AI in Publishing: what did we learn from the panel session?

UNSILO hosted two panel sessions about AI in academic publishing at this year’s Frankfurt Book Fair.

As Geoffrey Bilder of Crossref, one of the panellists, memorably stated, “I had resolved never to appear on a panel session about AI or blockchain again!”. But he did appear, and made a great contribution to this panel session.

The two sessions were chaired by Toby Green, formerly head of publishing at the OECD. Each panel comprised three participants, and there were also suggestions and comments from the floor. The panel was loosely based around a survey carried out by UNSILO on the take-up of AI and attitudes towards it by publishers. As Toby Green pointed out in his introduction, most publishers were using AI (or planning to use it) to save money or time with their in-house processes. While this undoubtedly provided some clear benefits, it was clear that most publishers weren’t planning to use AI to add new functionality for the end user.

Nonetheless, there were some impressive descriptions of workflow improvements. Dave Flanagan gave some fascinating examples of how Wiley had been able to improve the quality of the data in some of their chemistry journal articles using AI tools. At the other end of the corporate spectrum, Josh Nicholson, from, with only a handful of staff, gave another quick win for AI. Their model is simple: they take thousands of academic articles and identify citations that support or refute a hypothesis. Using these manually annotated examples as a training set, they then apply these to a test set and measure their results. Once the system has been trained, it is able to identify confirmations and refutations in new articles submitted to it, and has now been rolled out to identify some 400 million citations from the published literature. The way Josh Nicholson described this process, in his engaging and straightforward manner, it seemed like simplicity itself, although I suspect quite a lot of skill is involved to get high-quality results.

Some themes emerged very clearly; one was the need for a human intermediate to manage and curate the AI. Another theme was the contrast between large and small companies. Large publishers, such as Wiley, will have their own internal departments developing AI tools, while a startup such as concentrates on building just one or two AI-based tools, which publishers then deploy. Small- and medium-sized publishers, such as BMJ, Helen King pointed out, have to make use of external partners to create AI tools. This appears to be an increasingly common arrangement; Smaller publishers choose API-based tools that bolt on to their existing platform and/or workflow.

The second panel, about trust and bias, also had some examples of small-scale wins: Geoffrey Bilder described some utilities now in use that improve the matching of references within Crossref by some 35%. When it came to trust and bias, there was less talk of quick wins. However, Leslie McIntosh pointed out that plagiarism can in some circumstances be a good thing. In the ‘methods’ section of an academic paper, authors are encouraged to write in a style and layout that is almost identical to the equivalent methods section of thousands of other papers. If such similarities occurred in the abstract or the main body of the article, that might read like plagiarism. One worrying finding from the survey was that fewer than 10% of publishers using AI tools were checking for bias. Such a figure seems remarkably low compared with the number of times ‘bias’ is mentioned in an AI discussion today.

When asked for what they wanted to see happening in the next five years, Angela Cochran asked for a single format for references (Endnote currently has around 6,000 output styles, so there is plenty of scope for simplification). At the same time, she pointed out that we are ‘abused’ by the lack of trust shown by some of the biggest Internet corporations, such as Google and Facebook, which has the consequence of making us distrust AI-based solutions Finally, Leslie McIntosh offered a more optimistic closing idea: we should build AI tools from those parts of the workflow that we trust, such as the peer review process, and then build on that. It was an encouraging tone on which to close.

Receive an email every time we publish a new blog post