Research productivity is soaring as AI tools become mainstream, but 80% of researchers depend on general-purpose chatbots like ChatGPT, while only a quarter have tried AI assistants specifically built for academic work, according to Wiley's 2025 ExplanAItions survey. Faulty citations, hallucinated references, and rising error concerns are forcing a rethink of the tools researchers reach for first.
The rapid rise-and growing pains-of AI in research
Wiley's survey of over 2,400 researchers found AI usage jumped from 57% to 84% in a single year, with use for research and publication tasks climbing from 45% to 62%. A December 2025 Science study analyzing nearly 2.1 million abstracts reported that AI-assisted authors published significantly more papers. In the social sciences and humanities, preprint output rose by nearly 60%, and for Asia-based researchers writing in English as a second language, gains reached 89% at some institutions. Yet the same study cautioned that AI-generated writing often used polished, complex language that could mask weaker research ideas, and such papers were less likely to pass editorial review.
Despite the productivity boost, confidence is slipping. Concerns about AI-generated errors and hallucinations rose from 51% to 64% year-on-year in the Wiley data, while security and privacy worries grew from 47% to 58%. A year ago, researchers believed AI outperformed humans on more than half of relevant tasks; that figure has fallen below one-third. As active use replaces early excitement, AI tools are being tested against real academic workflows, and the results are mixed.
Why generic AI stumbles on scholarly work
General-purpose tools aim for breadth, not depth. They generate fluent text across countless topics, but scholarly communication demands precision: every citation must point to a real, verifiable source. That is where the mismatch becomes costly. A paper in the Journal of Dental Sciences highlighted that earlier mainstream models fabricated 51% of citations, and even newer versions like GPT-4 still produce hallucinations at rates around 18%. Another analysis presented at NeurIPS 2025 found hallucinated citations in at least 53 published research papers that had passed peer review.
The problem reflects how these systems generate responses from pattern-matching across training data, not from curated academic databases. For a researcher, a fabricated citation can damage credibility and slow career progression. This gap is especially dangerous for authors writing in a second language or those at institutions without strong editorial support, who may not catch the errors. Instead, dedicated Research tools built from the ground up for academic writing draw on verified databases and follow scholarly conventions.
The case for research-specific AI
Unlike general-purpose chatbots, AI built for scholarly work pulls from curated research databases and understands how research arguments are structured, how claims are supported, and how different fields evaluate evidence. Solutions like Paperpal, developed by Cactus Communications, incorporate AI disclosure requirements that journals and funders now mandate. A 12-month CACTUS study on Wolters Kluwer's journal Medicine found that manuscripts pre-screened by journal AI tools were accepted at rates up to 27% higher than those that were not.
For researchers navigating a new discipline, writing in English as a second language, or working at institutions with limited editorial infrastructure, the right tool can mean the difference between full participation in scholarly conversation and submission marred by preventable errors. Professionals looking to integrate such tools responsibly can benefit from structured AI for Science & Research guidance that clarifies how to evaluate and adopt research-specific assistants.
The institutional vacuum
Universities are struggling to keep up. A Frontiers study of AI policies across 40 institutions found that while every university emphasized academic integrity and ethical AI use, guidelines remain fragmented. An Elsevier survey of more than 3,200 researchers across 113 countries revealed that only 32% said their institution had clearly defined AI rules. Meanwhile, 73% of researchers in the Wiley survey want publishers to issue clearer guidance on responsible AI use.
Without well-established guardrails, the burden falls on individual researchers, many of whom face intense pressure to publish and lack clear directives. Editorial teams are left to handle growing submission volumes with insufficient screening tools. The path forward involves building transparency and compliance into the systems that support research, rather than leaving researchers to navigate the ambiguity alone.
Why this matters for science and research professionals
Using AI in research workflows is no longer optional, but which AI you choose directly affects publication success and professional reputation. General-purpose tools may speed up drafting, yet they introduce citation risks that research-specific assistants are designed to mitigate. Science and research professionals should evaluate any AI tool by three criteria: can it trace outputs to real, verifiable sources, does it flag uncertainty instead of inventing plausible text, and is it compatible with funder and journal disclosure standards. As acceptance rates and citation accuracy can make or break a career, selecting the right AI tool becomes a core research skill.
Your membership also unlocks:
