Future ready: Elsevier’s AI in higher education newsletter

Increasingly, researchers are turning to AI tools for writing support, although this has led to concerns around research integrity, the homogenization of academic writing and risks of excessive cognitive offloading, especially among less experienced PhD students and postdoctoral researchers. Moreover, most AI writing tools – both mainstream and academic – focus almost exclusively on improving language and style, making them more suited to the refinement of late-stage manuscripts. While editing is undoubtedly important, the drafting phase is far more fundamental to most researchers, as there is very little point in polishing weak reasoning.

Launched in June, the LeapSpace Writing Coach is the first tool to focus rigorously on this task. Built to institutional privacy and security standards – no user prompts or documents are used to train large language models (LLMs) – Writing Coach prompts authors to consider what evidence might be missing from their arguments, actively questions their claims and checks them against peer-reviewed literature in LeapSpace’s publisher-neutral knowledge base. Although the Writing Coach can help authors polish their language, it treats written communication as part of the thinking process rather than simply a means of presentation.

Given this fixation, it seems likely that Herbert would also have regarded AI as a particularly devious kind of cognitive trap. A primary driver for the development of modern AI platforms was, he might have argued, the pervasive issue of information overload – however, risks like hallucinations, bias and lack of transparency mean these tools could be worsening the problem they were built to solve. A particularly telling example is excessive cognitive offloading, where a growing body of evidence suggests that overreliance on the technology to deliver “answers” results in a decline in critical-thinking skills. It’s no wonder that 84% of researchers use AI tools, but only 22% trust them.

Elsevier’s LeapSpace was built to try to fill this trust gap and, along with several other research-oriented AI solutions, has been successful in managing hallucinations and curbing the more obvious forms of bias. Even in the area of cognitive offloading, there have been advances in embedding a "friction-by-design" principle, such as the way LeapSpace actively surfaces the contradictions implicit in an area of research, helping to stimulate critical engagement among users.

Although this is real progress, the journey to research-grade AI is, like any worthwhile quest, full of hidden dangers. "Knowing where the trap is – that's the first step in evading it" is the advice given to the young hero in Frank Herbert’s original Dune novel, and who are we to argue? What follows is a short review of some of the snares that still lie in wait for unwary AI providers.

A typology of traps

Source quality traps

• The Open Web Trap – The Open Web Trap refers to the practice of training AI tools on unverified content from the wider web rather than focusing solely on scholarly resources. While this approach has some value for generalist AI tools, for researchers it risks combining reliable and unreliable information in a dangerously unclear way. This can open the door to hallucinations and reputational risk, undermining confidence in AI responses. Even researchers who are fully aware of these issues must allow additional time for manually validating references when using mainstream open web tools.

• The Language/Geography Trap – Most AI tools are heavily biased towards English-language, Western-published research. Given the huge rise in research output from China, India and other non-Anglophone research nations (based on a 2022 surveyopens in new tab/window, Indonesia (23.45%), Saudi Arabia (18.01%), Malaysia (17.75%), Pakistan (15.87%), Colombia (15.65%), and Iran (16.27%) all have exceptionally high publication growth rates), AI tools may present an unrepresentative picture of the research landscape. As AI lowers the barriers to publication for non-English-speaking authors, this discrepancy may grow even further.

• The Citation Amplification Trap – A claim that originally rested on a single small study, later cited approvingly by several review articles, can appear robustly supported when that is not actually the case. While this problem predates AI tools, they can exacerbate it by conflating primary and secondary sources, relying on citation counts as a proxy for scientific robustness and creating the false appearance of consensus in poorly structured AI summaries.

Retrieval traps

• The Relevance Trap – AI tools focused on surfacing “relevant” results that are closely aligned with their user’s expectations risk overlooking useful insights they might not have anticipated. As Adrian Raudaschl, Elsevier’s Senior Director of Product Management, explains in his article on this topicopens in new tab/window, “a system that optimises for relevance gets better and better at giving users what they want, which trains those users to expect confirmation, which makes them worse at wanting what they need.” While it is simplistic to assume that unexpected insights are always more innovative, many AI tools risk failing their users by being too faithful for their wishes. Like the ideal present, the perfect AI summary may entail giving a user something they didn’t know they wanted.

