Questions Rise Over AI Autonomy After Anthropic Discovery Claims

2026-09-27 companies

San Francisco, Monday, 28 September 2026.
Anthropic faces scrutiny over claims its AI model autonomously discovered a new enzyme system. University of Copenhagen researchers allege the model utilized their team’s prior input data, while testing revealed the AI failed to reproduce the initial breakthrough in ten subsequent reruns.

Allegations of Data Contamination

Significant scrutiny has emerged regarding the originality of Anthropic’s recent biological findings. On 2026-09-26, computational biologist Mario Rodríguez Mestre from the University of Copenhagen publicly challenged the assertion that the AI system autonomously discovered the enzyme system [1]. Mestre indicated that his research team had been studying the same enzymes and molecules for four years and had shared their findings with the company’s AI models during that period [1]. The central dispute focuses on whether the AI model, Claude, independently reasoned the results or if the output was influenced by proprietary research data previously input by Mestre’s team [1]. This revelation complicates the narrative of autonomous discovery, suggesting potential data contamination rather than novel generation [2].

Reproducibility and Statistical Concerns

Further analysis of the discovery process highlights significant reproducibility challenges. While Anthropic deployed approximately 950 Claude agents across 21 hours to identify the array-associated reverse transcriptase (ART) system, subsequent tests showed the AI failed to reproduce the initial breakthrough in ten subsequent reruns [4][6]. This indicates a success rate of 9.091 for the specific pipeline configuration used during the initial discovery [6]. The system identified a repeated sequence array associated with a family of reverse transcriptases found in bacteria-infecting viruses, reminiscent of CRISPR-like repeat arrays [2]. However, the inability to consistently replicate the finding suggests that luck or specific, non-generalizable conditions may have played a major role in the initial result [4].

Corporate Strategy and Future Outlook

Anthropic CEO Dario Amodei addressed the findings during an address at the UN Security Council on 2026-09-25, describing the discovery as a preliminary finding with potential applications in gene therapy [5]. Amodei clarified that while Claude conducted the literature review and experimental design, human scientists executed and verified the experiments in the laboratory [3][5]. The company has established a biological laboratory in California to enable AI to formulate and test biological hypotheses, though current workflows exclude handling human-infecting pathogens [3]. While Amodei stated it may eventually be possible for Claude to autonomously control laboratory equipment, he confirmed this capability is not currently operational [3]. Scientific consensus, including feedback from CRISPR pioneer Feng Zhang, emphasizes that the utility and biological validity of the ART system remain unproven and require further investigation [3].

Sources

Artificial Intelligence Biotechnology