oslo.town is one of the many independent Mastodon servers you can use to participate in the fediverse.
An online home for the people of Oslo, Norway 🇳🇴 but a gateway to the world.

Server stats:

228
active users

#icssi

0 posts0 participants0 posts today

"Elite" peer review was studied by Daniel Larremore and colleagues using an anonymised dataset from AAAS of 112,000 papers submitted to Science and Science Advances. A few more desk rejects for women authors, but no difference in peer review. Clear preference for work from prestigious institutions: biased or just better? Big teams also do better.
#ICSSI #ScienceofScience #PeerReview #ScienceMagazine #ScienceAdvances #DanielLarremore #AAAS

How do different values affect peer review and how do these influence editorial decisions, asks Daniel Scott Smith? Very different reviews can make decisions uncertain.

Across STEM, they find accuracy and novelty are the most important values in peer review reports, social factors (biases) have little effect, and dissensus increases editorial gatekeeping. It doesn't matter if reviewers disagree.

Juan Mateos-Garcia of Google DeepMind wants equitable use of their tools like AlphaFold. They solved protein structure prediction in 2020 and this is helpful for drug discovery, but researchers in LMIC have barriers to adoption.

Used OpenAlex & PDB to study 20.6K papers citing AlphaFold. LMIC researchers are underrepresented, but study of diseases affecting LMICs is overrepresented - esp. work on natural products.

Corporate research is dominating AI science, says Nur Ahmed, but they're not engaging with responsible AI research. They're choosing speed over safety.

They used supervised machine learning on 6 million papers and found few firms were working on responsible AI. A handful attended responsible AI conferences vs hundreds at other AI conferences.

When they do engage, it's on explainability not societal impacts and human rights.

"Let's *delve* into this topic..."

Science of AI-mediated science - Shahan Ali Memon gave a very engaging talk on how AI is being used in science, how it changes how we do and think about science, and how these AI tools are different to those we've used before.

He was too modest to advise a representative of the UK government in the audience on how much of his metascience budget should pivot to AI.

Elle O'Brien at UMich says there are barriers to data science in academia. Does generative AI break this down?

Many researchers are using LLMs for tips on using software even when good documentation exists. They're not software literate.

They wrongly think chatbots do web search, can code, have databases, can calculate, and give faithful explanations. Does that make their use more risky?

A study of "beauty bias" in scientific careers presented at #ICSSI made me uneasy. Scraping faces from Google Scholar, business school websites, and job applicants and using machine learning to rate attractiveness without their consent seems disrespectful, esp. as the names of the "beautiful" institutions (which correlates with school ranking) are released. A small business school might not appreciate being labelled as ugly! 🧵
#ScienceofScience #Attractiveness #CareerProgression #ScienceCareers

How does industry shape academic science? Hongyuan Xia used a natural experiment to compare scientists in places that narrowly miss getting big industrial plants (like Amazon) built vs those that get them. Used citation of patents by the companies to define relevant fields.

Industry presence increases impact, number of papers in the field, and commercialisation. Why? Funding, collaboration, and researchers shifting focus.

China's rising leadership in global science: Christopher Esposito and UCLA colleagues studied leadership using author contribution statements. Extracted items were clumped using K-means clustering. They then trained a neural net to predict leadership among international collaborations in a large sample of papers.

In US-China collaborations, comparing 2010-12 vs 2020-23 shows a drop in US leaders and a rise in Chinese leaders.

"Helicopter science" and datasets: Sarah Bratt and colleagues studied the use of Global South (GS) data by Global North researchers (in 3200 Scientific Data papers) , using OpenAlex and an LLM, ReFinEd, to identify entities in Geonames.

The % of GS authors on GS data papers has risen, but on papers using GS data only 35% have a GS researcher as 1st or last author.

p.s. Saying "ding" instead of "next slide" was new to me!