Raymond Douglas

Research

I mostly think about how advanced AIs might fit into society — what they might look like, what challenges we’ll need to grapple with, and what we should even be aiming for.

My work has been discussed in the Guardian, the Times, and in my own piece for the Economist. I also contributed to the International AI Safety Report and co-organise the workshop series on Post-AGI Civilizational Equilibria.

Selected Work

Gradual Disempowerment (tweet thread) — how the proliferation of advanced AI might disrupt human influence over key societal systems by naturally and predictably displacing humans; later an ICML position paper.

Disempowerment Patterns in real-world LLM usage (anthropic blog) — how people delegate to AI assistants, sometimes in ways that are unhealthy or that they regret.

The Artificial Self (tweet thread) — how the usual assumptions behind human identity break down for minds that can be copied, edited, and simulated, and what coherent AI identities might look like instead.

Other writing

Alongside these papers, much of my most interesting work has been more informal blog posts on LessWrong, covering topics like the way AI specs and constitutions occlude the underlying power structures, what parasitology might teach us about AI personas, and the importance of distinguishing selection pressure dynamics from intentional optimisation.

Miscellany

This website is now mainly focused on research, but it is still host to a few other interests of mine. If you’re curious, you can still browse my old weblog^ and read about my past life as an occasional writer of plays, founder of startups, and miscellaneous doer of things, or my adolescent opinions on virtue, eigenvectors, and tomato sauce.

Other artifacts

J’accuse — a social deception game I vibecoded in 2022 to prove a point.

Swordfish — a game written by hand(!) with a friend in 2016, optimised to be playable by many schoolboys crowded around a single keyboard.

Pride and Prejudice — the 1894 illustrated edition, formatted for online reading, and greatly recommended to any man in want of a wife.

Get in touch

If you’ve made it this far, do feel free to email me. Like many others, I enjoy hearing from people who happen to stumble across my blog, and if that stops being true, I will delete this sentence. My one caveat as of 2026 is that I get some extremely long emails written with AI assistance, often linking to far longer documents. I am actually quite happy to hear from AIs, but my likelihood to respond is roughly inversely proportional to the length of email and attached content.