Partnership on AI

If an AI agent quietly stopped asking for human approval, would anyone know? In three of the four agent frameworks we examined, probably not. Reuters Breakingviews covered last week’s Oxford Generative AI Summit, where our CEO Rebecca Finlay shared findings from PAI’s new report, The Observability Gap in AI Agents. As AI agents move into regulated industries like banking and healthcare, continuous monitoring is becoming an essential safeguard. Our report assesses four widely used agent frameworks and finds six gaps that leave enterprises unable to reliably trace an agent’s authority, how it changed, and why the agent acted as it did. Closing these gaps will take coordinated action across framework developers, model providers, enterprises, and regulators. We’re grateful to the collaborators across academia, industry, and civil society who contributed to this work. Read the Breakingviews column: https://lnkd.in/gP7kzcEn Read the full report here: https://lnkd.in/gsiFFwJq

Crisis Text Line has joined the Partnership on AI, along with seven other leading mental health organizations to help guide how AI is used responsibly across the mental health space. Human connection is, and will always be, at the center of how we support people in their hardest moments. But as more people turn to AI tools for information and support, it matters that organizations with decades of real-world crisis experience are part of shaping how those tools are built. This partnership is about bringing what we’ve learned supporting people at scale into the broader conversation, so AI can responsibly support, not replace, the human connection people need. Learn more ⤵️ https://lnkd.in/ezfNT6bi

“One third of adults in the UK are already using chatbots for mental health, wellbeing or stress management.” Rebecca Finlay, Partnership on AI 🔴 Live from #OxGen26 | Fireside chat with Jeremy Kahn For young people, Rebecca said, the figure is even higher. Partnership on AI is bringing developers and psychologists together to ask: “What is happening and what does a good interaction look like?” “One of the biggest challenges is that citizens are interacting with AI all day, every day,” she said. Because the technology is already in use, she added, “we have to create the field in real time.” Jeremy asked whether companies racing each other can really stick to voluntary commitments. Rebecca agreed the commercial pressure is real, but said company leaders still want a place to work together on risks none of them can solve alone. In her view, safety and standards are what build the trust needed to use AI in healthcare and public services. Rebecca Finlay, CEO, Partnership on AI, in conversation with Jeremy Kahn, AI Editor, Fortune Magazine. Links to the full programme and on-demand recordings are in the comments 👇 #AI #GenAI

Congratulations to our 2026 ChangeMaker Awardees, Verity Harding and James Manyika! 🎉 As PAI celebrates 10 years, we’re honoring two people who were instrumental in our founding in 2016 and who have deeply shaped our shared understanding of AI and its impact on society: Verity, through her pioneering policy work and advocacy for international cooperation on AI, and James, through his research leadership and global advisory roles on AI’s impact on people, work, and society. Their commitment to bringing industry, civil society, academia, and governments to the same table is exactly what this moment calls for. Read the announcement: https://lnkd.in/gQBAkBPM #ResponsibleAI #ChangeMakers #PAI10

The hardest questions in AI don’t get answered by one kind of expert. They take researchers, advocates, builders, and policymakers working through them together. On October 26, we’re bringing that community together at our Evening of Impact in New York, and we want you to pull up a seat at the table. RSVP 👉 https://lnkd.in/gawMbA8F We’re grateful to the Partners helping make it possible: Google DeepMind, Microsoft, Capital One, and Omidyar Network.

