Shruti Kumar (name changed) is a professor at the Institute of Medicine who works on diagnosing neglected tropical diseases that infect nearly 1 million people each year. Professor Kumar said that in recent years he has been receiving more requests from scientific publishers to review research papers, and because the process is so time-consuming, he has made a conscious decision to only accept requests related to specific subfields of his field.
Indeed, as STEM journals rapidly expand globally and the pressure to publish more papers increases, qualified experts to review papers are becoming increasingly scarce, and leading academic publishers are turning to artificial intelligence (AI) rather than their peers for help.
Approximately 15,000 years
Professor Kumar said that AI could help in detecting plagiarism. “Some publishers check the plagiarism rate before sending manuscripts to reviewers, saying for example that 40% plagiarism was detected with AI. This is useful information for reviewers and will save them time to do the same.”
In 2020, reviewers around the world spent about 130 million hours (the equivalent of about 15,000 years) on the review process, according to Diksha Gupta, director of global social strategy at the American Chemical Society. This places an immense burden on reviewers who must balance their peer review responsibilities with their own academic and research endeavors.
However, “the peer-reviewed journal system has not been able to adapt and provide an appropriate system that can take advantage of AI technology while controlling the downsides,” according to a recently published paper. innovationco-authored with Gautam Dejiraj, a structural chemist at the Indian Institute of Science in Bengaluru. The rate of annual data generation exceeds the number of academic papers published each year, the paper added.
There is no substitute for humans
The global scientific community needs to experiment with alternatives to the peer-reviewed journal process “to ensure that scientific productivity is not diminished by errors and oversights induced by AI-generated knowledge being accepted as scientific findings,” the authors added.
“Although AI cannot replace human reviewers or make final editorial decisions, it can play a valuable supplementary role,” Dr. Gupta said. “AI-integrated tools help accurately match manuscripts with appropriate subject matter experts and provide a preliminary assessment at the pre-review stage. This ensures that only submissions of sufficient quality and relevance advance to full peer review, reducing unnecessary workload for reviewers.”
AI can meaningfully support peer review, but only when leveraged responsibly and for limited tasks, said Shane Rydquist, vice president of delivery and solutions at Cactus Communications Ltd. He added that AI will be able to support researchers’ workflows, managing routine tasks such as literature searches and data organization, identifying subtle patterns in complex data, and uncovering unexpected connections between distant fields that humans would never encounter.
In fact, AI is good at detecting text similarities, image manipulation, and patterns of data fabrication, so AI’s role should be to augment human expertise, for example in plagiarism and integrity checks, Dr. Rydequist said. It also helps with screening, assessing submission quality, formatting compliance, adjusting scope, analyzing expertise, identifying potential reviewers based on publication history, detecting bias by flagging potentially problematic or biased language, and identifying patterns of conflicts of interest.
Dr. Lidquist added, “The key is augmentation, not replacement.”
amplification risk
That said, he added, human judgment is still needed to “assess conceptual novelty and significance, evaluate contextual methodological soundness, make nuanced judgments about appropriateness for the journal’s audience, and provide constructive feedback that advances the science.”
From a publisher’s perspective, the integration of such AI systems is still in the development stage, according to Dr. Gupta, and “rigorous testing and validation is essential before these tools can be deployed at scale.”
However, there is one concern that arises. innovation Papers run the risk of amplifying mistakes that creep into machine-generated summaries, allowing future authors to cite and spread “fundamentally incorrect or misunderstood science.”
This “inevitably reflects papers that cannot be replicated, but in most cases there is no clear indication,” the authors write.
“It’s not the only reliable source of information.”
AI models also have biases that are difficult to understand and control, such as biases due to inclusion/exclusion choices in datasets, assumptions of algorithmic processes, and socio-economic factors built into the operating institutions developing the AI.
“Designing a system to counter this is a difficult and ongoing undertaking, and the current peer-reviewed journal system appears ill-equipped to do so,” the authors added.
“We have already encountered cases where people have used generative AI to spread false citations. For example, large-scale language models (LLMs) may generate seemingly plausible but non-existent references, which can create misleading chains of evidence,” Dr. Rydquist said.
When using generative AI platforms and tools, Dr. Rydequist continued, people tend to overlook subtle technical errors that a human expert would immediately notice. “And LLM tends to overrepresent highly cited but potentially flawed studies and underrepresent new and corrective studies. This is why human oversight remains essential. Without critical evaluation, AI can rapidly accelerate misinformation.”
December 18th newspaper science He suggested that despite both excitement and concern about the use of AI in academia, “empirical evidence remains fragmented” and the impact of LLM is not fully understood. The paper said that as LLMs begin to “reshape scientific production”, the importance of English fluency will take a backseat, “but the importance of robust quality assessment frameworks and methodological scrutiny will be paramount.”
“AI, at its current stage of development, will not be the only source of truth for reference and decision-making,” Dr. Gupta said.
“Clearly a human domain”
He added that with the evolution of intelligent machines, the intelligence of humans, especially experts and scientists, will become even more valuable, so we need to be careful and deliberate about how to use these tools wisely, without getting biased towards concepts and fundamentals.
Dr. Gupta explained that a simple and effective strategy to minimize errors in AI-driven data synthesis is to avoid relying on a single model. Using models in combination provides more balanced and accurate results.
“Equally important is ensuring that the source dataset comes from a genuine and trusted database.”
Several technical safeguards can further minimize AI-related errors, Dr. Rydquist added. For example, he said, always verify AI-generated citations and summaries of data against primary sources.
So, does AI help or hinder creativity? First, AI can only make incremental discoveries. However, they cannot achieve fundamental discoveries to generate truly original hypotheses like humans.
In fact, Dr. Rydquist says, AI may inadvertently limit creativity and lateral thinking. “Taking deeper into a problem often yields insights, but AI shortcuts can rob scientists of this generative friction.”
However, the key difference is that while AI is good at solving well-defined problems within established frameworks, “true creativity often involves reframing the problem itself or questioning basic assumptions, which is still clearly the domain of humans,” he added.
“The challenge is not just to embrace technological advances, but to maintain the human spirit that fosters true innovation,” Dr. Gupta said.
TV Padma is a science journalist based in New Delhi.
