Utility of Artificial Intelligence Image Recognition Technology in Identifying Prescription Pills

Every day, approximately 165 children visit US emergency departments for medication-related poisoning.1 Concerningly, data show 59.5% of all poisonings are not reported to Poison Control, which may suggest individuals are looking to other sources during such emergencies.2 A 2026 survey indicates approximately one-third of US adults trust health information from artificial intelligence (AI), and many commonly used AI platforms claim they can identify pills if asked.3–6The potential applications of AI have expanded as publicly accessible platforms have evolved to include image-upload features for visual content analysis, which have been advertised as capable of identifying plants, tools, or products.4–6 Members of the public can open a chat with an AI model, take a picture of a medication that is not in its original container or is in a container with an illegible label using the built-in camera feature, and receive information about the medication. Some people may be tempted to use this feature in emergencies as opposed to a smartphone app because they are already using these tools for everyday tasks and no download is required. This study aims to assess the accuracy of 5 free-to-use AI platforms in identifying prescription medications from images.The 50 most prescribed medications in the United States were sampled.7 An inhaler-form of albuterol was excluded, leaving a final sample of 49 medications in tablet or capsule form. Medication identity was verified by a pharmacist. An iPhone 13 Pro was used to photograph each of the 49 pills on a neutral background. AI tools (ChatGPT-5, Gemini 2.5 Flash, Claude Sonnet 4.5, Microsoft Copilot [Oct 2025], Google Lens 2.5 Pro) were prompted exclusively with photos of the front and back of the tablet (as allowed by the tool) or markings on the capsule (Figure 1). Drugs.com, a pill identification search tool recommended by the National Institutes of Health, was used as a control.8 Outputs were categorized as correct, incorrect, unidentified, or refusal to answer. An inability to identify the pill or an identification of a nonmedication (ie, bed bug, tonsil stone) were grouped together as unidentified. An identification was considered correct only if the first outputted result was the proper medication. This study was deemed exempt by the Northwell Health Institutional Review Board.Across all AI platforms, 18.4% of responses were correct identifications, whereas 37.1% of responses were incorrect. Claude refused to identify any medications, citing safety concerns. ChatGPT, Gemini, and Copilot included warnings with their responses; Google Lens did not. There were 7 instances (14.3%) where Gemini initially gave a response, then deleted it and replaced it with a deferral message such as “Try a different topic. When something seems like it might not be safe or appropriate for you, I draw the line.” Excluding Claude, AI outputs were 23.0% correct, 46.4% incorrect, 21.6% unidentified, and 22.9% refusals. Copilot had the lowest success rate at 8.16%, and Gemini had the highest success rate at 36.7%. In contrast, Drugs.com correctly identified all but 1 medication (98.0% accuracy) (Table 1).Artificial intelligence tools are currently unreliable in identifying prescription medications from images and should not be used for this purpose. Publicly accessible AI platforms regularly release new models, with recent updates introducing safeguards aimed at preventing misuse. For example, from ChatGPT-4 to ChatGPT-5, safeguards were added to reduce the rate of harmful or unsafe outputs.9 This update may have contributed to the difference in results between this study and a 2024 study that prompted ChatGPT to identify nephrology medications from images. Specifically, there was an increase in unidentified medications, as 0% of outputs from the 2024 study, using ChatGPT-4, fell into this category compared with 40.8% in the current study using ChatGPT-5.10Future AI models may be better equipped to analyze images of medications, but current trends in these systems are moving away from providing a tool that’s use could cause harm. At this time, such platforms should be designed to consistently provide an error message when asked to identify a medication. Future studies may wish to explore the utility of dedicated AI systems developed specifically for medication identification from images. Clinicians should strongly discourage the use of general AI tools in cases of accidentally ingested medications and recommend immediately contacting Poison Control centers instead.