A single particular person claims to have authored 113 educational papers on artificial intelligence this 12 months, 89 of which shall be offered this week at considered one of the world’s main convention on AI and machine studying, which has raised questions amongst laptop scientists about the state of AI analysis.
The writer, Kevin Zhu, lately finished a bachelor’s diploma in laptop science at the College of California, Berkeley, and now runs Algoverse, an AI analysis and mentoring firm for top schoolers – lots of whom are his co-authors on the papers. Zhu himself graduated from highschool in 2018.
Papers he has put out in the previous two years cowl topics like utilizing AI to locate nomadic pastoralists in sub-Saharan Africa, to evaluate skin lesions, and to translate Indonesian dialects. On his LinkedIn, he touts publishing “100+ prime convention papers in the previous 12 months”, which have been “cited by OpenAI, Microsoft, Google, Stanford, MIT, Oxford and extra”.
Zhu’s papers are a “catastrophe”, mentioned Hany Farid, a professor of laptop science at Berkeley, in an interview. “I’m pretty satisfied that the entire factor, prime to backside, is simply vibe coding,” he mentioned, referring to the apply of utilizing AI to create software program.
Farid known as consideration to Zhu’s prolific publications in a current LinkedIn post, which provoked dialogue of different, comparable instances amongst AI researchers, who mentioned their newly common self-discipline faces a deluge of low-quality analysis papers, fueled by educational pressures and, in some instances, AI instruments.
In response to a question from the Guardian, Zhu mentioned that he had supervised the 131 papers, which had been “crew endeavors” run by his firm, Algoverse. The corporate prices $3,325 to high-school college students and undergraduates for a selective 12-week on-line mentoring expertise – which entails assist submitting work to conferences.
“At a minimal, I assist evaluation methodology and experimental design in proposals, and I learn and remark on full paper drafts before submission,” he mentioned, including that initiatives on topics similar to linguistics, healthcare or schooling concerned “principal investigators or mentors with related experience”.
The groups used “commonplace productiveness instruments similar to reference managers, spellcheck, and typically language fashions for copy-editing or bettering readability”, he mentioned in response to a question about whether or not the papers had been written with AI.
Bot watchers in turmoil
The evaluation requirements for AI analysis differ from most different scientific fields. Most work in AI and machine studying does not go endure the stringent peer-review processes of fields similar to chemistry and biology – as a substitute, papers are usually offered much less formally at main conferences similar to NeurIPS, considered one of the world’s prime machine studying and AI gatherings, the place Zhu is slated to current.
Zhu’s case factors at a bigger subject in AI analysis, mentioned Farid. Conferences together with NeurIPS are being overwhelmed with rising numbers of submissions: NeurIPS fielded 21,575 papers this 12 months, up from underneath 10,000 in 2020. One other prime AI convention, the Worldwide Convention on Studying Representations (ICLR), reported a 70% improve in its yearly submissions for 2026’s convention, practically 20,000 papers, up from simply over 11,000 for the 2025 convention.
“Reviewers are complaining about the poor high quality of the papers, even suspecting that some are AI-generated. Why has this educational feast misplaced its taste?” requested the Chinese language tech weblog 36Kr in a November post about ICLR, noting that the common rating reviewers had awarded papers had declined year-over-year.
In the meantime, college students and lecturers are dealing with mounting strain to rack up publications and sustain with their friends. It is unusual to produce a double-digit quantity – a lot much less triple – of top of the range educational laptop science papers in a 12 months, lecturers mentioned. Farid says that at occasions, his college students have “vibe coded” papers to up their publication counts.
“So many younger folks need to get into AI. There’s a frenzy proper now,” mentioned Farid.
NeurIPS evaluations papers submitted to it, however its course of is far faster and fewer thorough than commonplace scientific peer evaluation, mentioned Jeffrey Walling, an affiliate professor at Virginia Tech. This 12 months, the convention has used giant numbers of PhD college students to vet papers, which a NeurIPS space chair mentioned compromised the course of.
“The truth is that usually occasions convention referees should evaluation dozens of papers in a brief time frame, and there is often little to no revision,” mentioned Walling.
Walling agreed with Farid that too many papers are being printed proper now, saying he’d encountered different authors with over 100 publications in a 12 months. “Lecturers are rewarded for publication quantity greater than high quality … Everybody loves the fantasy of tremendous productiveness,” he mentioned.
On Zhu’s Algoverse’s FAQ web page, solutions discusses how the firm’s program might help candidates’ future school or profession prospects, saying: “The abilities, accomplishments, and publications you obtain right here are extremely regarded in educational circles and may certainly strengthen your school utility or résumé. This is very true in case your analysis is admitted to a prime convention – a prestigious feat even for skilled researchers.”
Farid says that he now counsels college students to not go into AI analysis, due to the “frenzy” in the discipline and the giant quantity of low-quality work being put out by folks hoping to higher their profession prospects.
after e-newsletter promotion
“It’s only a mess. You possibly can’t sustain, you may’t publish, you may’t do good work, you may’t be considerate,” he mentioned.
Slop flood
A lot glorious work has nonetheless come out of this course of. Famously, Google’s paper on transformers, Attention Is All You Need – the theoretical foundation for the advances in AI that led to ChatGPT – was offered at NeurIPS in 2017.
NeurIPS organisers agree the convention is underneath strain. In a remark to the Guardian, a spokesperson mentioned that the progress of AI as a discipline had introduced “a big improve in paper submissions and heightened worth positioned on peer-reviewed acceptance at NeurIPS”, placing “appreciable pressure on our evaluation system”.
Zhu’s submissions had been largely to workshops inside NeurIPS, which have a distinct choice course of than the predominant convention and are usually the place early-career work will get offered, mentioned NeurIPS organisers. Farid mentioned he did not discover this a substantive rationalization for one particular person to put his title on greater than 100 papers.
“I don’t discover this a compelling argument for placing your title on 100 papers that you would not have probably meaningfully contributed to,” mentioned Farid.
The issue is larger than a flood of papers at NeurIPS. ICLR used AI to evaluation a big quantity submissions – leading to apparently hallucinated citations and suggestions that was “very verbose with a number of bullet factors”, in accordance to a current article in Nature.
The sensation of decline is so widespread that discovering an answer to the disaster has turn into the topic of papers itself. A May 2025 position paper – a tutorial, evidence-based model of a newspaper op-ed – authored by three South Korean laptop scientists that proposed an answer to the “unprecedented challenges with the surge of paper submissions, accompanied by rising considerations over evaluation high quality and reviewer duty”, gained an award for excellent work at the 2025 Worldwide Convention on Machine Studying.
In the meantime, says Farid, main tech corporations and small AI security organisations now dump their work on arXiv, a web site as soon as reserved for little-viewed preprints of math and physics papers, flooding the web with work that is offered as science – however is not topic to evaluation requirements.
The price of this, says Farid, is that it is virtually inconceivable to know what’s truly going on in AI – for journalists, the public, and even consultants in the discipline: “You haven’t any likelihood, no likelihood as a mean reader to strive to perceive what is going on in the scientific literature. Your signal-to-noise ratio is mainly one. I can barely go to these conferences and determine what the hell is going on.”
“What I inform college students is that, if what you’re making an attempt to optimize publishing papers, , it’s truly actually not that arduous to do. Simply do actually crappy low-quality work and bomb conferences with it. However if you’d like to do actually considerate, cautious work, you’re at a drawback since you’re successfully unilaterally disarmed,” he mentioned.
Disclaimer: This article is sourced from external platforms. OverBeta has not independently verified the information. Readers are advised to verify details before relying on them.
