Karin Verspoor

I helped build a government AI system. DOGE fired me, rolled the AI out to the whole agency, and implied the AI can do my job and the jobs of the others they've fired.

It can't. But, what DOGE accidentally revealed about themselves in the process is fascinating. 🧵

What are we looking at on the screen?

Zero‐ and few‐shot prompting of generative large language models provides weak assessment of risk of bias in clinical trials Existing systems for automating the assessment of risk-of-bias (RoB) in medical studies are supervised approaches that require substantial training data to work well. However, recent revisions to RoB...

Will share when it’s up. In the meantime here is something relevant to the use of LLMs for quality assessment of medical research:

“Zero- and few-shot prompting of generative large language models provides weak assessment of risk of bias in clinical trials”
doi.org/10.1002/jrsm...

Given the challenges that LLMs have in processing numbers correctly, and the prevalence and importance of numbers in scientific papers, I suspect this will not work well.

We have a paper upcoming at NAACL’25 in May showing the limitations of LLMs for numerical reasoning.

I made the shift from long commute on train+tram to an 8-minute walk. Highly recommend! I can even pop home for lunch!

Portraits de femmes scientifiques YouTube video by FemmesetSciences

Today is International Women's Day 👭🧪🔭. With this video we hope to motivate girls to engage in scientific careers, please circulate/ Cette vidéo en français/anglais a vocation à être diffusée dans les classes 👩‍👩‍👧‍👦. www.youtube.com/watch?v=uKlT...
#InternationalWomensDay2025 #JourneeDesDroitsDesFemmes

All 3 papers were sent to me by (different) @springernature.com "Discover" journals. Obviously they are being targeted by desperate authors. Stronger editorial pre-screening is needed.

Another day, another completely incomprehensible paper to review, with no meaningful research question or results. More inappropriate references.

The lit review was almost funny (if it weren't so bad) -- presented in alphabetical order by author's last name! When would that ever be a good idea?!

I’m sure it is happening all the time, sadly. My concern is how often such papers are getting through peer review and getting published.

Fake papers are contaminating the world’s scientific literature, fueling a corrupt industry and slowing legitimate lifesaving medical research ‘Paper mills’ are contaminating the world’s scholarly output with fake papers that hinder research. Lifesaving biomedical fields are especially hard hit.

Fake papers are contaminating the world’s scientific literature, fueling a corrupt industry and slowing legitimate lifesaving medical research

@joelving.bsky.social @gcabanac.cpesr.fr and Cyril Labbé have the stories, the data, and this receipts in our partnership with @theconversation.com

@retractionwatch.com How do we avoid getting these papers published in the first place?!

Update: The second paper received a “revise” decision, despite clear violations of editorial policies I listed in my review. I have contacted the EIC to request a review of that decision.

NB none of the other reviewers commented on the fabricated and inappropriate references.

The system is broken.

For those of us who don’t exclusively publish in or need to read biomedical journals, Google Scholar provides an important (necessary) resource. I try to teach my students to scrutinise both the source and the quality of the work.

So it is a matter of time investment, not innovation? Interesting.

Thanks; Google Scholar is also up so I can find most things.

I guess my concern is what's going on with PubMed!

Screenshot of web connection error showing that pubmed is down.

#PubMed is down today. Coincidence?!

I just reviewed two papers. One was almost certainly written with heavy assistance from genAI. The other was full of missing, misleading and incorrect references, many apparently planted.

I suspected as much from the abstracts.

I hate that I have to spend time policing science.
#researchintegrity

Super excited to be a part of this collaboration!

This #AI image from an unknown source at an unknown time has been making the rounds and finally got to me. It is amusing because of how many ways it is wrong ...

For me, it reminded me of challenges we faced working on references in recipes. Recipes are hard!
aclanthology.org/2022.finding...

No, they just want people to cite it. That's different.

@gcabanac.cpesr.fr in case you have not yet seen this, it is up your alley.

I always use train-DEVELOPMENT-test I suppose because of the same discomfort. Validate is arguably more common the dev step. I suppose because we use it to validate choices during training?

My guess is it is in a book/influential tutorial somewhere that way and that drove everything…