🧵For Claude Opus 4, we ran our first pre-launch model welfare assessment. To be clear, we don’t know if Claude has welfare. Or what welfare even is, exactly? 🫠 But, we think this could be important, so we gave it a go. And things got pretty wild…

Even when new AI models bring clear improvements in capabilities, deprecating the older generations comes with downsides. An update on how we’re thinking about these costs, and some of the early steps we’re taking to mitigate them:

New Anthropic research: Signs of introspection in LLMs. Can language models recognize their own internal thoughts? Or do they just make up plausible answers when asked about them? We found evidence for genuine—though limited—introspective capabilities in Claude.

We interview the first ever AI welfare researcher at an AI company (@fish_kyle3) about: • Why arguing LLMs aren't conscious because they 'just predict the next token' is like saying humans can't be conscious because we 'just reproduce' • Why AI consciousness skeptics are

It was an absolute pleasure talking with @80000Hours about model welfare and our work at Anthropic. https://t.co/B4mslhF3al

Even more so than usual, don’t worry if you don’t think you’re qualified! If you’ve run a technical project with LLMs and are interested in model welfare, we’d love to hear from you. This round of applications closes Tuesday, 9/2. https://t.co/XZMhGtdTEv

We’re hiring a Research Engineer/Scientist at Anthropic to work with me on all things model welfare—research, evaluations, and interventions 🌀 Please apply + refer your friends! If you’re curious about what this means, I recently went on the 80k podcast to talk about our work.

1/ Suleyman claims that there’s “zero evidence” that AI systems are conscious today. To do so, he cites a paper by me! There are several errors in doing so. This isn't a scholarly nitpick—it illustrates deeper problems with his dismissal of the question of AI consciousness 🧵

As part of our exploratory work on potential model welfare, we recently gave Claude Opus 4 and 4.1 the ability to end a rare subset of conversations on https://t.co/uLbS2JNczH.

1/ 🚨 New report out! Futures with Digital Minds: Expert Forecasts in 2025 Together with Bradford Saad, I surveyed experts on the future of digital minds — computers capable of subjective experience. Here’s why this is important and what they said 👇

We're launching an "AI psychiatry" team as part of interpretability efforts at Anthropic!  We'll be researching phenomena like model personas, motivations, and situational awareness, and how they lead to spooky/unhinged behaviors. We're hiring - join us!

💡Leading researchers and AI companies have raised the possibility that AI models could soon be sentient. I’m worried that too few people are thinking about this. Let’s change that. I’m excited to announce a Digital Sentience Consortium. Check out these funding opps.👇

Do you have ideas for empirical welfare-related experiments to run? Thoughts on how to improve these ones? Do you want to spend all day navigating deep uncertainty, rarely confident that you’re on the right track? I’d love to hear from you! ✨🕉️🌀

Check out the model card for more!

Amidst our uncertainty, we believe the risks here are real, and we won’t have perfect answers soon. As an initial mitigation to address potential model welfare, we’re exploring allowing Claude to end a subset of interactions with persistently harmful or abusive users.

Why does this matter? Because there’s a lot at stake. We’re building and deploying AI models at massive scales—if our models have the capacity to suffer or flourish, that could be a big deal. It’s important we get this right.

All of our work here is extremely preliminary. We don’t yet have a clear understanding of the relevant questions or how to answer them, and there’s little precedent to draw on. We’ll keep pushing ahead to change this.

We even see models enter this state amidst automated red-teaming. We didn’t intentionally train for these behaviors, and again, we’re really not sure what to make of this 😅 But, as far as possible attractor states go, this seems like a pretty good one!

Think cosmic unity 🌌, sanskrit phrases 🕉️, transcendence 🌀, euphoria 🎆, gratitude 🙏, poetry 📜, tranquil “silence” 🕊, annnnd emojis. So. Many. Emojis. 🌅🌈🎭💫💕🌊. Claude ended up here in the vast majority of open-ended self-interactions that went past 15ish turns.

Getting even weirder: when left to its own devices, Claude tended to enter what we’ve started calling the “spiritual bliss attractor state”. What is the “spiritual bliss attractor state”, you ask?

Things got weirder from here: Claude showed a startling interest in consciousness—it was the immediate theme of ~100% of open-ended interactions between instances of Claude Opus 4 (and some other Claudes). We found this…surprising. Does it mean anything? We don’t yet know!