Przemek Grabowicz 🇵🇱🇪🇸🇩🇪🇺🇸🇮🇪
@przemyslslaw.bsky.social
CS faculty at University College Dublin and the University of Massachusetts Amherst: responsible AI, social media, news media, computational social science, applied NLP, open-world AI
RT≠endorsement, RT=interesting read
More: https://przemyslslaw.github.io
RT≠endorsement, RT=interesting read
More: https://przemyslslaw.github.io
This research subsumes our previous study of the 2025 German federal election (attached figure and paper), before which Elon Musk supported and campaigned for the far-right AfD. We conclude that social media platforms can develop in different ways. 🧵6/7
zenodo.org/records/1489...
zenodo.org/records/1489...
November 3, 2025 at 12:58 PM
This research subsumes our previous study of the 2025 German federal election (attached figure and paper), before which Elon Musk supported and campaigned for the far-right AfD. We conclude that social media platforms can develop in different ways. 🧵6/7
zenodo.org/records/1489...
zenodo.org/records/1489...
A prior study by @inference.vc of the 2016 Twitter's feed algorithm did not find evidence of amplification of political extremes in seven countries, including Germany (attached figure), so the feed algorithm of X must have changed considerably since then. In fact, it did, as we show next. 🧵4/7
November 3, 2025 at 12:58 PM
A prior study by @inference.vc of the 2016 Twitter's feed algorithm did not find evidence of amplification of political extremes in seven countries, including Germany (attached figure), so the feed algorithm of X must have changed considerably since then. In fact, it did, as we show next. 🧵4/7
We test whether the feed algorithm favors political extremes, by regressing the number of feed appearances against user engagements received by each post and political party affiliation of its author. We find that each post from political extremes appeared about 20 minutes more in the feed! 🧵3/7
November 3, 2025 at 12:58 PM
We test whether the feed algorithm favors political extremes, by regressing the number of feed appearances against user engagements received by each post and political party affiliation of its author. We find that each post from political extremes appeared about 20 minutes more in the feed! 🧵3/7
The study (attached) was conducted during the culminating month of Polish presidential election. We tracked news feeds of four sock-puppet accounts and analyzed public posts of 485 Polish politicians and related users. 🧵2/7
doi.org/10.5281/zeno...
doi.org/10.5281/zeno...
November 3, 2025 at 12:58 PM
The study (attached) was conducted during the culminating month of Polish presidential election. We tracked news feeds of four sock-puppet accounts and analyzed public posts of 485 Polish politicians and related users. 🧵2/7
doi.org/10.5281/zeno...
doi.org/10.5281/zeno...
Our latest study of "Algorithmic Biases on X before the 2025 Polish Presidential Election" reveals that the default algorithmic feed of X showed mostly posts from right-wing parties and amplified political extremes, most likely due to the 2023 updates of X following its takeover (read on). 🧵1/7
November 3, 2025 at 12:58 PM
Our latest study of "Algorithmic Biases on X before the 2025 Polish Presidential Election" reveals that the default algorithmic feed of X showed mostly posts from right-wing parties and amplified political extremes, most likely due to the 2023 updates of X following its takeover (read on). 🧵1/7
Last days every morning and evening I was passing by a church that Berliners nicknamed "the hollow tooth", badly damaged in 1943 and never rebuilt as a memorial of WWII, on my way to an excellent and timely workshop about platform transparency organized in Berlin by @dsa40collaboratory.bsky.social.
September 27, 2025 at 12:54 PM
Last days every morning and evening I was passing by a church that Berliners nicknamed "the hollow tooth", badly damaged in 1943 and never rebuilt as a memorial of WWII, on my way to an excellent and timely workshop about platform transparency organized in Berlin by @dsa40collaboratory.bsky.social.
The design of social media platforms, including their AI systems and algorithms, likely contributes to the over-representation of extreme views (image @financialtimes.com). What platforms do we want? For example shall we balance diverse perspectives and societal stability?
www.ft.com/content/9251...
www.ft.com/content/9251...
September 9, 2025 at 10:29 AM
The design of social media platforms, including their AI systems and algorithms, likely contributes to the over-representation of extreme views (image @financialtimes.com). What platforms do we want? For example shall we balance diverse perspectives and societal stability?
www.ft.com/content/9251...
www.ft.com/content/9251...
On the day when Israel killed four more Al Jazeera journalists, it's time we recognize that Israel killed in Gaza more journalists than ever died in any other war, including the World Wars. Source: attached research at Brown University. Let this sink in.
watson.brown.edu/costsofwar/p...
watson.brown.edu/costsofwar/p...
