Patents are legal documents that describe new inventions, including chemicals that are still in development. These documents often contain valuable information about potentially harmful substances long before they appear in consumer products or the environment. This makes patents a key resource for early warning systems (EWS), which help authorities identify hazardous chemicals before they pose a risk to people or the planet. However, much of the chemical information in patents is hidden in images, such as drawings of molecular structures and chemical reactions. These images are not searchable by text-based systems, making it difficult for computers to analyze them automatically. A proposed workflow involves using artificial intelligence (AI) to detect and interpret chemical structures in patent images. Once detected, the AI would convert the structure into a machine-readable format, allowing it to be compared with known chemicals and their associated risks. If the AI identifies a potentially dangerous or completely new chemical structure, it would flag it for further examination by human experts. This process would allow scientists to determine if the chemical is already known, if it resembles other hazardous substances, or if it has features linked to specific dangers. In a recent study published in the journal Chemical Research in Toxicology, researchers tested three popular AI tools on chemical structures from two specialized data sets obtained from the European Patent Office's Espacenet database. One set focused on general organic chemistry, and the other on per- and polyfluoroalkyl substances (PFAS), a group of synthetic chemicals known for their environmental persistence and potential health risks. The AI-generated structures were then reviewed by five chemistry experts. The results showed that the AI performed well on standard chemical structures, with accuracy rates between 74% and 78%. However, when it came to PFAS structures, the AI's performance dropped significantly, with none of the tools correctly identifying 26 out of 43 unique structures. The AI tools struggled with the complex and sometimes outdated ways that chemicals are depicted in patent images. Older patents often had blurry or distorted images, which made it harder for the AI to interpret them accurately. If an AI misreads a chemical structure, the consequences can be serious. It could lead to unnecessary precautions being taken or, worse, allow a truly dangerous chemical to go unnoticed. Because of this, the researchers emphasize that any future automated system must include quality control mechanisms and confidence scores to ensure reliability. The study highlights that AI is not intended to determine whether a chemical is harmful, but rather to act as a translator, converting chemical drawings into digital formats that can be compared with known substances and flagged for further review. While the current technology is not yet ready for full automation, the study provides a clear understanding of its strengths and limitations. This knowledge is essential for developing more effective systems in the future, which could help improve early warning systems for hazardous chemicals.