Partnership on AI

Millions of people now turn to AI chatbots during moments of mental health crisis. Policymakers and the public are still debating what that means, but AI companies are making decisions that affect people today. In March, we convened frontier AI companies, mental health clinicians, researchers, and people with lived experience to examine how chatbots currently respond to suicide and self-harm. Our new analysis maps the common intervention types in use across the industry today, and the challenges companies are still working through. Read the full analysis:

AI adoption is accelerating faster than public trust in it. A robust AI assurance ecosystem, the norms, tools, and independent experts that measure and verify whether AI systems are trustworthy, could close that gap. The pieces exist today, but they're fragmented and incomplete. In our report, we map what a functioning assurance ecosystem actually requires and offer 12 concrete recommendations for policymakers, from funding cross-sector collaboration to supporting AI Safety Institutes in assessing frontier systems. This is Part 1 of a three-part series on AI assurance. Read it here: https://buff.ly/JjUSyPL #AITrust #AIPolicy #ResponsibleAI

Even Google DeepMind, with all its resources, says overseeing its own AI agents will take "increasingly costly measures." If that's the case for the most advanced agents inside a frontier lab, consider the banks and hospitals now putting agents into onboarding, compliance, and patient care. Same problem, far less to spend on it. Vinh Nguyen and I have a new piece in Tech Policy Press. Regulators and AI labs agree we need to monitor agents in real time. The infrastructure to actually do that at scale doesn't exist yet. Why it's so hard. It's not "build more dashboards": 🟡 Logging is expensive, and the data is scattered. To really see what an agent is doing, you'd capture every model call, every reasoning step, every tool invocation, every hand-off to another agent. Retaining all of that quickly becomes impractical. And even then, the data doesn't sit in one place. It's spread across actors who don't share it. 🟠 Detection adds its own cost. Catching failures in real time increasingly means using AI to watch AI, which piles on compute and engineering overhead. 🔴 A human still has to review the flags, and that's where it breaks. In cybersecurity, 71% of security operations professionals report burnout. They drown in alerts, most of them false positives, while critical signals get missed. Deciding whether an agent made a defensible call or a discriminatory one takes domain expertise, not just the ability to triage. Our argument ➡️ enterprises don't have to wait for all of this to be solved. The near-term move is to tier monitoring by risk, not by engineering budget. Imagine if a hospital chose which patients to monitor by the price of the sensors instead of how sick the patient is. We must match monitoring to the stakes. The only question is whether enterprises build this now or wait for the failure that forces them to. Read the piece here 💡 https://lnkd.in/gmAxxV96 Partnership on AI Council on Foreign Relations Justin Hendrix #AIagents #AIgovernance #enterpriseAI

We've gotten remarkably good at building AI agents, but we're still figuring out how to monitor them once they're out in the world. New op-ed from PAI's Madhulika Srikumar and Vinh Nguyen (Council on Foreign Relations), out in Tech Policy Press: "We Can't Monitor AI Agents at Scale. Here's What It Will Take." https://lnkd.in/e2iGAsrC Agents are already executing transactions, pulling sensitive data, and making calls that directly affect real people, and yet regulators are assuming a level of real-time human oversight that the infrastructure hasn't caught up to. Madhulika and Vinh argue we've run this playbook before, in cybersecurity, which got ahead of the problem by learning to tier and sample its monitoring. They argue AI agents need that same discipline, plus a shared format for what agent traces should capture. So, will companies build this kind of infrastructure before something goes wrong publicly, or only after?

When the early names look like this, you post them early! An early look at some leading voices invited to join #OxGen26, with many more global voices to be announced over the coming weeks: 🔹 Sir Nigel Shadbolt — Chair of AI@Oxford 🔹 Sir John Lazar — President, Royal Academy of Engineering 🔹 Carl Benedikt Frey — Leading Oxford Economist on AI & Work at the Oxford Internet Institute, University of Oxford 🔹 Daniele Magazzeni — Group Chief AI Officer, UBS 🔹 Rebecca Finlay — CEO, Partnership on AI Two days on the adoption, societal impacts and future of AI — 15–16 October 2026, Jesus College Oxford, University of Oxford. Tickets are now live → oxgensummit.org/register #OxGen26 #AI #Oxford

