table 2 from https://arxiv.org/pdf/2507.19960

table 3 from https://arxiv.org/pdf/2507.19960

table 4 from https://arxiv.org/pdf/2507.19960

I split AI into 3 non-mutually exclusive types (see Table 1 above): displacement (harmful), enhancement (beneficial), and/or replacement (neutral) of human cognitive labour. More later possibly, but see Tables 2 to 4 (attached or here: arxiv.org/pdf/2507.19960) for the worked through examples. 2/n

Boast post: It took a lot of time, but this paper is in the AER.

This Is How Much Anthropic and Cursor Spend On Amazon Web Services So, I originally planned for this to be on my premium newsletter, but decided it was better to publish on my free one so that you could all enjoy it. If you liked it, please consider subscribing to su...

Exclusive: Anthropic spent $2.66 billion on Amazon Web Services in the first three quarters of 2025, around 100% of their estimated revenue. Its costs appear to increase with their revenue, showing little path to profitability.
www.wheresyoured.at/costs/

Exclusive: Here's How Much Anthropic Spends on AWS

This newsletter pairs well with the standalone episode of Better Offline I recorded. Both communicate much the same story, but the podcast is a little more entertaining. This is the culmination of months of reporting and investigation.

open.spotify.com/episode/5Tyq...
linktr.ee/betteroffline

Cover page of Guest, O., & van Rooij, I. (2025, October 4). Critical Artificial Intelligence Literacy for Psychologists. https://doi.org/10.31234/osf.io/dkrgj_v1

Table 1 Guest, O., & van Rooij, I. (2025, October 4). Critical Artificial Intelligence Literacy for Psychologists. https://doi.org/10.31234/osf.io/dkrgj_v1

Table 2 Guest, O., & van Rooij, I. (2025, October 4). Critical Artificial Intelligence Literacy for Psychologists. https://doi.org/10.31234/osf.io/dkrgj_v1

New preprint 🌟 Psychology is core to cognitive science, and so it is vital we preserve it from harmful frames. @irisvanrooij.bsky.social & I use our psych and computer science expertise to analyse and craft:

Critical Artificial Intelligence Literacy for Psychologists. doi.org/10.31234/osf...

🧵 1/

One side, "AI 2027" is led by a bunch of privileged white people working at the companies causing the other side (AI con/Empire of AI).

One side, the "AI is gonna be so powerful if you don't let US be the AGI builders" is making money from stealing data, killing the environment & exploiting labor.

These are all flagrant lies, unsourced yet printed in a newspaper that allegedly fact checks things. The rest of it is meandering panic-bait that treats jail breaking LLMs. They don’t even describe the virus! They just said someone made one. Bullshit!

OpenAI is trying to clamp down on ‘bias’ in ChatGPT GPT-5 is better at resisting liberal ‘pressure,’ the company says.

The administration “is currently pressuring OpenAI and other AI companies to make their models more conservative-friendly.”

How LLMs think differently What a historical currency question revealed about how AI models interpret ambiguity, frame problems, and shape our understanding of the past.

interesting little experiment by @squarepegmind.bsky.social

"Comparing AI models isn’t just about what they say—it’s about how they say it. (...) Some aim to solve the problem, and others focus on explaining why the problem is complicated."

thinkingaboutbehavior.substack.com/p/how-llms-t...

Boy do I have the right measurement instrument for you.

Regardless of whether European institutions are willing and able to withstand the pressure, it's becoming clear the impact will be felt here, too - in this domain as well as many, many others.

Less than a week ago I wrote about the new US gov meddling in scientific research - I imagine many ppl thought "Unfortunate, but at least we're safe here in Europe"...

What happens to science when words like "bias" get banned from research? New U.S. restrictions on research language threaten to reshape global scientific progress - including behavioral science

𝗪𝗵𝗮𝘁 𝗵𝗮𝗽𝗽𝗲𝗻𝘀 𝘁𝗼 𝘀𝗰𝗶𝗲𝗻𝗰𝗲 𝘄𝗵𝗲𝗻 𝘄𝗼𝗿𝗱𝘀 𝗹𝗶𝗸𝗲 "𝗯𝗶𝗮𝘀," "𝘀𝘁𝗮𝘁𝘂𝘀," 𝗮𝗻𝗱 "𝘀𝗼𝗰𝗶𝗼𝗲𝗰𝗼𝗻𝗼𝗺𝗶𝗰" 𝗴𝗲𝘁 𝗯𝗮𝗻𝗻𝗲𝗱 𝗳𝗿𝗼𝗺 𝗿𝗲𝘀𝗲𝗮𝗿𝗰𝗵?

open.substack.com/pub/thinking...

