Had an agent just start writing a CSV by hand today (unclear what relation the data had to reality) when the function it wrote didn’t work correctly in its environment. Stay safe out there!

Why is use of AI and LLMs so divisive for academics? I think it's because ~half of them vote centre-left and ~half of them vote environmental-left When I did a nat rep survey in the UK, these two had opposite views of economic growth and tech optimism jamesbreckwoldt.substack.com/p/what-do-th...

A modest-sized but truly random sample of opinions is a better measure of opinions in the population than an enormous but self-selected sample of opinions from the same population

Some of these are quite popular but * military analysis: Anders Puck Nielsen * astrophysics: Dr Becky, Cool Worlds * AI/machine learning: Welch Labs * environmental science: Simon Clark * beautiful coding projects: Sebastian Lague * piano tutorials: mangold project * NFL analysis: Brett Kollman

This is very annoying as an American citizen as the US also gets grumpy about traveling through its border on a non-US passport. so I’ll have to figure out the right sequence of passports to show not to hurt any government’s feelings

Interesting that the policy ends up coming down to "at least paraphrase what the AI wrote". I think there's a lot to be said for this to make sure at least one human brain paid enough attention to rewrite what the AI said. But fascinating how quickly we're having to grapple with all of this

The most downloaded paper on EconStor in Feb. 2026 was: "Briggs, Ryan C.; Mellon, Jonathan; Arel-Bundock, Vincent (2026) : It must be very hard to publish null results, I4R Discussion Paper Series, No. 281, Institute for Replication (I4R), s.l." hdl.handle.net/10419/336819 @i4replication.bsky.social

Conditionally accepted at the APSR (w/ @scottclifford.bsky.social & @patrickpliu.bsky.social): Why does political information so often change beliefs but NOT attitudes? We highlight the role of belief relevance, or the extent to which beliefs bear on attitudes.

This proposes a way of using AI agents to produce research. Ok. But this bit is a pipe dream: "And human scientists should retain authority over — and responsibility for — framing the question, validating the path and signing off on conclusions." Here's why... /1

this but for scientists who have p-hacked

I really do mean every day too

This undersells the fact that you would sometimes do zoom meetings during the kayak!

I kayaked to work every day for 3 years

The main thing I think AI is going to do is generically put pressure on all of these institutions. That opens up space for changes in general including ones that might not be directly tied to AI.

Something tells me this very legitimate agent may not have read my book

Time to repost some causality memes #IdentificationK9

I remember reading a paper at one point that drew a distinction between polling vote share error (how far off the vote shares were) and polling narrative error (when the wrong party was said to be ahead). Does anyone remember the reference?

Looking at a couple of panel studies my current working hypothesis is that sustained participation in panel studies decreases respondents life satisfaction.

Problem about the loneliness epidemic is, it's everywhere except in representative survey data. Let's look at where the claim comes from. 1/

I think it’s worth bearing in mind that graduate premium is a measure of difference not the level of achievement within the graduate group.

I've been interested when we were going to reach the capability threshold for this. Once the tires have been kicked, all social science journals should adopt computational reproducibility on submission. Huge reduction in reviewer burden by eliminating a whole class of errors up front.

No. There isn’t. Those of us who work on this issue try to debunk that regularly. But there is a persistent level of loneliness that is very damaging for those who experience it - and the solutions they need bring wider good. There is also some evidence that some people’s loneliness is deepening.

but it turns out that this seems to be driven by that question battery being long and interviewers who were paid per interview encouraging respondents to skip it! journals.sagepub.com/doi/abs/10.1...

Good summary for the lack of a loneliness epidemic here ourworldindata.org/loneliness-e...

Case in point, @pengzell.bsky.social just sent me 5 papers proving that there’s not actually any evidence for a loneliness epidemic. My mind is highly changed

My experience with reading journal reviews on day 1 versus a week later

Maybe it’s just the more academic side of Bluesky discourse but I find my mind being changed pretty often and see it happening for others too

Agree that abstracts aren’t all that matters (but I do think they matter a lot). We’re working on the followup paper now where we’re collecting all tests run in papers and the way the authors discuss them in the text.

This is a really fair point and we’ll definitely fix this on the next iteration

I see this view a lot but I think AI will change RA work not eliminate it. I think AI is going to make some tasks unnecessary (data extraction) but others more viable (having RAs directly develop a data collection interface or check assumptions in an analysis)

To a second approximation, only the journal of experimental political science does (we checked over 150 journals) bsky.app/profile/ryan...

Yes, but have you considered that I — a very normal Bluesky user — have a strong belief that AI is bad? I am very smart

To a first approximation literally nowhere in political science likes to talk about nulls in their abstracts

We spent months training grad student RAs and GPT-5 mini still beat them by a lot

I am on paternity leave, so not promoting this properly, but lots of Britons have stable but idiosyncratic views , as shown in this book doi.org/10.1093/9780...

Yes that’s definitely true. I have a lot of cites to my “only use this method if you run a validation for your use case” papers that read “Mellon (2014) shows that this method is valid for my use case”

I think the LLM hallucinated citations problem is going away (because it’s fundamentally easy to check if a paper exists using an API) but the much harder problem (especially for faculty reviewing student work) will be hallucinating the content of citations

It's a bit annoying that the default python setup for datascience (jupyter notebooks) is a bad fit for git version control (you end up with inconsequential chunk output merge issues). I wish they'd gone with a stricter separation of code and output like rmd.

If it makes you feel better seahorses are more closely related to horses than sharks

Psychologists stop trying to do psychological science no wait

When you stop borrowing our methodologies

Economists stop trying to do political science

Tim Harford has a great podcast on this where he shows that people didn’t lose that much money on the tulip bubble but lost their shirts investing in railroads (despite the fact that those were a transformative innovation) timharford.com/2023/01/caut...

By the way this is separate from whether you think AI companies stocks are overvalued. The technology at its present level is going to reshape academic research even if Sam Altman goes bankrupt

I know Bluesky is more LLM sceptical, but I definitely had the experience both that gpt-5.2 is superhuman at getting proofs to compile and that it’s not that well aligned (and is chasing compilation success not proving the substantive claims I wanted to)

I've been playing with Lean to prove consistency of a 2-stage estimator in a sociology paper. I have a question for people who do lean proofs. If you understand the theorem and the assumptions it relies on, how much can you treat the an LLM lean proof as a black box

I’ve had an idea floating around my head that there might be a niche audience that would be interested in watching me live stream researching and writing social science papers in realtime on twitch. Terrible idea or kind of fun?

I'm sure someone else has suggested this, but it's time for mandatory DOIs in reference lists. This will make it easier to identify hallucinated references

Does anyone see an issue with using parenthesis color highlighting when typesetting equations like this in latex? ChatGPT seems to think it’s uncouth but I think it would be a clear improvement.