At a recent gathering of the Association for the Advancement of Artificial Intelligence (AAAI), computer scientist Odd Erik Gundersen presented findings that underscore a growing concern in the field: the vast majority of AI research cannot be independently verified. According to a report covered by Science magazine, only 6% of 400 algorithms presented at two AI conferences in recent years included the source code necessary for replication. This scarcity of transparency poses a fundamental challenge to the scientific method, which demands that experiments be reproducible under identical conditions.
The implications extend beyond academic integrity. Without access to code, other researchers cannot test whether an AI system performs consistently across different hardware or datasets. This variability is particularly pronounced in machine learning, where algorithms learn from experience. Nan Rosemary Ke, a Ph.D. student at the University of Montreal, explained to Science that randomness can lead to misleadingly favorable results. “There’s randomness from one run to another,” she said. “[You can get] really, really lucky and have one run with a really good number. That’s usually what people report.”
The lack of reproducibility is not a new problem, but its prevalence in AI is striking. The report found that only one in three researchers shared the data used to test their programs, and just half provided a summary of the algorithm, known as pseudocode. This incomplete disclosure makes it nearly impossible for peers to validate claims or build upon prior work.
Why Sharing Is Hard
Several factors contribute to the reluctance to share code. Some algorithms are proprietary, owned by companies that guard their intellectual property. Others depend on unpublished underlying code, making it difficult to release a standalone version. Researchers may also withhold code out of fear of competition, or simply because the code is lost—stored on a broken laptop, a misplaced disk, or even destroyed by a pet, as one anecdote suggests.
Gundersen emphasized that the push for transparency is not about assigning blame. “It’s not about shaming,” he told Science. “It’s just about being honest.” His call for change reflects a broader movement within the scientific community to address the “replication crisis” that has affected fields from psychology to medicine.
Real-World Consequences
The stakes are high. AI systems are increasingly deployed in critical roles, from smartphone search to nuclear reactor management. If an algorithm behaves unpredictably when transferred to a new environment, the consequences could be severe. Trust in AI, already fragile, depends on the ability to verify that these systems perform as intended.
As the field continues to expand, the demand for reliable AI grows. The inability to replicate results not only undermines scientific progress but also erodes public confidence. For AI to fulfill its potential in society, researchers must embrace a culture of openness, sharing both code and data to ensure that every algorithm can be tested, validated, and trusted.