Nature Portfolio

We have officially launched our new 1-day Nature Masterclasses workshop on Harnessing Generative AI in Academic Publishing! Last week, I had the pleasure of conducting two 1-day workshops for Peking University Health Science Center in Beijing. The first is our flagship Getting Published that focuses on scientific writing and manuscript structure. The second was our new workshop on generative AI. We have been including generative AI as a 2-hour module in our Getting Published workshop, but given the interest in the research community, we have expanded it into a full 1-day workshop as well. We first covered why outputs vary when using large language models (LLMs) that can lead to hallucinations or inappropriate output. We then reviewed common limitations and biases that researchers should be aware of along with strategies to mitigate them. Next, we discussed context engineering to improve how we communicate with LLMs. And finally we reviewed a number of common use cases of LLMs in research and publishing workflows. Thankfully, even though it was our first time presenting it, the feedback was very positive with 97% rating the quality as either excellent (84%) or good (13%). We will be updating the content to include agentic workflows and the role of AI as an assistant rather than just a tool in future workshops as that was a hot topic for discussion during the workshop. It's never a dull day developing useful solutions for the research community at Springer Nature! It was wonderful to work with all the great participants in the workshop, so thank you so much for attending! #GenerativeAI #AI #AcademicPublishing #LifeAtSpringerNature #BePartOfProgress

✨New paper out in Nature Portfolio today. ✨ For 8 weeks around the 2024 US election, we randomly assigned 2,000 people to use social media algorithms we built ourselves. The question: do engagement-based algorithms amplify intergroup, moral, and emotional (IME) content — and does that distort how we perceive political norms? We also built and tested an intervention. Rather than amplifying outgroup posts to "diversify" feeds (which can backfire), our "diversified extremity" algorithm reduces the outsized influence of extreme users — the small group who post most of the toxic political content — to make feeds more socially representative. 💠Two cool things: This is Nature's first published Stage 2 Registered Report. Everything you see was pre-registered and agreed upon before we ran the study. (I have a lot of thoughts now on the pros and cons of RRs — maybe a separate post.) We also wanted to show one way to run an algorithm RCT independent of industry. We engineered our own ranking algorithms on Bluesky, giving us full control and transparency — less black box, no conflicts of interest, no worry about the algorithm quietly changing mid-study. 💠Results: First: engagement-based feeds amplified IME and toxic content versus a chronological feed. The biggest jumps were in moral outrage (up to +79%) and political content (up to +57%) after the election. Our diversified extremity algorithm reduced this exposure. Because we estimated the true base rate of English-language content across the study, we can show this is a genuine overrepresentation of what's on the platform. Second: engagement-based feeds raised perceived partisan animosity (especially after the election) and reduced norm-perception accuracy. The surprising part — people actually underestimated how acceptable others found toxic political content. Our best explanation: amplified toxic posts often came wrapped in visible condemnation (replies, etc). Seeing others push back may have signaled toxic posting was less acceptable, even as the vibe felt more hostile. Third, null: no evidence that algorithms changed people's own engagement with toxic or IME content. That behavior was rare and concentrated — the top 5% of users drove 75% of all toxic engagement. Fourth: it's often assumed that dialing down divisive content hurts user experience. Our data suggest that tradeoff may be overstated — the diversified extremity feed cut toxic and IME content while maintaining login frequency and even improving overall platform enjoyment. Plenty of limitations. Bluesky's population changed dramatically mid-study. Our participants were earlier users, but later joiners expressed more outrage. How do findings generalize to other platforms? see discussion! 📄 Paper: https://lnkd.in/gCc56Ce7 Huge thanks to my incredible collaborators: Joshua Jackson Nour Kteily Eli Finkel Curtis Puryear Meriel Doyle Mark Torres Abdo Elnakouri Trevor Spelman Jake Teeny Victoria Parker.

✨New paper out in Nature Portfolio today. ✨ For 8 weeks around the 2024 US election, we randomly assigned 2,000 people to use social media algorithms we built ourselves. The question: do engagement-based algorithms amplify intergroup, moral, and emotional (IME) content — and does that distort how we perceive political norms? We also built and tested an intervention. Rather than amplifying outgroup posts to "diversify" feeds (which can backfire), our "diversified extremity" algorithm reduces the outsized influence of extreme users — the small group who post most of the toxic political content — to make feeds more socially representative. 💠Two cool things: This is Nature's first published Stage 2 Registered Report. Everything you see was pre-registered and agreed upon before we ran the study. (I have a lot of thoughts now on the pros and cons of RRs — maybe a separate post.) We also wanted to show one way to run an algorithm RCT independent of industry. We engineered our own ranking algorithms on Bluesky, giving us full control and transparency — less black box, no conflicts of interest, no worry about the algorithm quietly changing mid-study. 💠Results: First: engagement-based feeds amplified IME and toxic content versus a chronological feed. The biggest jumps were in moral outrage (up to +79%) and political content (up to +57%) after the election. Our diversified extremity algorithm reduced this exposure. Because we estimated the true base rate of English-language content across the study, we can show this is a genuine overrepresentation of what's on the platform. Second: engagement-based feeds raised perceived partisan animosity (especially after the election) and reduced norm-perception accuracy. The surprising part — people actually underestimated how acceptable others found toxic political content. Our best explanation: amplified toxic posts often came wrapped in visible condemnation (replies, etc). Seeing others push back may have signaled toxic posting was less acceptable, even as the vibe felt more hostile. Third, null: no evidence that algorithms changed people's own engagement with toxic or IME content. That behavior was rare and concentrated — the top 5% of users drove 75% of all toxic engagement. Fourth: it's often assumed that dialing down divisive content hurts user experience. Our data suggest that tradeoff may be overstated — the diversified extremity feed cut toxic and IME content while maintaining login frequency and even improving overall platform enjoyment. Plenty of limitations. Bluesky's population changed dramatically mid-study. Our participants were earlier users, but later joiners expressed more outrage. How do findings generalize to other platforms? see discussion! 📄 Paper: https://lnkd.in/gCc56Ce7 Huge thanks to my incredible collaborators: Joshua Jackson Nour Kteily Eli Finkel Curtis Puryear Meriel Doyle Mark Torres Abdo Elnakouri Trevor Spelman Jake Teeny Victoria Parker.