When can we say AI made a scientific discovery?

In September 2026, Anthropic, an AI company, announced that its AI system, called Claude, had made a discovery in molecular biology. The system operates within a lab where 950 AI agents analyze vast databases of DNA sequences to identify patterns that might lead to scientific breakthroughs. For example, these agents might flag unusual sequences that could encode enzymes with useful properties. Anthropic claimed that after 21 hours of analysis, the AI identified a repeating pattern around a known enzyme, a pattern Anthropic stated had not been catalogued before. This pattern was described as being reminiscent of the structure that led to CRISPR, a revolutionary gene-editing technology. However, the announcement sparked debate about whether this constituted a genuine scientific discovery or merely an assistive tool for human scientists.

The Anthropic announcement highlighted a key issue: defining what qualifies as an AI-made scientific discovery. While the AI identified a new pattern in DNA sequences, some biologists argued that this alone does not constitute a discovery. Lucas Harrington, a biologist, noted that the real discovery lies in understanding the function of such patterns. For instance, finding a cluster of genes is often the first step, but determining what the system does is far more challenging and scientifically meaningful. This debate underscores a broader tension: AI companies frequently present their systems as capable of making discoveries independently, whereas traditional scientific progress typically emerges from collaboration and iterative experimentation. The article suggests that the standard for what counts as a discovery may be too high or too vague when applied to AI.

Further complicating the issue, a biologist named Mario Rodríguez Mestre from the University of Copenhagen stated in late September 2026 that his team had already discovered the pattern Anthropic’s AI flagged. Mestre, who had used Claude in his research, questioned whether Anthropic’s team might have learned from his work. Anthropic denied this, but Mestre decided to stop using the AI tool. This incident highlights the importance of verifying claims and avoiding redundancy in scientific research. It also raises ethical questions about how AI systems are trained and whether they might inadvertently use or replicate existing discoveries without proper attribution.

The article contrasts AI-driven research with traditional scientific methods. AI systems like Claude are designed to process massive datasets quickly, identifying patterns that humans might miss. For example, Anthropic’s system analyzed 200,000 candidate DNA sequences and narrowed them down to a few worth exploring, a process that is scientifically valuable but does not necessarily equate to a discovery. Meanwhile, OpenAI claimed in September 2026 that its AI agents had solved a high-profile mathematical problem worth $1 million. However, critics argued that the problem solved might not be the most significant one in the field, or that the AI’s solution lacked proper acknowledgment of prior work. This debate illustrates the challenges of evaluating AI’s contributions to science and the risk of shifting goalposts when assessing breakthroughs.