The term “social media influencer” may call to mind Instagram accounts shilling hair growth gummies and cute outfits—but in reality, influencers influence all types of things. Including artificial intelligence research trends.
Mainstream interest in AI and machine learning is at an all-time high, and the industry is responding—churning out thousands of AI and ML works for conferences and journals. The AI/ML community is also particularly active in posting non-peer-reviewed preprints via online platforms like ArXiv. Given this glut of work, what rises to the top and receives attention?
The answer, at least in part, is: the research that a pair of highly influential users of X (formerly Twitter) choose to highlight, according to a new preprint from researchers at University of California at Santa Barbara.
The UCSB paper analyzed more than 8,000 AI and ML papers, considering both social media mentions and the number of citations. Reviewing tweets from December 2018 to October 2023, the researchers concluded AI/ML papers shared by two specific influencers had median citation counts two to three times higher than those of the control group.
This is crucial because academic citations aren’t just about recognition in one’s field; they also affect decisions like research funding and tenure at academic institutions. And it’s a change from the status quo. As recently as 2018, a study of conference papers showed that a paper’s review score—meaning acceptance to top conferences—was a primary indicator of future citation count.
Now, “the correlation between influencer tweets and citation count—and not review scores—points to a shift in how the community finds and reads papers,” this new work concludes.
Two influencers with an outsized effect on AI
The researchers selected two influencers as case studies. Both share AI/ML papers consistently and have a significant following on X (formerly Twitter): @_akhaliq and @arankomatsuzaki. “These influencers have emerged as pivotal figures in navigating the flood of information, akin to journalists in civic society, highlighting and contextualizing significant works for the community,” the authors write.
That curation is a helpful—and, of course, unpaid—service from these influencers given the deluge of research. But “an over-reliance on a select group of curators may inadvertently skew the research landscape, emphasizing certain topics or perspectives over others,” the researchers write. Inadvertent bias in sharing certain labs’ or researchers’ work may entrench a lack of geographic, gender, and/or institutional diversity, the paper adds.
Awareness is the first step to busting this social media echo chamber, says lead author Iain Xie Weissburg, a first-year master’s student in UCSB’s Electrical and Computer Engineering program.
“We wanted to help the community recognize this and be vigilant in ensuring that research remains an evenly leveled domain,” he tells Spectrum. “As it stands now we all tend to get our information from a select few, we conclude that these are the hot topics, and then we often select our research based on that hype.”
The point is not to shame or place undue responsibility on these influencers or others, Weissburg is careful to note. “It’s that the publication and conference systems need to adapt to the vast increase in the volume of AI/ML research, which we can see continuing for the foreseeable future—especially with the influx of generative AI in the public sphere,” Weissburg says.
The analysis highlights not only social media’s expanding influence in AI/ML research, but also the importance of an evolving ecosystem to bring diversity of thought to today’s digital academic landscape.
The volume of AI papers is overwhelming
The researchers’ selection of just two influencers is “far from perfect” methodology, says Delip Rao, an independent researcher affiliated with the University of Pennsylvania and University of California at Santa Cruz. He adds “these two folks tend to tweet out papers from big labs and famous names. So, it is not clear who is influencing whom.”
Still, he agrees with the overall conclusion that a small group of influencers have an outsize effect, “which is problematic for science.” Citation counts are meant to be driven by experts who deeply understand the work they’re citing and the context, he notes, adding that it’s “unrealistic” to expect even expert influencers to bring this type of rigorous review to the social media activity that’s outside of their day jobs.
That’s especially true given the proliferation of AI/ML research as of late, as one of the two featured influencers acknowledged recently. “I hate when arXiv casually releases like 500 papers almost every day,” Aran Komatsuzaki wrote on X.
The “sheer volume of papers published daily is overwhelming, making it impractical for individuals to sift through arXiv feeds,” Komatsuzaki, the chief technology officer of Teraflop.ai, tells Spectrum in an interview. “The community relies on curators like @_akhaliq and myself to highlight a selection of noteworthy papers each day…. [I’m] cautious about not promoting research with unconvincing, weak, or dubious results.”
The other influencer cited, Ahsen Khaliq or “AK,” is a machine learning engineer at Hugging Face. “I think there is some shift in the community toward finding/discovering new research and citing it through Twitter or other social media rather than conferences/reviewer scores—although each has their own place in the community,” Khaliq says in an interview.
Can the AI social media bubble be popped?
The paper’s core conclusion, Weissburg says, is that researchers, conference organizers, and academic institutions must be aware of the shifting norms as preprint platforms and social media accounts change the landscape of sharing research—especially in the AI/ML space. The authors also argue that conferences and peer-review processes may have to evolve to ensure that quality research is effectively distributed.
Weissburg says that engineers working in the field should at times resist the urge to hop on the hype train, lest they ignore other important areas of research. Influencers, too, can take note that “paper-sharing isn’t a zero-sum game. You have an effect, and you don’t necessarily always have to share the biggest companies and most famous researchers. It’s important we have a diverse community in terms of different ideas and different backgrounds.”
In future work, Weissburg hopes to explore possible outsize social-media effects in other areas of science. There’s also an opportunity to explore the underlying mechanisms of social media—like how algorithms surface content to users—as they relate to academic recognition.
Rao would like to see new ideas for new ideas for surfacing and disseminating quality papers. “When there is excessive production, it is natural to rely on curators, and the community’s reliance on such folks is a cry for help,” he says. “We need better ways to combat this information deluge, and the answer is, hopefully, not influencers.”