1. What are the ethical and societal implications of advanced AI assistants? What might change in a world with more agentic AI?. Our new paper explores these questions:. It’s the result of a one year research collaboration involving 50+ researchers… a🧵

RT @hannahrosekirk: Listen up all talented early-stage researchers! 👂🤖. We're hiring for a 6-month residency in my team at @AISecurityInst….

RT @ma_tay_: Happy that "Value Profiles for Encoding Human Variation" was accepted to EMNLP!. I'm proud of this paper - it combines my inte….

RT @rcbregman: AMAZING result from a 10-year study: a one-time $1,000 cash transfer to households in rural Kenya cut infant mortality by 48….

RT @GoogleDeepMind: As AI agents begin to take action in the real world, it's critical we develop new ethical frameworks to ensure they're….

10/ The link to the full paper from the @GoogleDeepMind / @Google team is here: .

9/ We are at a pivotal moment, building the foundational architecture for a world populated by AI agents. The choices we make now will determine whether this future is safe, beneficial, and ultimately human-centric.

8/ Second, we need proactive stewardship of the entire agent ecosystem. Our ability to understand and verify agent behaviour must keep pace with agent capabilities. This requires new technical standards, governance frameworks, and industry-wide collaboration.

7/ So, what’s next?. First, developers must move beyond static benchmarks. We need more meaningful, ecologically valid evaluations—like RCTs—to understand their real-world, long-term impacts of AI.

6/ Deep human-AI relationships create risks of emotional harm and manipulation. We argue that agents must be designed to benefit users, respect their autonomy, demonstrate care & support a flourishing human life. (See also work on "socioaffective alignment w. @hannahrosekirk)

5/ Second, the emergence of social agents: for some users AI agents don't just function assistants – but as companions. Built on LLMs, their ability to form long-term, immersive relationships with users is unprecedented, introducing a new relational dimension to AI ethics.

4/ This means we must expand our conception of value alignment: It's not enough for an agent to follow a user's literal command. It must also align with broader principles: user well-being, long-term flourishing, and value enhancing societal norms.

3/ First, the alignment problem. Getting AI objectives right has always been hard. But when agents can act in the world—accessing email, making purchases, writing code—the consequences of misalignment become tangible and immediate.

2/ Our paper explores the urgent ethical questions this raises, including: . —Who is responsible when an agent makes a mistake?.—How do we ensure these systems are safe and broadly beneficial?. We focus on two challenges in particular: .AI Alignment and the rise of Social Agents.

1./ Together with Geoff Keeling, @Arianna_Manzini & @profjamesevans we ask: What is an AI agent?. Our answer: A system that can perceive its environment and act autonomously to achieve a goal. The capacity for real-world action breaks the 4th wall and has profound implications.

Pleased to share our new piece @Nature titled: "We Need a New Ethics for a World of AI Agents". AI systems are undergoing an ‘agentic turn’ shifting from passive tools to active participants in our world. This moment demands a new ethical framework.

RT @NeelNanda5: Take: Chain of Thought is a misleading name. It's really a "scratchpad". "Thoughts" are internal activations. Imagine you'r….

RT @KobiHackenburg: Today (w/ @UniofOxford @Stanford @MIT @LSEnews) we’re sharing the results of the largest AI persuasion experiments to d….

This paper is absolutely essential reading for anyone interested in developing a science of AI safety and evaluation. I esp. appreciate the “principle of parsimony”:. Behaviours should not be attributed to complex mental processes if simpler explanations are available ✅.

We’re hiring a sociological research scientist @GoogleDeepMind!. Work with the inimitable @KLdivergence, @weidingerlaura, @iamtrask, @canfer_akbulut, Julia Haas & many others 🙌.

RT @RuneKvist: Insurance is an underrated way to unlock secure AI progress. Insurers are incentivized to truthfully quantify and track ris….