This year, Stanford University organized Agents4Science, the first open conference to accept papers written entirely by artificial intelligence. 

The conference’s mission was to serve as a “sandbox” to explore the future of AI-driven scientific discovery and review processes. Among the accepted were three papers written by Liner, an AI company that draws from over 200 million journals and ranked first in SimpleQA’s Benchmark for the world’s most accurate search engine in March of 2025. 

According to Kyum Kim, head of Liner’s United States operations, about 1% of Texas A&M students and faculty use their large language model, or LLM, a machine learning AI program that draws from massive data sets to understand and generate human-like text. 

Like most LLMs, Liner has free and paid options for its users. Its main features include agents — tools that automate tasks — targeted for literature review searches, citation recommendations and even peer reviews. 

“AI can augment the process of finding information and knowledge creation, but it should be done by humans at the end of the day,” Kim said. “Liner is filling a gap by drawing from trustworthy sources that give people accurate information about what’s out there. From there, our users are creating new research with it — that’s what we’re built for.”

A major challenge in using broad-spectrum LLMs, like ChatGPT or Gemini, for academic research is that they often “hallucinate.” Hallucination refers to the generation of fabricated information that AI draws from limited, unverified sources. 

For example, Kim said that if other LLMs are prompted to write a paper about a highly specialized research topic, they sometimes generate false references to “fill in the gaps,” thus undermining the LLM’s credibility.  

A way to reduce this problem is retrieval-augmented generation. Nuclear engineering graduate student Zaid Abulawi described the process in terms of mathematical vectors.

“Let’s say you have a book with three chapters: one about birds, the second on lions and the third about cooking,” Abulawi said. “AI will convert the information in those chapters into vector equations and store them in a ‘space.’ Because birds and lions are both animals, their information will be stored close together and within easy reach, with cooking further away. AI hallucinates when it reaches for information in a specific vector format that is not readily available, and instead draws on the closest similar vector — often generating false information.”

Assistant Professor of Computer Science and Engineering Tomer Galanti, Ph.D., said he’s seen peer-reviewed journals publish papers later found with false, AI-generated citations. These instances usually result from a game of authorship and ego, or when scientists hide the use of AI to receive full credit, Galanti explained. 

“The future of AI is here,” Galanti said. “Everyone, including universities, is going to adopt it. It’s inevitable, and we need to be thinking about how to deal with it responsibly. … I think that if you discover something with a simple AI prompt, it is still your scientific contribution. AI isn’t interested in credit. But people still need to disclose their use of it as a tool, and that’s where hidden information and bruised egos can come into play.”

Abulawi’s dissertation utilizes machine learning to model nuclear reactors, but he also builds LLM agents. 

“I used to code from scratch,” Abulawi said. “I know every line of code and how it is supposed to work, but using AI makes the process much more efficient. What used to take me weeks now takes a couple of hours.” 

As a graduate student who uses ChatGPT almost every day, he also emphasized the importance of treating AI solely as a learning tool, echoing Galanti. 

“If you want to do something right, you will take the hard path and learn it first,” Abulawi said. “It can be tempting to turn to AI when you’re on a deadline and so on, but as researchers and students, we must not only rely on these models.”

According to a 2025 study by the Pew Research Center, the number of teens using AI for schoolwork has doubled since 2023. Galanti said he is most concerned about students studying critical STEM topics, like healthcare or engineering, who use AI as the backbone of their education.

Galanti also said he often sees students present material or figures that are completely AI-generated in his classes. To challenge them, he asks his students to explain the complex metrics of their findings, actively facilitating learning and encouraging his students to dig into the material for themselves. 

“I think every piece of technology should be used with a grain of salt,” Galanti said. “This is where universities should apply unique measures to address the ways AI could be used by both students and researchers. People will try to cut corners and find shortcuts, and we need to be ready as educators.”

In writing-based disciplines, AI has sparked issues with plagiarism and fully-generated content. Science & technology journalism graduate student Grace Cote, who serves as a teaching assistant and co-instructor for biomedical writing courses at A&M, uses AI to search dense literature and identify connections between research topics. 

But among her students, she sees AI take the form of a ghostwriter. 

“AI can be a double-edged sword when it comes to writing,” Cote said. “Oftentimes, it can give you nice phrasing and even polish a sentence, but it can’t mimic the true human experience. Especially for assignments like literature reviews or medical narratives, generated writing comes off as cold, detached and relatively easy to spot.”

Of the instructors who teach the required writing courses for A&M’s future medical professionals, Cote said that she and her peers have had to send many students to the Aggie Honor System for abusing AI. 

She compared the situation to 1920s prohibition: If universities try to ban AI use, students will still find dishonest ways to incorporate it into their writing. Instead, Cote said educators need to teach them how to use it responsibly.

“When things get overwhelming or when we approach burnout, AI can lighten the load by helping with mundane tasks and make parts of our day easier,” Cote said. “But it should ultimately be used as a partner for clarity, not ghostwriting. Nothing can replace the touch of a human.”​​

As A&M works to integrate LLMs like Liner, ChatGPT and Gemini into its classrooms and research labs, Abulawi stresses the importance of utilizing programs that will give accurate results.

“Large language models are really cool; they can surprise you with their intelligence but also with how dumb they are,” Abulawi said. “Until now, we were the only ones guiding the thinking process. And now, we have a complex system that aids in that. But I don’t want to live in a world where everything is AI-generated — every tool has advantages and disadvantages. We should be critical of how we are using it, especially when it comes to conducting trustworthy research.”