Artificial intelligence (AI) and large language models (LLMs) are now able to generate a large volume of academic finance papers that are almost identical to those written by humans, as revealed by a recent study.published in the Journal of Economic Literature.
Subscribe to our newsletterfor the most recent science and technology news updates.
Creating an AI research workflow
Study co-authors Mihail Velikov, an assistant professor of finance at Penn State's Smeal College of Business, and Robert Novy-Marx, the Lori and Alan S. Zekelman Distinguished Professor of Business Administration at the University of Rochester, created a system for streamlining academic research. Within about 12 hours, they produced almost 400 complete,publication-ready finance papers—employing artificial intelligence to develop hypotheses and compose the manuscript. The research illustrated how AI can speed up the creation of academic papers, while also highlighting worries regarding its possible effects on the scholarly community and the essence of scientific discovery.
AI is now capable of generating a large number of papers efficiently, and this will transform the way we create and share knowledge," Velikov stated. "This serves as an early indication of what's ahead with current AI advancements.
The scientists did not originally intend to develop an AI system for generating financial research papers. Velikov's academic work centers on discovering irregularities in the stock market—data patterns that deviate from established theoretical models of financial market behavior. He and Novy-Marx were engaged in a data analysis project, examining corporate financial records for possible indicators that could forecast which stocks would surpass market performance.
They discovered over 30,000 possible indicators. They tested the forecasting ability of each indicator, which involved checking them against 200 known irregularities mentioned in financial research. After this evaluation, they reduced the list to 95 signals that were genuinely new.
Velikov subsequently input the data into a website he created, which was capable of producing a template report following the analysis. The generated reports resembled academic papers that describe new anomalies. The only element that was absent was a hypothesis or explanation for the potential reasons behind these anomalies.
Allowing an AI language model to compose the essays
It was late 2023 when I realized that large language models could be highly effective at generating stories to account for these anomalies," Velikov stated. "A combination of data mining and large language models might result in a significant number of convincing scientific papers.
So, they gave it a try. The researchers employed Anthropic's LLM Claude Opus 4, which was the most recent version available then, to transform the template reports into scholarly articles, using the data from the data mining initiative. For each of the 95 signals, they directed the LLM to create descriptive titles for the predictors and generate four separate papers, each with a unique hypothesis and theoretical explanation to account for the findings related to the same signal.
In total, the researchers published 380 papers. Each paper contained an abstract, introduction, data, results, conclusion, and references. The AI-generated papers, along with the complete code used to create them, arepublicly available on GitHub.
Pressures on peer review and guidelines
The ability of AI to efficiently produce academic papers has sparked questions and concerns regarding academic research and the peer-review process, according to Velikov. Overall, the number of submissions to journals and conferences has increased significantly in recent years, placing a heavy burden on peer reviewers. As AI and large language models become more powerful and widely used, he noted that the scientific peer-review system must evolve to address these advancements.
Currently, with agent-based AI systems, this can be accomplished more effectively, and the papers are significantly improved," Velikov stated. "This will likely elevate the standards. It may also transform how we share and assess research.
HARKing dangers and the future of employment
The research pointed to another issue of worry, Velikov stated. In theAI-generated papers, the LLM developed the hypotheses once a pattern in the data had already been recognized—a method referred to as HARKing or forming hypotheses after the results are known. This is an established practice in academic settings, which Velikov noted is generally seen in a negative light, but AI alters the extent to which HARKing can happen. If AI provides justifications for findings in the data, it could lead to concerns about what qualifies as a scientific contribution, especially since AI can still produce hallucinations—when large language models create incorrect or deceptive information and present it as true.
Although the researchers concentrated on financial studies, they mentioned that the consequences of these results could apply to other areas.
I don't believe we'll all lose our jobs and be replaced by AI," Velikov stated. "However, I feel our roles will change significantly, and the more we focus on learning how these systems function, the more effective our research will become. The better we'll perform in our positions.
More information:Robert Novy-Marx et al., Artificial Intelligence-Driven (Finance) Research,Journal of Economic Literature (2026). DOI: 10.1257/jel.20251821
GitHub: github.com/velikov-mihail/AI-Powered-Scholarship
Supplied by Pennsylvania State University
This narrative was first released on.
No comments:
Post a Comment