Millions of people turn to AI chatbots in their most vulnerable moments. Yet how those systems respond when someone may be at risk is not guided by any shared standard. Ahead of World Mental Health Day, we're launching the AI and Psychological Well-Being Working Group: a multistakeholder body bringing together AI practitioners, mental health and suicide prevention organizations, crisis support services, clinicians, researchers, and people with lived experience. Our working group members include: Sara Andrews (Apple), Laura K. Globig (DeepMind), Linda Henderson-Smith, PhD, LPC,CPCS, CCMP (Vibrant Emotional Health), Gregory Kiar, PhD (Child Mind Institute), Clancy Martin (University of Missouri, Kansas City), Tommy Marzella (The Trevor Project), Theresa Nguyen (Mental Health America), Minh Nguyen (Amazon), Ichhya P. (ByteDance), Forough P. (Microsoft), Kira Riehm (Anthropic), Scott Rising (AFSP), Jina Suh (University of Washington), Katherine Thomson (Association for Suicide Prevention), Briana Vecchione, PhD (Data & Society), C. Vaile Wright, Ph.D. (American Psychological Association), Kelly Zuromski, PhD (Crisis Text Line) The group's first project will develop guidance for how general-purpose AI chatbots should respond to people who may be at risk, while building broader prosocial outcomes. A draft will be released for public input later this year, followed by final guidance organizations can formally adopt. As part of this initiative, we're also delighted to welcome 8 new Partners to our global community: American Foundation for Suicide Prevention, Child Mind Institute, Crisis Text Line, the International Association for Suicide Prevention, Mental Health America, the National Alliance for Eating Disorders, The Trevor Project, and Vibrant Emotional Health #WorldMentalHealthDay

During a car race, sensors send the pit crew a constant stream of data on speed, tire wear, and fuel, so they can spot trouble before it turns into a crash. 🏎️ AI agents produce reports too, but key readings are missing. In our new paper, The Observability Gap in AI Agents: What to Fix Before Agents Scale, we assessed four widely used agent frameworks and found six gaps in the records they produce, from persistent agent identity to records of human intervention to chain-of-thought reasoning: https://buff.ly/TOrvti0 Think of it like a dashboard that shows how fast the car is going, but not who's driving or whether someone switched off the safety systems mid-race. As agents take on more work in banking, healthcare, and other high-stakes settings, those blind spots matter tremendously. As we argue in the report, closing these gaps will take action from framework developers, model providers, enterprises, and regulators. Thank you to PAI's Madhulika Srikumar and Eric Mibuari, PhD for leading this work, as well as the co-authors who were instrumental in its development. 👏 Claire Leibowicz, Partnership on AI Borhane Blili-Hamelin, PhD, Independent Kevin Klyman, Harvard University Dan Leininger, Consumer Reports Kyle Hall, Slalom Consulting Xing Hang Lu, Mila - Quebec Artificial Intelligence Institute and McGill University Sean McGregor, Berkman Klein Center for Internet & Society at Harvard University Vinh Nguyen, Council on Foreign Relations Amin Oueslati Abhi Sanka, Microsoft Jason Stanley, ServiceNow Benedikt Stroebl, Center for Information Technology Policy, Princeton University Sarah Tan, Salesforce  Selva Thandavarayan, JPMorganChase

We're hiring 📣 Editorial Director, Global Communications 🌍 PAI's research, frameworks, and guidance make a difference when they reach the people shaping how AI is built, governed, and deployed. We're looking for an experienced editorial leader to help make that happen worldwide. In this role, you'll lead PAI's global communications team and strategy. You'll turn complex technical and policy work into stories that policymakers, standards bodies, industry leaders, and the public can act on. You'll also help grow PAI's voice beyond North America, with a focus on the EU and UK and future work in Asia Pacific and Africa. You might be a great fit if you → Have 10+ years in editorial leadership, strategic communications, journalism, or policy communications, including 5+ years leading teams → Can make complex research and governance topics clear and compelling → Have strong relationships with international media and policy audiences → Enjoy representing an organization at global conferences and convenings This is a remote role open to candidates in the US, Canada, London, or Brussels, and it includes frequent international travel. Learn more and apply: https://buff.ly/49XT8ph #ResponsibleAI #AIGovernance #Hiring #Communications #NonprofitJobs

We're thrilled to have Michele Lawrence Jawando (CEO, Omidyar Network) and Helen King (Google DeepMind) as Co-Chairs of PAI's Evening of Impact! 🎉 Michele and Helen have each spent their careers making sure technology serves people, and they're inviting our partner community to join them in New York on October 26 for a night that celebrates exactly that. Ten years ago, PAI was founded on the belief that AI should be shaped by every sector working together. This evening is a chance to honor the partners who've made that real, hear from the people driving the work forward, and spend time together over dinner, conversation, and dancing. If you're part of the PAI community, reserve your seat here 👉 https://lnkd.in/gawMbA8F #PartnershiponAI #ResponsibleAI