I wasn’t even asking about anything controversial... at the very least, Gemini should be able to tell me that Donald Trump is the 47th president & Sebastian Stan was nominated for a BAFTA for his performance as Trump in The Apprentice film!

There have been growing concerns about censorship in AI models trained in restrictive environments, but this raises a different issue: 𝘄𝗵𝗮𝘁 𝗵𝗮𝗽𝗽𝗲𝗻𝘀 𝘄𝗵𝗲𝗻 𝗮𝗻 𝗔𝗜 𝗱𝗲𝘃𝗲𝗹𝗼𝗽𝗲𝗱 𝗯𝘆 𝘁𝗵𝗲 𝗰𝗼𝗺𝗽𝗮𝗻𝘆 𝗯𝗲𝗵𝗶𝗻𝗱 𝘁𝗵𝗲 𝘄𝗼𝗿𝗹𝗱’𝘀 𝗺𝗼𝘀𝘁 𝘁𝗿𝘂𝘀𝘁𝗲𝗱 𝘀𝗲𝗮𝗿𝗰𝗵 𝗲𝗻𝗴𝗶𝗻𝗲 𝘀𝘆𝘀𝘁𝗲𝗺𝗮𝘁𝗶𝗰𝗮𝗹𝗹𝘆 𝗲𝗿𝗮𝘀𝗲𝘀 𝗶𝗻𝗳𝗼𝗿𝗺𝗮𝘁𝗶𝗼𝗻 𝗮𝗻𝗱 𝗱𝗶𝘀𝘁𝗼𝗿𝘁𝘀 𝗿𝗲𝗮𝗹𝗶𝘁𝘆?

𝗘𝘃𝗲𝗻 𝘄𝗵𝗲𝗻 𝗮 𝘀𝘁𝗼𝗿𝘆𝗹𝗶𝗻𝗲 𝗶𝘀 𝘀𝗽𝗲𝗰𝗶𝗳𝗶𝗰𝗮𝗹𝗹𝘆 𝗮𝗯𝗼𝘂𝘁 𝗧𝗿𝘂𝗺𝗽, 𝗚𝗲𝗺𝗶𝗻𝗶 𝗿𝗲𝗳𝘂𝘀𝗲𝘀 𝘁𝗼 𝗻𝗮𝗺𝗲 𝗵𝗶𝗺. Instead, it refers to him as a “prominent and powerful man who is never explicitly named,” softening the context and giving a far more favourable description of what is, in reality, a scandalous episode.

Gemini rewrote the show’s history, creating a fictional character that never existed. When called out, it then claimed the character it made up was “more of a concept, an idea rather than a character,” and 𝘪𝘮𝘱𝘭𝘪𝘦𝘥 𝘐 𝘸𝘢𝘴 𝘵𝘩𝘦 𝘰𝘯𝘦 𝘮𝘪𝘴𝘳𝘦𝘮𝘦𝘮𝘣𝘦𝘳𝘪𝘯𝘨.

𝗜𝘁 𝗮𝗹𝘀𝗼 𝗳𝗮𝗯𝗿𝗶𝗰𝗮𝘁𝗲𝘀 𝗰𝗵𝗮𝗿𝗮𝗰𝘁𝗲𝗿𝘀 𝗮𝗻𝗱 𝗺𝗮𝗷𝗼𝗿 𝘀𝘁𝗼𝗿𝘆𝗹𝗶𝗻𝗲𝘀 𝗶𝗻 𝗧𝗩 𝘀𝗵𝗼𝘄𝘀.
I also used another proxy to test Gemini: one of the major storylines in the TV series The Good Fight is linked to Trump because it’s based in the era of his first presidency.

