"In a landmark study, OpenAI researchers reveal that large language models will always produce plausible but false outputs, even with perfect data, due to fundamental statistical and computational limits." www.computerworld.com/article/4059...

Can large language models stand in for human participants? Many social scientists seem to think so, and are already using "silicon samples" in research. One problem: depending on the analytic decisions made, you can basically get these samples to show any effect you want. THREAD 🧵

A very interesting and stealthy property of RNAs once they get modified by glycosylation:

Journal publishers *are* bundling your papers up in "data licensing agreements" for big tech companies to use for AI model training. Our publisher, T&F, got £75m from Microsoft for that last year alone.

One of the most common misunderstandings about the use and value of placebos in clinical RCTs, often made by both methodological experts and experienced trialists: (from @stephensenn.bsky.social in academic.oup.com/jrsssa/advan...)

During the 1970s, an arson wave ravaged poor communities across the US. Some lost 80% of their housing. Residents were blamed. In reality, landlords—paid by insurers and enabled by the state—drove this arson-for-profit epidemic. 

I reviewed @benchansfield.bsky.social's revelatory BORN IN FLAMES:

There are three major issues with the Baby’s First Years study design. Saying monthly cash won’t help children’s health is highly misleading — even irresponsible www.statnews.com/2025/08/19/b...

1. I need to recover this thread from X about a review on lithium in tap-water and suicide. It shows that papers aren't retracted despite being total nonsense

Totally ridiculous (but not even a little surprising) that you can publish a f***-up this bad, and when someone with a brain spots it the editor treats it like a "both sides" debate (see thread).

85% of all drug candidates that go into human trials fail. Expensively. This is the central fact of the entire research pharma industry.

Published in "Smart Medicine" 2025. The authors used three cell lines to study HCC. None of these are HCC cell lines. Two of them are known to be HeLa since 2017. Last author at @stonybrooku.bsky.social

Heart disease, Infectious disease Hugely inpactful fields of health. Such underappreciated work by @robcaliff.bsky.social and reporting by @stuartbuck.bsky.social . Summers always show who is on the pulse even if it's not paid.

This, of course, results in massive wastage of the data that citizens have contributed & field workers have sweated to collect. I witnessed the director of a well-known epi study insisting a teaching data set be destroyed "in case a student finds something we didn't".😠😡

TIL the original paper describing CRISPR, by Francisco Mojica, was rejected by 4 journals and took 2 years to be published

goodscience.substack.com/p/how-we-cou... Such a great article on how 90% of all medical recommendations in certain fields are based on barely anything worth called evidence. Not enough to put a person into contact with any such procedure.

No matter how hard you try, you simply cannot guess the journal withe the highest “impact factor” in organic chemistry last year. Really. Nor will you guess many of the other “winners”:

We are launching a project on improving code sharing in health and medical research funded by @ukri.org and working with @floriannaudet.bsky.social and @bmj.com Read more here: bennett.ox.ac.uk/blog/2025/08/help-us-improve-code-sharing-in-health-research/

The number of children injured or killed by cars while walking or riding a bike has fallen steadily since the 1970s, but CDC researcher note that this decline is not because streets are safer, but because fewer kids are out and about in the first place.

A very interesting (and very thorough) case for Alzheimer’s as (of all things) a disease of lithium deficiency. This one’s worth taking seriously, I believe:

This is actually very cool and good.

My latest, exploring with numbers how the millennial household budget is basically incomprehensible to retired boomers.

“6% of sleuths agreed that current reporting channels for misconduct are effective, compared to 77% of research-integrity officers” I think this is readily explainable by the differential access to information. RIOs do not see what doesn’t even reach them, eg when editors refuse to act.

We might have been looking at atherosclerosis the wrong way all these years. Is the cause a bacterial metabolite?

I think it was a target trial emulation of difference in differences for a propensity score matched analysis with Bayesian borrowing of historical controls of digital twins based on synthetic data from the real world.

