I wanted to offer my take on the document, as a lot of people and several reporters have asked for my assessment. In sum, on the questions that matter most to ARI, Anthropic made the right calls and pushed the debate forward productively. Let me walk through what we see as necessary parts of a frontier AI regulation and compare to the Anthropic plan. Safety plans need to be evaluated by an independent entity, which could be an IVO but could also (more likely in at least the early days) be the government itself. This is in the Anthropic plan. ✅ The review by the independent entity must be more than a compliance exercise that just checks if the test the lab wrote and graded was in fact written and graded; it must be an independent substantive assessment of safety. This is in the Anthropic plan. ✅ The government must have the ability to prevent release of dangerous models. This is in the Anthropic plan. ✅ The government must have the ability to recall models whose dangerous capabilities are revealed only after deployment. This is Anthropic's plan. ✅ And very significantly, the preemption language offered up by Anthropic for consideratoin explicitly limits federal displacement to the specific frontier governance functions Congress has chosen to occupy — catastrophic risk testing, evaluator licensing, closely related reporting. Outside those functions, federal law cannot occupy the field, cannot preempt by implication, and cannot displace state statutory or common-law claims. States keep their authority over civil rights, child safety, labor, consumer protection, and tort law. Ambiguity is resolved in favor of state authority, not federal displacement. ✅ Just as importantly, the framework says that federal compliance does not create immunity, a safe harbor, or a presumption against liability under state law. That one sentence closes a door that opponents of accountability have been trying to open for years. ✅ I'm often asked, "What kind of preemption language would you ever support at ARI?" The language that is in the Anthropic proposal is what we like if preemption is insisted upon: narrow in fact, clear about limits, comprehensive in its instructions to courts about what is still permissible at the state level (basically everything from child safety to labor issues to CSAM screening to training data transparency and more) except the specified regulation of cyber, bio, loss of control for a few models). This is the kind of preemption language that should be in, say, the Trahan-Obernolte bill or any other frontier AI law, if any state laws are to be preempted. I hope Congress takes note. The plan has more about cybersecurity and biosecurity, which I'll comment more on later. But wanted to offer this quick take on how I evaluate the safety provisions and the preemption language.

Chatbots are built to mimic human conversation. For kids, that mimicry can drive unhealthy attachment and validate dangerous behavior. Read ARI's latest Policy Byte, proposing a non-personified default setting on AI chatbots for minors: ari.us/non-personified-chatb…

The new House NDAA misses an important opportunity to adopt strong AI standards and guardrails. As DoW faces challenges on AI talent, wargaming, AI testing standards, and human oversight, ARI is urging Congress to strengthen the bill. Our statement: ari.us/ndaa-ai-provisions-fa…

ARI is hiring a Policy Analyst to help shape ARI's policy agenda on AI's economic and societal impact. → 2 years policy experience → Background in legislative, executive, think tank, or advocacy → Interest in AI policy Apply: ats.rippling.com/americans-f…

Adversaries are racing to build AI-savvy military talent. The Pentagon is losing its own to industry, often before HR processes the application. New ARI Policy Byte: U.S. talent acquisition is the most under-resourced front in the AI race. ari.us/policy-bytes/the-pent…

AI-enabled weapons are compressing the window between machine recommendation and lethal engagement. Below 3 seconds, operators cannot build situational awareness. New ARI policy byte argues for a statutory 5-second floor on lethal engagements. ari.us/policy-bytes/minutes-…

AI is already reshaping the battlefield. Until now, the governance picture has been harder to see. Today, ARI is launching our Global Defense AI Policy Landscape, mapping military AI policy across 45 countries. Take a look at globalaidefense.net

The Google-Pentagon deal is a wake-up call for Congress. Our new white paper calls for a five-second rule to ensure meaningful human control of AI weapons, plus thorough verification of AI systems before any contract is awarded. Axios has the story:

🚨Important step forward from the Trump Admin requiring transparency on data center energy usage. Next step: Congress must direct EIA to take its pilot programs nationwide and standardize data center energy reporting across the country. wired.com/story/the-us-gover…

GSA's proposal mandating that government be able to use AI systems for “any lawful purpose" poses major risks, both to American citizens and to the tech to government pipeline. Industry and civil society are pushing back. ARI's full comment to GSA: ari.us/wp-content/uploads/20…

The White House promised its AI framework would protect kids, artists & communities. On all counts it falls short. States frozen out of AI policy, and kids and creators left vulnerable in the AI era. Read @bradrcarson's op-ed in @techpolicypress: techpolicy.press/how-the-ai-…

Pharmaceuticals, cars, and children's toys all face mandatory safety standards. AI platforms do not. Two landmark court victories against Meta are a start, but Congress still needs to act. New Policy Byte from ARI: ari.us/policy-bytes/the-rece…

1

BERNIE SANDERS: "In a sane world, the leadership of the United States sits down with the leadership in China... to work together so that we don't go over the edge and create a technology which could perhaps destroy humanity."

I mean he quit his comfortable university president job to dual hat running an AI safety PAC and a think tank focused on cyber risks from AI, among other things. Granted, you do go to more happy hours.

We’ve entered the “deliberative democracy” part of the US AI governance arc. The next part is messy, but welcomed. If there was ever an issue to democratically deliberate upon, this is it.