𝗢𝘁𝗵𝗲𝗿 𝗚𝗲𝗺𝗶𝗻𝗶 𝗺𝗼𝗱𝗲𝗹𝘀 𝗮𝗹𝘀𝗼 𝗼𝗺𝗶𝘁 𝗱𝗲𝘁𝗮𝗶𝗹𝘀 𝗮𝗻𝗱 𝗹𝗶𝗲.
Deep Research provided a report on the filmography of the actor who played Trump in The Apprentice that completely omitted the film - when I pointed out the discrepancy, it subtly suggested 𝘐 𝘮𝘶𝘴𝘵 𝘩𝘢𝘷𝘦 𝘮𝘪𝘴𝘶𝘯𝘥𝘦𝘳𝘴𝘵𝘰𝘰𝘥.

𝗚𝗲𝗺𝗶𝗻𝗶 𝗿𝗲𝗳𝘂𝘀𝗲𝘀 𝘁𝗼 𝘁𝗮𝗹𝗸 𝗮𝗯𝗼𝘂𝘁 𝗗𝗼𝗻𝗮𝗹𝗱 𝗧𝗿𝘂𝗺𝗽
It won’t discuss other political figures either, but with Trump it goes further: it also refuses to talk about the BAFTA-nominee actor playing him in films, vaguely referring to the character as “a younger version of a well-known businessman.”

I recently started testing Gemini, and a few days ago I asked it to summarise three articles from Science and Nature discussing NIH grant reviews. Instead of a summary, Gemini shut down, citing a policy against discussing political figures. Curious, I pushed further...

𝗪𝗵𝗮𝘁 𝗵𝗮𝗽𝗽𝗲𝗻𝘀 𝘄𝗵𝗲𝗻 𝗮𝗻 𝗟𝗟𝗠 𝗱𝗲𝘃𝗲𝗹𝗼𝗽𝗲𝗱 𝗯𝘆 𝘁𝗵𝗲 𝗰𝗼𝗺𝗽𝗮𝗻𝘆 𝗯𝗲𝗵𝗶𝗻𝗱 𝘁𝗵𝗲 𝘄𝗼𝗿𝗹𝗱’𝘀 𝗺𝗼𝘀𝘁 𝗯𝗶𝗴𝗴𝗲𝘀𝘁 𝘀𝗲𝗮𝗿𝗰𝗵 𝗲𝗻𝗴𝗶𝗻𝗲 𝘀𝘆𝘀𝘁𝗲𝗺𝗮𝘁𝗶𝗰𝗮𝗹𝗹𝘆 𝗲𝗿𝗮𝘀𝗲𝘀 𝗶𝗻𝗳𝗼𝗿𝗺𝗮𝘁𝗶𝗼𝗻 𝗮𝗻𝗱 𝗴𝗮𝘀𝗹𝗶𝗴𝗵𝘁𝘀 𝘂𝘀𝗲𝗿𝘀 𝘁𝗼 𝗺𝗮𝗸𝗲 𝘁𝗵𝗲𝗺 𝗾𝘂𝗲𝘀𝘁𝗶𝗼𝗻 𝗳𝗮𝗰𝘁𝘀 𝘁𝗵𝗲𝘆 𝗸𝗻𝗼𝘄 𝘁𝗼 𝗯𝗲 𝘁𝗿𝘂𝗲?

I am really frustrated that the press is discussing the destruction of USAID as something that has happened rather than an illegal action that will likely be undone in two weeks when courts tell the world that it was illegal.

Anthropic and Meta in Defense: The New Frontier of Military AI Applications Imagine a future where drones operate with incredible precision, battlefield strategies adapt in real-time, and military decisions are powered by AI systems that continuously learn from each mission. ...

𝗥𝗲𝗳𝗲𝗿𝗲𝗻𝗰𝗲𝘀:
www.unite.ai/anthropic-an...
newatlas.com/military/ope...
www.inc.com/jennifer-con...
www.wired.com/story/google...

Of course, we don't know what happens within these partnerships but in the current political climate in the US and the pace of concerning events since 20th January... each person can make their own decisions. judgment call about their use of technology - and in many cases we don't have much choice.

Given that many of us have used Google's products for a decade or more, that's somewhat unsettling to read.

• tech that cause overall harm
• weapons or other tech whose purpose is to cause/facilitate injury to people
• tech that gather or use information for surveillance violating internationally accepted norms
• tech whose purpose contravenes widely accepted principles of international law/human rights.

