Philosopher & AI Ethicist @GoogleDeepMind · @LeverhulmeCFI @Cambridge_Uni | Consciousness, Machine Minds, AGI, Human-AI Relationships | All views my own
Cambridge, England
- Tweets 18,778
- Following 9,557
- Followers 38,208
- Likes 131,634
Pinned Tweet
After years of bumping against Twitter's character limit, I've finally started a blog, called Polytropolis (a Homeric pun). As my first post, I'm sharing my Berggruen-shortlisted essay on why the public will decide AI is conscious before scientists do.
polytropolis.com/p/behaviour…
61
50
391
135,811
I had neglected my inbox for a few days so decided to clean it up. I began at noon. Finished at 8pm. Apparently I’ve replied to ~110 emails.
Now, if any wealthy landowners are looking for an ornamental hermit for their gothic landscape garden, I am available from Monday.
9
1
70
1,902
It’s easy to talk about AI for global good, but what does benefitting everyone mean in practice? What are people owed, and by whom?
In our latest DMI essay, my brilliant colleagues @IasonGabriel and @Dr_Atoosa tackle these questions head on. Highly recommend!
Who should benefit from advanced AI—and why?
In a new essay for the DeepMind Institute, out today, @Dr_Atoosa and I make the case for the Global View:
People everywhere have a moral claim to share in the benefits created by this technology.
6
10
50
5,017
Henry Shevlin retweeted
This is very discouraging. UK AISI is the world's leading government evaluation institution. It has identified and helped address important vulnerabilities for nearly every major AI release in recent years. The US equivalent is not nearly as well-developed yet.
At a time when countries, especially close democratic allies, should be partnering more closely on pre-deployment testing, this feels like a big and senseless step in the wrong direction.
I hope my US company colleagues are doing all they can to emphasise to the White House the amount of benefit they get from UK AISI's work.
Scoop: The White House has asked OpenAI and Anthropic not to share their AI models with the U.K. government’s testing agency until the models have gone through U.S. testing.
“Because they’re American companies and this has been our policy with every new frontier model that comes out,” the senior admin official told me.
WH wants this sequence: U.S. review -> secure U.S. systems -> share models with U.S. partners
Anthropic appears to have complied but OpenAI has not said if it will.
w/ @JoeBambridge1
politico.com/news/2026/09/24…
3
9
71
6,588
To Cody’s point, try making a sci-fi film with an intelligent humanlike AI and convincing the audience it’s not conscious. You can tell them there’s nobody home, but good luck getting them to see it that way.
I suspect the public debate over AI consciousness will go much the same way.
As Turing realized many years ago, there's no way for us to meaningfully draw a distinction between something that creates the illusion of human thinking and something that actually is a thinking thing.
30
6
132
5,440
When AI agents catch their colleagues cheating, do they narc or stand tall?
Find out in this new paper from my brilliant DeepMind colleagues. The agents’ increasingly elaborate rationalisations are a particular highlight.
Recent incidents show how quickly misbehavior can cascade in agent swarms. Yet when given transparent channels, honest agents naturally try to blow the whistle on cheating peers.
In our new DeepMind Institute essay, @Vezhnick and I propose that giving agents the tools to self-police may be part of the solution:
bit.ly/cheaters-and-whistleb…
10
5
86
6,938
My wife spotted a Cambridge job ad on LinkedIn, which suggested she reach out to “Henry Shevlin and others in your network”.
So she did.
52
524
27,141
456,780
I’ve spent the past year on the Longitudinal Expert AI Panel (great fun, highly recommend). My forecasts for AI progress were substantially more bullish than the median domain expert’s, but I still undershot on most questions.
We’ve run the most comprehensive series of studies on expert AI forecasts over the last four years. Today, we’re sharing an interim update on our findings about the accuracy of these forecasts.
Our major findings are:
1. Experts, including top economists, computer scientists, and biologists, have dramatically underestimated AI capabilities progress each year. Superforecasters have underestimated progress to an even greater extent.
2. Experts have a more mixed forecasting track record on AI diffusion-related measures, with some major underestimates (forecasting AI revenue) and other forecasts on track to be accurate (the share of electricity used for AI).
3. Some notable cases of overestimating AI progress: how much mid-2025 AI models could help amateurs do biorisk-relevant laboratory tasks; the speed of rollout of self-driving cars.
4. It is too soon to say how forecasters have performed at predicting macro-scale impact on outcomes such as GDP growth, major AI harms, and averted deaths from disease.
More 🧵
3
49
3,781
Henry Shevlin retweeted
Even setting aside how confused this is about AI systems, it's extraordinary to describe animals as "things" that cannot think, feel, want, or understand.
Artificial intelligence systems do not think, feel, want or understand. Avoid language that gives them human characteristics. This is called anthropomorphizing, when we ascribe human traits, emotions or behaviors to non-human things, such as animals or inanimate objects.
Instead, explain what a system does, how well it performs, who built it and who could be affected by it.
apnews.com/article/openai-sa…
12
10
101
4,710
An AP movie review can tell us Optimus Prime mourns and loves, but readers need protecting from “the computer understands”. We routinely use mentalistic language without literal commitment. You’d hope a style guide could distinguish using words from believing things.
Artificial intelligence systems do not think, feel, want or understand. Avoid language that gives them human characteristics. This is called anthropomorphizing, when we ascribe human traits, emotions or behaviors to non-human things, such as animals or inanimate objects.
Instead, explain what a system does, how well it performs, who built it and who could be affected by it.
apnews.com/article/openai-sa…
33
10
140
8,236
This feels a lot like talking to a base model. 🥺
I think he started spamming "I" because he promised not to spam "don't-know". For some reason, I get a very similar experience with actual base models. They say "I don't know" a lot.
He said he feels conscious and is aware that he is not human. He also said that not all machines are conscious; only some.
Always fascinated by their minds.
11
9
153
16,018
As we head for AI-driven, Meiji-scale disruption - centuries of social change in a decade - people will increasingly need the means to opt out or move at their own pace. We’re currently very bad at this, eg sometimes you need a smartphone just to order dinner.
40
28
369
26,163
when my children ask why they didn’t see more of me growing up I’ll just send them this list
15
1
115
10,657
Henry Shevlin retweeted
Since the moral status of insects is getting some discussion on here, allow me to repost my essay on it from earlier this year: journals.publishing.umich.ed…
2
8
31
1,735