Patrick Butlin

5. 'Higher-order representation in AI' (unfortunately slightly dated already): philosophymindscience.org/index.php/phim…

1. 'Desire in AI': philarchive.org/rec/BUTDIA 2. 'Are any machines conscious today?': philarchive.org/rec/BUTAAM-2 3. 'Testing for consciousness in current AI': philarchive.org/rec/BUTTFC 4. 'Consciousness and AI' encyclopaedia entry: oecs.mit.edu/pub/zf1nbs6d/

Some recent papers:

Many thanks to the editor and reviewers for @TrendsCognSci and especially to my co-authors, including @rgblong @Yoshua_Bengio @birchlse @davidchalmers42 @ConstantAxel @georgejwdeane @EricElmoznino @kanair @MatthiasMichel_ @Liad_Mudrik @meganakpeters @eschwitz and others!

The new paper is here: sciencedirect.com/science/articl…

New paper on AI consciousness! Here we present the theory-derived indicator method for assessing AI systems for consciousness. Link below.

We're thrilled to announce the first Eleos Conference on AI Consciousness and Welfare. Join us Nov 21-23, 2025 in Berkeley, CA for discussions on AI welfare with leading researchers from @nyuniversity, @Google, @AnthropicAI, & more.

New Article: "Principles for Responsible AI Consciousness Research" by Butlin and Lappas jair.org/index.php/jair…

The challenge of AI moral status is part of a broader, global challenge: developing AI responsibly and preparing for its impact on society. I’m especially excited about helping Eleos to work out what success in meeting this challenge looks like—and taking actions to achieve it.

The questions I’ll be working on at Eleos, about the conditions for consciousness and the grounds of moral status, are deeply interesting and important. I’m looking forward to renewing my collaboration with @rgblong and continuing to build the community of AI welfare researchers.

I’m happy to announce that I’m joining @eleosai. At Eleos, I’ll continue my work on the philosophy and science of AI minds and moral status. [1/3]

lesswrong.com/posts/BEZzG56i… @rgblong @OwainEvans_UK @flxbinder

@TheDavidSJ @rgblong @conjurial @EricElmoznino Distinguishing environment from system is a very interesting issue, something I’d like to think more about, but perhaps GWT can just say there has to be two loops.

@patrickbutlin @rgblong @conjurial @EricElmoznino From a computational functionalist perspective, how do we distinguish the “environment” from the “system”? What if the tokens are just being used for an inner monologue before producing the “real” output?

@patrickbutlin @rgblong et al argue that transformers lack the recurrence indicator of phenomenal consciousness required by Global Workspace Theory. However, autoregressive decoder-only transformers are recurrent via the token sequence. What am I missing? arxiv.org/abs/2308.08708

@rgblong @TheDavidSJ @conjurial @EricElmoznino I agree with Rob about modules. But also, I’m not sure the environment a system acts on can be its workspace. Otherwise any system with a causal loop through an environment would selectively “write to” and “read from” a workspace. I think GWT requires an inner loop.

@TheDavidSJ @conjurial @patrickbutlin and as with all relayed issues I’ll page @EricElmoznino 😎

@sergio_fsf Thanks Sergio, I hope you're well!

@rgblong @davidpapineau @patrickbutlin Very good! Patrick is just great

excited to talk about this project with @patrickbutlin this afternoon at the Association for the Scientific Study of Consciousness #ASSC26

Just out: onlinelibrary.wiley.com/doi/full/10.11… Best read together with: hrcak.srce.hr/288525

@tom4everitt @MishaLaskin @MishaLaskin What happens if you train the transformer on learning histories from more than one RL algorithm?

@tom4everitt @MishaLaskin I think so, although a strange one. It seems to learn an RL algorithm during training then use it to select actions. Or maybe it creates a simulacrum which is an agent.

In our new work - Algorithm Distillation - we show that transformers can improve themselves autonomously through trial and error without ever updating their weights. No prompting, no finetuning. A single transformer collects its own data and maximizes rewards on new tasks. 1/N