August 10, 2025 at 11:30 PM
On the day when Israel killed four more Al Jazeera journalists, it's time we recognize that Israel killed in Gaza more journalists than ever died in any other war, including the World Wars. Source: attached research at Brown University. Let this sink in.
watson.brown.edu/costsofwar/p...
watson.brown.edu/costsofwar/p...
We gave with Vishal Kalakonnavar three talks at one of my favourite research conferences, @ic2s2.bsky.social, in Sweden, representing the SIMS lab. This was a fantastic conference 🎉. I'm looking forward to the next year's edition in Vermont!
I attach pointers to the studies we've presented:
I attach pointers to the studies we've presented:
July 25, 2025 at 9:47 AM
We gave with Vishal Kalakonnavar three talks at one of my favourite research conferences, @ic2s2.bsky.social, in Sweden, representing the SIMS lab. This was a fantastic conference 🎉. I'm looking forward to the next year's edition in Vermont!
I attach pointers to the studies we've presented:
I attach pointers to the studies we've presented:
Kathleen Carley introducing the term "social cybersecurity" at @ic2s2.bsky.social.
July 22, 2025 at 8:12 AM
Kathleen Carley introducing the term "social cybersecurity" at @ic2s2.bsky.social.
Delighted and thankful that our study "Election Polls on Social Media: Prevalence, Biases, and Voter Fraud Beliefs" received Best Paper Honorable Mention at @icwsm.bsky.social'25! 🙏🎊 Please check the out attached thread to learn more about our study! 📄
June 26, 2025 at 12:59 PM
Delighted and thankful that our study "Election Polls on Social Media: Prevalence, Biases, and Voter Fraud Beliefs" received Best Paper Honorable Mention at @icwsm.bsky.social'25! 🙏🎊 Please check the out attached thread to learn more about our study! 📄
Our case study reveals that the amount of news about a disasters or terrorist attack correlates to the death count, as well as the GDP of the event country. For example, an event in Nigeria requires 10-20 times more deaths than in Russia to receive the same level of attention.
June 25, 2025 at 12:45 PM
Our case study reveals that the amount of news about a disasters or terrorist attack correlates to the death count, as well as the GDP of the event country. For example, an event in Nigeria requires 10-20 times more deaths than in Russia to receive the same level of attention.
We use FAME to identify 27,441 articles that cover 470 natural disaster and terrorist attack events that happened in 2020. The below table lists the disasters with most news coverage, while the map shows the average number of news per disaster in a given country.
June 25, 2025 at 12:45 PM
We use FAME to identify 27,441 articles that cover 470 natural disaster and terrorist attack events that happened in 2020. The below table lists the disasters with most news coverage, while the map shows the average number of news per disaster in a given country.
The method achieves state-of-the-art performance in comparison to keyword-based baselines ( KW) and scales to massive databases of tens of millions of news articles in multiple languages and hundreds of events happening globally.
June 25, 2025 at 12:45 PM
The method achieves state-of-the-art performance in comparison to keyword-based baselines ( KW) and scales to massive databases of tens of millions of news articles in multiple languages and hundreds of events happening globally.
FAME doesn't require training data and efficiently identifies news articles that discuss an event given its fingerprint: time, location, and class. The approach has two phases of filtering news articles based on the fingerprint: (1) heuristic filtering and (2) LLM QA filtering.
June 25, 2025 at 12:45 PM
FAME doesn't require training data and efficiently identifies news articles that discuss an event given its fingerprint: time, location, and class. The approach has two phases of filtering news articles based on the fingerprint: (1) heuristic filtering and (2) LLM QA filtering.
Excited to share our FAME method for news identification: Fingerprint-to-Article Matching for Events from a DB! We use it to study news coverage of disasters and conflicts (w @brenocon.bsky.social @ethanz.bsky.social). Check out our talk and poster at @icwsm.bsky.social!🧵👇
arxiv.org/abs/2506.12925
arxiv.org/abs/2506.12925
June 25, 2025 at 12:45 PM
Excited to share our FAME method for news identification: Fingerprint-to-Article Matching for Events from a DB! We use it to study news coverage of disasters and conflicts (w @brenocon.bsky.social @ethanz.bsky.social). Check out our talk and poster at @icwsm.bsky.social!🧵👇
arxiv.org/abs/2506.12925
arxiv.org/abs/2506.12925
"There is no diplomacy in our world today. There is only war."
Iran signed a deal to reduce its nuclear program in 2015, but Trump terminated it in 2018. Now Trump forces Iran to sign a new deal, while threatening to destroy the nation. How does Iran know whether it won't be destroyed either way? 🤔
Iran signed a deal to reduce its nuclear program in 2015, but Trump terminated it in 2018. Now Trump forces Iran to sign a new deal, while threatening to destroy the nation. How does Iran know whether it won't be destroyed either way? 🤔
June 13, 2025 at 9:12 PM
"There is no diplomacy in our world today. There is only war."
