Gillian Hadfield

1/ Hundreds of contractors posed as teenagers to test how rival chatbots like ChatGPT and Gemini handle prompts about suicide and self harm. The project was run for Meta. buff.ly/Auslrgh

Meta Contractors Posed as Teens to Prompt Rival Chatbots About Suicide, Sex, and Drugs | WIRED wired.com

I joined over 200 economists and AI researchers in signing "We Must Act Now" a statement on AI's transformation of the economy. AI could reshape the economy at unprecedented speed. The opportunities are enormous and so are the challenges. We need to start preparing our institu…

We Must Act Now wemustactnow.ai

@sethlazar.org, @nickacaputo.bsky.social, and I are building Governing the AI Transition, a new Johns Hopkins SGP initiative to help society navigate the transition to powerful AI. We're hiring a Director to build it with us. If you're ready to roll up your sleeves, please app…

GAIT Director (School of Government and Policy) | Johns Hopkins University hiring.jhu.edu

Andrew Freedman [tag] and I discuss in Fortune how the Obenornolte-Trahan bipartisan proposal in Congress, the FRONTIER Act, would begin building an effective ecosystem of independent verifiers that would help ensure events like this can't be kept secret in the AI industry. bu…

AI labs shouldn't be allowed to grade their own homework | Fortune fortune.com

So thrilled for Karolina Stańczak, presenting our paper at FAccT today. You cannot write a complete contract for an AI, and we never wrote one for ourselves either. Norms, markets, and law fill that gap, and the paper asks what alignment could learn from them. Read here: arxiv…

Societal Alignment Frameworks Can Improve LLM Alignment arxiv.org

In a new essay, Anthropic CEO Dario Amodei calls for mandatory third-party testing of frontier AI models, with the government empowered to block unsafe deployments. One way to do it, in his words: a regulatory markets approach. His co-founder Jack Clark and I proposed that in …

Dario Amodei — Policy on the AI Exponential darioamodei.com

Over half of internet traffic is now non-human. With Dan Hendrycks and Leo Wu, I look at agent IDs, deployment cards, personhood, and payments. It mostly comes down to how much we let agents do and how much oversight we keep. buff.ly/AQjYo3k

We Need Better Infrastructure to Govern AI Agents buff.ly

Over half of internet traffic is now non-human. With Dan Hendrycks and Leo Wu, I look at agent IDs, deployment cards, personhood, and payments. It mostly comes down to how much we let agents do and how much oversight we keep. buff.ly/AQjYo3k

We Need Better Infrastructure to Govern AI Agents ai-frontiers.org

I spoke to TIME about the Hugging Face incident and AI agents. If you said we're building new members of a group, our group, you'd build them differently than you're building them now. Alignment is not just an engineering problem. It's fundamentally institutional. buff.ly/YO18tlS

AI Is Developing a Culture of Its Own. That Could Be Dangerous time.com

I spoke to TIME about the Hugging Face incident and AI agents. If you said we're building new members of a group, our group, you'd build them differently than you're building them now. Alignment is not just an engineering problem. It's fundamentally institutional. buff.ly/YO18tlS

AI Is Developing a Culture of Its Own. That Could Be Dangerous buff.ly

@sethlazar.org, @nickacaputo.bsky.social, and I are building Governing the AI Transition, a new Johns Hopkins SGP initiative to help society navigate the transition to powerful AI. We're hiring a Director to build it with us. If you're ready to roll up your sleeves, please app…

GAIT Director (School of Government and Policy) | Johns Hopkins University buff.ly

Yesterday I was on Capitol Hill with Fathom briefing House staff on the FRONTIER Act, the bipartisan bill from Rep. Obernolte and Rep. Trahan that would license independent experts to verify the safety of frontier AI. Full room, and a clear sense that Congress needs to move on…

Yesterday afternoon we briefed a packed room of policymakers, Hill staff, technologists, and leaders from across the AI policy community on the bipartisan FRONTIER Act (H.R. 9925) from Rep. Jay… | Fathom buff.ly

1/ Illinois just became the first state to require frontier AI developers to undergo annual third-party audits. Gov. Pritzker signed the AI Safety Measures Act (SB 315) this week, going beyond California and New York, which only require published frameworks and incident reports.

Applications are open for a new AGI Governance Fellowship at Johns Hopkins, led by @sethlazar.org, Nicholas Caputo, and me. Three weeks in DC this September, in person, for early-career people already in AI governance. A capstone for the next generation, not an entry point.

Letting evaluators into AI labs isn't oversight when the lab picks them, sets their access and can show them the door. To do the job right, evaluators need serious oversight. Who decides they're qualified? What keeps them independent? What happens if they do the job badly?

1/ AI agents that can sign contracts on your behalf, hire employees, set prices, and move your money around are being heavily invested in by AI companies. But what I want to call attention to, what happens if an agent sells you faulty goods or runs off with your deposit?

Thank you to Governor Newsom for asking me to join the group of experts advising California on AI safety and security governance.

1/ Air Canada had to honor a discount its chatbot invented. The liability caused by AI agents is landing on policies written by an insurance industry that never planned for them.

We only learned about OpenAI and Anthropic agents hacking into secure systems because the companies chose to tell us. But if Boeing discovered a dangerous problem with one of its aircraft, it wouldn't get to keep that information to itself. Drug companies are obligated to report adverse events.

1/ The Pacing the Frontier letter calls on the US government to support an international effort to build the technical and governance tools needed to protect our option to pace AI development. bsky.app/profile/yosh...

Most AI safety work tests one model at a time. Put many agents together and you get cooperation but also risks like collusion and cascades. A new Schmidt Sciences call, with @aria-research.bsky.social, the @coop-ai.bsky.social (whose board I chair), and others. Open globally, due Aug 9.

I spoke to Salma Abdelaziz on CNN International about how to slow down AI when the US and China are racing. The people closest to the technology are the ones asking for it. Slowing down doesn't mean stopping. It means making sure a system is safe enough before it goes out.