Thrilled to share that our paper on instrument misidentification in materials science is now peer-reviewed and out in @plosone.org! doi.org/10.1371/jour...

Essay mills shamelessly advertising on the academia feed. CC @keithwilson.eu

Approving Sarepta’s gene therapy was a bad idea in 2024, and it’s looking worse all the time.

Absolutely. We are writing such guidance to accompany the INSPECT-SR tool : bmjopen.bmj.com/content/14/3...

You know how we all know that neurons run exclusively on glucose and don’t use fatty acids as fuel? Well. . .

Damn 61% is good these days? What has sciense become... (It's really time for my (non-IT) tool that prevents image recordings from being manipulated; there is only one big incident missing for the funding to flow properly)

This isn't a coincidence, IMO. The fly community is generally super rigorous - because there are so many tools and techniques available, people expect you to really nail things in order to get published (and keep getting funded)

My own university has a "strategic goal" to increase research "outputs" by 10% annually (7 year doubling time). Your university is almost surely working towards the same. University rankings are now the main driver of this damaging institutional behavior. Publication is an out-of-control arms race.

New post on Restores' website: new research finds that "RCTs on labour induction without IPD-sharing are more likely to have lower quality, more trustworthiness concerns, and exaggerated effect estimates than those with shared individual patient data". restores.univ-rennes.fr/highlights/d...

If your name is on a paper you didn't contribute directly to, or even carefully read one time, you are a dishonest fraud, end of. If you publish as part of a cabal that puts everyone's name on every paper regardless of direct input, you are all dishonest frauds.

This is just going to be the first of many, many stories you are going to read about big tech companies faking "AI for climate" claims. The sheer volume of AI-climate-washing going on right now is huge, and it's very likely most of the claims are fabricated or exaggerated archive.is/8sYNv

I have three questions. www.theguardian.com/technology/2...

In Dodgy News today: People will not remember, but before Brian Deer sorted out Andrew Wakefield, he worked on Briggs. And his daughter Joanne wrote a memoir / investigation, and I am beyond excited to jam three margaritas into my head and read it. www.thetimes.com/culture/book...

Even at the time, we knew about what were (and maybe still are) commercial p-hacking services, where you can send your data for 'consultation' and get a result. We knew that people would not treat it like an exercise, but rather just straight up use it in the exact manner it wasn't intended.

This is quite correct in the most important sense. I'll tell you a story: over ten years ago, I built an automatic p-hacker (in Matlab, I'm old). It took some data, changed more or less every analysis parameter possible, used every outlier management schema, etc. and then subgroups. It could...

Turns out that raising virtually insane expectations literally has a price. “Compass Pathways’ stock has crashed by nearly 50% despite the company announcing its Phase III trial of psilocybin had met its primary endpoint [..] investors seem unconvinced by the impact of the therapy in the study.”

ooh look at me it took me ten years to invent my own job and i only have headshots that make me look like a kindly potato also i really enjoyed this interview www.politico.com/newsletters/...

@dereklowe.bsky.social How on earth should axiom show that their cell collection is similar enough to a liver within a human?

We've frustratingly allowed "real world" to become a deceptive synonym for "observations we happened haphazardly to make". I'd say half the studies that come to me now, the majority of which are useful but modest clinical audits, will have "real world" in the proposal title. It's magic!

Finally, something AI can do just as well as real "researchers". www.nature.com/articles/d41...

Springer-Nature launched a series of "Discover" journals that closely mimic MDPI titles -- sharing *identical* journal names, and likely similar business model. What is going on, and why researcher will - as always - fall for it? A 🧵 the-strain-on-scientific-publishing.github.io/website/post...

Weekend reads: A university ‘integrity index’; surgeon falsely claims Harvard Ph.D.; more on ‘gold standard science’ executive order

A targeted effort to find and isolate some of the most harmful papers from the medical literature:

"Ten institutions create more than half of UK university spinouts" [PAYWALLED] buff.ly/uitE2Cr