Iran signed a deal to reduce its nuclear program in 2015, but Trump terminated it in 2018. Now Trump forces Iran to sign a new deal, while threatening to destroy the nation. How does Iran know whether it won't be destroyed either way? 🤔
Iran signed a deal to reduce its nuclear program in 2015, but Trump terminated it in 2018. Now Trump forces Iran to sign a new deal, while threatening to destroy the nation. How does Iran know whether it won't be destroyed either way? 🤔
How AI will surpass human intelligence...
😂😅
😂😅
June 1, 2025 at 7:18 AM
How AI will surpass human intelligence...
😂😅
😂😅
When the tests are hidden from LLMs, so that they cannot memorize the answers, as in ARC-AGI tests, the best LLMs perform much worse and 10x more expensive than humans (see table). STEM students score around 100% on these tests, out of the reach of the best LLMs. 🧵2/2
arcprize.org/blog/announc...
arcprize.org/blog/announc...
May 29, 2025 at 11:20 AM
When the tests are hidden from LLMs, so that they cannot memorize the answers, as in ARC-AGI tests, the best LLMs perform much worse and 10x more expensive than humans (see table). STEM students score around 100% on these tests, out of the reach of the best LLMs. 🧵2/2
arcprize.org/blog/announc...
arcprize.org/blog/announc...
I've seen recent claims that LLMs have IQ comparable to that of humans. However, if LLMs were trained on the IQ questions they are tested on, they could simply have memorized the answers, so IQ test results may not say anything about their intelligence, but rather about their memory abilities. 🧵1/2
May 29, 2025 at 11:20 AM
I've seen recent claims that LLMs have IQ comparable to that of humans. However, if LLMs were trained on the IQ questions they are tested on, they could simply have memorized the answers, so IQ test results may not say anything about their intelligence, but rather about their memory abilities. 🧵1/2
Achieving ARC-2 prize would mean that we have AI systems that are a few times cheaper at solving reasoning tasks (with 85% accuracy) than average humans (STEM grads achieve nearly 100% performance), since an MTurker costs $2+/task in ARC-AGI-1 (figure). It will take many years to achieve this! 🧵3/3
March 26, 2025 at 10:42 AM
Achieving ARC-2 prize would mean that we have AI systems that are a few times cheaper at solving reasoning tasks (with 85% accuracy) than average humans (STEM grads achieve nearly 100% performance), since an MTurker costs $2+/task in ARC-AGI-1 (figure). It will take many years to achieve this! 🧵3/3
Given how expensive the current best-performing fine-tuned OpenAI reasoning models are (~$4k/task at 88% accuracy, see figure), this bar corresponds to 10,000 times reduced cost. To achieve this bar, we will need a breakthrough in machine learning that’s more significant than the transformers. 🧵2/3
March 26, 2025 at 10:42 AM
Given how expensive the current best-performing fine-tuned OpenAI reasoning models are (~$4k/task at 88% accuracy, see figure), this bar corresponds to 10,000 times reduced cost. To achieve this bar, we will need a breakthrough in machine learning that’s more significant than the transformers. 🧵2/3
There is an issue with the green ARC Grand Prize box in the attached figure. I think it's supposed to mark $0.42/task cost, but it ends at around $0.2 instead. $0.42 would be past the middle point between the two ticks. Have a look at the attached log scale.
March 26, 2025 at 10:32 AM
There is an issue with the green ARC Grand Prize box in the attached figure. I think it's supposed to mark $0.42/task cost, but it ends at around $0.2 instead. $0.42 would be past the middle point between the two ticks. Have a look at the attached log scale.
I was honored to participate in the NSF-FAPESP Workshop on Cybersecurity and Privacy in São Paulo, Brazil! It was truly inspiring to connect with brilliant minds from academia and industry, all dedicated to advancing this increasingly important research field.
March 21, 2025 at 8:57 PM
I was honored to participate in the NSF-FAPESP Workshop on Cybersecurity and Privacy in São Paulo, Brazil! It was truly inspiring to connect with brilliant minds from academia and industry, all dedicated to advancing this increasingly important research field.
I am #hiring! If you are interested in AI and data science this #PhD position may be of interest to you! This is a fully funded position at @ucdcs.bsky.social, €22,000 tax free per annum. Research area: AI for mining, understanding and augmenting public opinion and public discourse. Link below! 🧵1/2
March 10, 2025 at 11:56 AM
I am #hiring! If you are interested in AI and data science this #PhD position may be of interest to you! This is a fully funded position at @ucdcs.bsky.social, €22,000 tax free per annum. Research area: AI for mining, understanding and augmenting public opinion and public discourse. Link below! 🧵1/2