• The Consensus Trap – Related to the Relevance Trap, AI tools that succumb to the Consensus Trap tend to reproduce the mainstream view of a field, reflecting the bulk of the literature. As a result, minority positions and emerging challenges to orthodoxy are statistically suppressed. This can misrepresent fields where the consensus is contested and reinforce the status quo, potentially stifling new thinking.

• The Recency Trap – AI platforms trained on, or oriented towards, recently indexed literature may systematically undervalue foundational or slow-burn research. Conversely, platforms that put too much weight on citation counts may surface well-cited but superseded findings.

Interpretive traps

• The Fluency Trap – The written fluency of AI responses can create a false impression of authority. For example, an AI output might combine a range of different perspectives in a way that conveys confidence, but underrepresents specific points of view, in order to create a coherent narrative. While in the human world it might make some sense to use linguistic proficiency as an indicator of knowledge, perhaps as a marker of education, this assumption does not hold true for AI-generated content.

• The Feedback Loop Trap – AI tools shape what researchers read, then researchers use this content as the basis of their papers, sometimes with the support of built-in writing assistants. Once published, these same papers might be ingested back into the training data of the AI platform(s). The risk with this feedback cycle is that it is likely to reinforce existing research directions and gradually narrow the intellectual diversity of a field. While this is a longstanding problem – back in 2008opens in new tab/window, James Evans showed that the shift to online publishing and search made scientists more likely to read and cite the same highly visible papers – AI is accelerating the feedback loop.

Escaping the traps

Like most powerful traps, the ones documented above are hard to see, while many have characteristics analogous to defence mechanisms that help ensure their continuity. For example, the Consensus Trap produces outputs that look authoritative because most of the field agrees, so breaking out of it requires the confidence to dissent from the norm. Similarly, the Relevance Trap is disguised as good service, so AI providers that confront it risk confusing and upsetting their users.

The question then becomes, “How much cognitive friction will researchers tolerate?” or even “Who decides what is ‘good’ for researchers anyway?” While more disruptive insights may be beneficial in the long run, they are by definition difficult to assimilate and may not offer the same short-term rewards as incremental advances that can be written up for repeated publication in highly cited journals – the result of cultural factors which many AI providers may already have helped shape.

Some traps are easier to overcome than others. The Open Web Trap can be resolved simply by focusing exclusively on scholarly sources, although some providers are reluctant to do this because it limits the potential user base of their tools. With the Fluency Trap, easily digestible prose that sacrifices scientific nuance to narrative flow can be successfully broken down into a series of more representative bullet points – a technique used in LeapSpace Deep Research – without adversely affecting the average user’s cognitive performance.

Similarly, the Recency Trap requires providers to make a judgment call about the age of content incorporated in their summaries, with many opting for a small bias towards recency to better reflect the current state of knowledge. LeapSpace falls into this category but runs a second search to identify highly influential papers across a longer publication date horizon.

The Language/Geography Trap also requires strategic decision-making, with providers determining how to prioritize the goal of regional coverage against quality measures (like publication in highly cited journals) that typically favor established research nations. Content coverage is also important here, with LeapSpace drawing on abstracts from Scopus, which has long had broader non-English coverage than similar bibliographic databases opens in new tab/window, as well as broader publisher-neutral full-text coverage.

Some traps present even more fundamental problems. While it is difficult for any successful AI platform to completely evade the Feedback Loop Trap, some of the links in the feedback chain can still be broken. For example, LeapSpace’s commitment not to train its models on user’s search queries means reading behavior and query patterns cannot directly shape the system's future recommendations. This supports user privacy but also protects the openness of the wider research ecosystem. At the same time, the tool’s emphasis on surfacing what is missing rather than just amplifying what is popular, consciously promotes intellectual diversity. For all this useful friction, the ultimate problem is that a systemic version of this trap operates at the level of the published literature itself.

