Tackle ‘AI slop’ in education research ‘or lose teacher trust’

A sharp rise in “unmistakably large language model-written” phrases in education papers is threatening to destroy school teachers’ trust in university research, a high-profile sleuth has warned.

For the past nine months Stephen Vainker, a school teacher who holds a PhD from the University of Exeter, has used his Substack page to chronicle the growing prevalence of what he calls  “LLM slop” in papers published in the journals of the British Educational Research Association (Bera).

Highlighting suspected AI-written phrases in more than 100 papers in Bera-affiliated journals, Vainker’s analysis has drawn particular attention to about 100 previously rarely used phrases which, since the advent of generative AI in 2023, have become commonplace in the society’s publications.

Bera said it had launched an investigation into the allegations but research integrity teams had decided no further action was required over the majority of papers flagged.

Among the overused phrases are, Vainker said, “underscoring the intricate interplay between” which was used just 26 times in Bera journals before 2023 but has appeared 3,004 times since then, and “underscores the critical role of”, which was used 59 times pre-2023 but 3,111 times since that year.

“This approach allows for a more nuanced understanding of” has appeared 791 times in Bera literature since 2023, having featured just 54 times prior, he added.

Similarly, “underscoring the need for a nuanced” appeared just 17 times in Bera journals before 2023 – but 583 times afterwards, while “tailored to diverse cultural contexts” featured six times before 2023 but 94 times since then.

In most cases, the authors of papers with suspected undeclared AI use are from China, said Vainker who believes this trend should prompt a full-scale review of peer review and editorial practices at Bera-run journals.

Asked by Times Higher Education if the prevalence of AI-authored text among Chinese scholars might suggest they use it to improve readability rather than fraud, Vainker conceded this was “plausible for about a third of cases”.

“But there is a good proportion of these Bera papers that are just terrible,” he argued. “In one case the results section was only eight lines long, while some paragraphs will just repeat identical sentences,” he said, adding: “They are just drivel.”

“If this continues Bera and university education research more generally will lose legitimacy with teachers,” warned Vainker, whose sleuth work has recently triggered a University of Melbourne investigation into alleged plagiarism by John Hattie, one of the world’s most influential educationalists. Hattie has consistently denied any wrongdoing. 

“Even 10 years ago Bera journals could be relied on to publish some really good stuff,” continued Vainker, noting that “their critiques of the Michael Gove reforms were very sharp and research on formative assessment was excellent – I don’t think you could say the same now”.

“Deeper and broader critical thinking” on classroom teaching had given way to “very applied studies that don’t actually say anything useful”, he said.

He also questioned whether Bera policies to “increase the range and diversity of authors being published” in its journals, under its updated Race Equity Policy, might have inadvertently led reviewers to be more lenient to authors from underserved backgrounds.

“If reviewers have to balance quality concerns with other considerations, that is going to cause problems. But it’s naïve to think all these authors from China who’d never published before 2023 can now turn into multiple submissions written in perfect English,” he said.

“I’m not against the aims of this policy but it’s reckless to have these targets unless you also have a really robust AI policy to ensure papers are rigorous,” he said.

A Bera journals spokesperson told THE that it had launched an investigation, along with its publisher Wiley, into the allegations raised by Vainker. However, although “some of these concerns are being investigated further and the outcomes of those investigations are pending…“for a majority of the papers flagged, Wiley’s research integrity team did not find any reason to pursue further action”, they said.

“The Bera journals follow Wiley’s AI guidelines, which state that certain uses of AI require disclosure but not in all circumstances. Undisclosed AI use does not always indicate fraud or manipulation; however, when combined with other signals, it can indicate that a paper needs further examination.”

On whether its race equity policy would affect reviewer decisions, they added that “journals are editorially independent from Bera and under no circumstances would Bera pressure editors or reviewers to accept any papers”.