What does "good AI" actually look like? For over a decade, the AI field has debated frameworks, principles, and pledges. But there's never been a shared, public resource that shows what responsible AI looks like in practice. At last week's UN Global Dialogue on AI Governance in Geneva, our CEO Rebecca Finlay co-lead a panel on safe, secure, and trustworthy AI, and sat down with Geneva Solutions to talk about why measurement, not just intention, is the next frontier for the field. Three themes stood out: 1️⃣ The need for an open, independent evidence base to inform policymakers and the public 2️⃣ Interoperability between AI systems built on shared standards and human rights principles 3️⃣ The need to actually measure whether AI progress is serving the public interest This backdrop is why PAI launched two new initiatives last week: the Global AI Progress Hub, a resource highlighting what responsible AI practices look like in action, and a Measures of Responsible AI Progress report, assessing how well AI systems are delivering for people and communities. As Rebecca put it, the goal is to draw lessons from sectors like civil aviation and pharmaceuticals — fields that took years to build common technical standards, but ultimately made progress safer and faster. The same logic applies to AI. Read the full interview here: https://lnkd.in/egYuxEav Kasmira Jefford #ResponsibleAI #AIGovernance #UNGD

Today "We Must Act Now: A Statement on AI's Transformation of the Economy" was released, and I was glad to be among the initial 200+ economists and AI researchers who signed it. I've been working on understanding and improving AI's impacts on workers since 2019, long enough to know how badly confident predictions in this space can age. In 2019, we were still using the phrase "deep learning." The field (me included!) was broadly convinced that white collar and creative work was safe from the chopping block. Four years later, the creative workers we thought were safe led the US's first strikes over AI: Hollywood writers with WGA and then actors with SAG-AFTRA. A month into that strike, Katya Klinova and I released Partnership on AI's Guidelines for AI & Shared Prosperity, developed with leaders from industry, labor, civil society, and academia (inc. Anton Korinek, Daron Acemoglu, Andrea Dehlendorf, Deborah Greenfield, Lama Nachman, and Dean Carignan), to understand the progress of LLMs behind the scenes. By then it was already clear the prior wave of research on "the future of work" (mine included!) had gotten it very wrong. Now, the publicly visible ripples from ChatGPT's release are building toward a potential tsunami. Underestimating the potential for radical transformation and job displacement – with all the hard-learned humility we should have about confident predictions – would be societal malpractice. So would be repeating other past mistakes: • not investing in public understanding of AI capabilities and impacts • waiting for degraded job quality and job losses before acting • targeting specific groups of workers instead of creating broad supports • leaving decision-making to solely to markets and corporate C-suites instead of bringing in workers and communities from the start Thankfully, we're in better shape than in 2019 and 2023, with bigger and bolder coalitions, greater public awareness, and more data shared by AI companies. Rebecca Finlay and I are supporting the Partnership on AI team (Michael, Eliza, Shachee, Gavriel) tackling these challenges: bringing key stakeholders in to shape that data, and leading multistakeholder scenario planning and recommendations on AI and work. AI capabilities are advancing rapidly, and the range of possible futures is only widening. The statement calls on economists, policymakers, and technology leaders to invest now in understanding AI's economic impacts and building the incentives, guardrails, and institutions to ensure AI complements humans and benefits society. I couldn't agree more. I'd add workers and labor as crucial decision-makers, and underscore policymakers, because these incentives and institutions won't build themselves. That's how we create the AI we want, and the worker and community supports we need for economic flourishing. Full (short!) statement linked in comments, please sign if you agree. And thanks Anton, Erik Brynjolfsson, Tom Cunningham, and Ajay Agrawal for circulating.

One message stayed with me after this week's Partnership on AI Partner Forum in Geneva: responsible AI isn't just about the technology we build—it's about the people it serves. It was an honor to join leaders from across sectors to discuss how AI governance can better support workers, strengthen public trust, advance health, and promote equity. These conversations couldn't be more timely as AI continues to evolve faster than the systems designed to guide it. What gives me hope is that we're moving beyond asking whether AI should be governed and toward asking how we can build it responsibly, together. Thank you to Partnership on AI for convening such a thoughtful dialogue alongside the ITU AI for Good Summit and the inaugural United Nations Global Dialogue on AI Governance. I'm grateful to have shared the stage with such insightful panelists (Brian Tse, Christy Hoffman, Alex Walden) and to contribute Crisis Text Line's perspective on keeping human connection at the center of responsible AI.

This week I joined a panel at the Partnership on AI's 2026 Partner Forum in Geneva, held alongside the ITU AI for Good Summit and the inaugural UN Global Dialogue on AI Governance. Our discussion touched on what responsible AI must deliver: for workers, for health, and above all how it must align with human rights, and who bears responsibility for getting there. UNI's contribution is always to focus on the role of workers' voice when it comes to the use of AI at work. There is no "responsible AI" without workers' at the table, and this remains a critical challenge.  Many thanks to Partnership on AI for convening this room, and to my fellow panelists for the candor. And Congratulations to AI for its reaching its 10 year birthday. Quite a milestone.    Partnership on AI Margaret Meagher Alex Walden UNI Global Union