A recent study released by an interdisciplinary team led by computer scientists Arvind Narayanan and Sayash Kapoor from Princeton University brings into sharp focus the potential harm that AI could do to scientific research. The researchers argue that the lack of properly outlined best practices in using machine learning within scientific fields is threatening the reproducibility of scientific results and, consequently, the very principles that underpin scientific processes.
Machine learning has been widely adopted across all scientific fields. However, according to the researchers, the rapid growth has not been met with a corresponding development of standards to ensure validity and reproducibility of results. They claim that thousands of studies using flawed machine learning methods have been published, which may cripple scientific research if not addressed timely.
As a solution, the team proposes a checklist of best practices, dubbed REFORMS (Recommendations for Machine-learning-based Science). The checklist, consisting of 32 questions across eight areas, provides guidelines on all aspects of a study, including stating study goals, computational reproducibility, data quality, data preprocessing, modeling, and data leakage.
Already, AI’s integration into scientific research has indicated various benefits. A survey from Nature showed that 66% of academics believed AI enables quicker data processing, 58% said it enhances computation, and 55% reported that AI saves time and money. However, concerns were also raised, with 53% of the academics worrying about replicability of results, 58% about bias, and 55% about AI enabling fraudulent research.
Without proper checks, even the most well-meaning of researchers can misuse AI, leading to “nonsense” scientific outputs. Narayanan and Kapoor’s guidelines aim to hold individuals accountable for their use of AI in scientific research. However, implementing these standards may be challenging due to the underreported nature of the reproducibility crisis, and the need for consensus among researchers, reviewers, and journals. Therefore, it is crucial for the scientific community to recognize this growing issue and provide shared guidelines to regulate the use of AI in scientific research.