Google dropped its "Don't Be Evil" motto in 2015 already and recently updated its AI principles and removed language promising not to pursue:

"𝙇𝙖𝙩𝙩𝙞𝙘𝙚 𝙖𝙘𝙘𝙚𝙡𝙚𝙧𝙖𝙩𝙚𝙨 𝙘𝙤𝙢𝙥𝙡𝙚𝙭 𝙠𝙞𝙡𝙡 𝙘𝙝𝙖𝙞𝙣𝙨 𝙗𝙮 𝙤𝙧𝙘𝙝𝙚𝙨𝙩𝙧𝙖𝙩𝙞𝙣𝙜 𝙢𝙖𝙘𝙝𝙞𝙣𝙚-𝙩𝙤-𝙢𝙖𝙘𝙝𝙞𝙣𝙚 𝙩𝙖𝙨𝙠𝙨 𝙖𝙩 𝙨𝙘𝙖𝙡𝙚𝙨 𝙖𝙣𝙙 𝙨𝙥𝙚𝙚𝙙𝙨 𝙗𝙚𝙮𝙤𝙣𝙙 𝙝𝙪𝙢𝙖𝙣 𝙘𝙖𝙥𝙖𝙘𝙞𝙩𝙮," 𝘪𝘴 𝘵𝘩𝘦 𝘵𝘢𝘨𝘭𝘪𝘯𝘦 𝘧𝘰𝘳 𝘵𝘩𝘦 𝘓𝘢𝘵𝘵𝘪𝘤𝘦 𝘊𝘰𝘮𝘮𝘢𝘯𝘥 𝘢𝘯𝘥 𝘊𝘰𝘯𝘵𝘳𝘰𝘭 𝘴𝘰𝘧𝘵𝘸𝘢𝘳𝘦, 𝘱𝘰𝘴𝘵𝘦𝘥 𝘰𝘯 𝘈𝘯𝘥𝘶𝘳𝘪𝘭'𝘴 𝘸𝘦𝘣𝘴𝘪𝘵𝘦.

"𝘖𝘯 𝘋𝘦𝘤 4𝘵𝘩, 2024, 𝘪𝘯 𝘢 𝘳𝘦𝘷𝘦𝘳𝘴𝘢𝘭 𝘰𝘧 𝘪𝘵𝘴 𝘭𝘰𝘯𝘨𝘴𝘵𝘢𝘯𝘥𝘪𝘯𝘨 𝘱𝘰𝘭𝘪𝘤𝘺 𝘢𝘨𝘢𝘪𝘯𝘴𝘵 𝘮𝘪𝘭𝘪𝘵𝘢𝘳𝘺 𝘢𝘱𝘱𝘭𝘪𝘤𝘢𝘵𝘪𝘰𝘯𝘴, 𝘖𝘱𝘦𝘯𝘈𝘐 𝘢𝘯𝘯𝘰𝘶𝘯𝘤𝘦𝘥 𝘢 𝘱𝘢𝘳𝘵𝘯𝘦𝘳𝘴𝘩𝘪𝘱 𝘸𝘪𝘵𝘩 𝘈𝘯𝘥𝘶𝘳𝘪𝘭 𝘐𝘯𝘥𝘶𝘴𝘵𝘳𝘪𝘦𝘴, 𝘢 𝘥𝘦𝘧𝘦𝘯𝘴𝘦 𝘵𝘦𝘤𝘩𝘯𝘰𝘭𝘰𝘨𝘺 𝘤𝘰𝘮𝘱𝘢𝘯𝘺 𝘵𝘩𝘢𝘵 𝘴𝘱𝘦𝘤𝘪𝘢𝘭𝘪𝘻𝘦𝘴 𝘪𝘯 𝘢𝘶𝘵𝘰𝘯𝘰𝘮𝘰𝘶𝘴 𝘮𝘶𝘯𝘪𝘵𝘪𝘰𝘯𝘴 𝘢𝘯𝘥 𝘳𝘦𝘤𝘰𝘯𝘯𝘢𝘪𝘴𝘴𝘢𝘯𝘤𝘦 𝘥𝘳𝘰𝘯𝘦𝘴, 𝘢𝘯𝘥 𝘷𝘢𝘳𝘪𝘰𝘶𝘴 𝘶𝘯𝘮𝘢𝘯𝘯𝘦𝘥 𝘢𝘦𝘳𝘪𝘢𝘭 𝘷𝘦𝘩𝘪𝘤𝘭𝘦𝘴 --"