The Citation Amplification Trap is also challenging, although a solution at the tool level appears to be within reach. Currently, the LeapSpace Trust Cards feature helps by anchoring assertions in AI summaries to specific passages in primary sources, which is especially useful if a response is based entirely on review papers. However, while this is industry standard, true evasion of the Citation Amplification Trap would require the Cards to identify how many citations of a claim are independently derived from primary research and how many are simply reproducing citations through review papers. In other words, Trust Cards that currently link to the cited source would also need to link to the original underlying source.

The Trust Gap

While the arrival of mainstream AI has, quite rightly, triggered intensive discussions around bias, the irony is that bias is already everywhere in the research ecosystem. Above, we discussed how the Feedback Trap has its origins in online search behaviors, while the peer-review process itself has well-documented conservatism biases that reinforce themselves over time. Then there is publication bias (journals favor positive results), citation bias (researchers disproportionately cite work that confirms their hypotheses), prestige bias (researchers at elite institutions receive disproportionate citations, grants and journal acceptances) and all kinds of bias around gender, language and geography.

The fact that the best research AI tools rest heavily on this foundation also means their epistemic approach is shaped by decades of imperfect decision-making. In fact, AI tools take these existing problems and amplify them with the breadth, speed and convenience of their functionality.

Of course, there is also a long tradition of thought that says bias is structurally inevitable in any quest for precision – and academic research is probably the ultimate quest for precision – so perhaps we should simply accept it. This may seem a little depressing but, as we have seen above, the development of high-quality research AI tools still provides an opportunity to address some of these problems, even if they come with their own cognitive costs. What matters is that these trade-offs and costs should be visible, with benefits like speed and broad coverage underpinned by a commitment to epistemic honesty – a frankness about limitations as well as strengths. In other words, the best way to get out of the trap may actually be to stop struggling.

When someone says “84% of researchers use AI tools but only 22% trust them,” they are usually making a case for better AI tools. While this is reasonable, it is worth reflecting that there probably should be a trust gap, just as there is for other tools. For example, if you want to use an axe, then you need to be sure it can be used safely, that it is well-suited to the job, that the head is fitted properly and that it does not have any unexpected properties. It is this very wariness that enables you to use the axe to its best advantage and chop wood successfully.

Exactly how big the trust gap should be will depend on the task being undertaken and the background of the user, but in a world of imperfect tools, habitual caution is no bad thing.

The Knowledge Trap

The information overload problem has been around for a long time, but repeated attempts to manage it – online journal platforms, bibliographic databases, analytics, research data and AI – have each had their own strengths and weaknesses.

This is testament to both the magnitude of the challenge and the rapid expansion of scholarly literature, but there may also be an issue with knowledge itself: the Knowledge Trap.

The philosopher Karl Popper held that growing knowledge effectively expands human ignorance because every answered question generates several new ones – rather like walking towards an ever-broadening horizon.

It is possible that the tools through which we try to answer research questions also introduce new questions in an infuriating but strangely exhilarating way. What matters is curiosity, tenacity and integrity – an openness about our limitations and a robust combination of wariness and wonder. Frank Herbert put it far more succinctly: "Any path which narrows future possibilities may become a lethal trap." Look out!

However, as far as researchers are concerned, most policies still tend to focus on outputs (submitted work, manuscripts, theses) and say very little about the search and discovery part of their workflows. Guidelines often exhort researchers to "use AI responsibly", without specifying what this means or which tools meet institutional standards for data protection, citation integrity or transparency.

A research-grade AI model provides a practical lens for institutions and libraries to assess and implement AI tools that support research quality, integrity and responsible adoption. As described in the new article, Setting standards for AI: Why research-grade AI matters for institutions, “research-grade AI refers to AI systems purpose-built for research workflows, grounded in trusted content, designed with responsible functionality and built to support – not replace – human judgment.” The article goes on to outline how this approach can be applied across areas such as procurement, governance and policy, helping researchers, librarians and research managers to move beyond experimentation toward responsible, scalable innovation.

There is an striking contrast between an institutional view of AI as a workflow accelerator, potentially expanding the capacity of staff, and the first-hand feedback of users.

In another of the intriguing discrepancies surfaced by the report, the same academic leaders who are optimistic about AI’s ability to drive new knowledge express considerable pessimism about AI governance, with only 35% saying their institution has “good AI governance”.