Artificial intelligence can now automatically read scientific data about herb-drug interactions with 90% accuracy for tables and 45% for images, according to research reviewed by Gram Research. A 2026 study testing 10 AI models found that the best systems could extract safety information from pharmacology papers reliably, though image quality remains a limiting factor. This technology could help experts monitor thousands of natural products for dangerous combinations much faster than manual review, potentially protecting millions of people who take supplements with prescription medications.
As more Americans take herbal supplements alongside prescription medications, the risk of harmful interactions grows. Researchers tested advanced artificial intelligence systems to automatically extract safety information from scientific papers—a critical step in protecting public health. According to Gram Research analysis, the best AI models successfully read 90% of data tables and 45% of charts from pharmacology studies, though image quality remains a challenge. This technology could help experts monitor thousands of natural products for dangerous combinations much faster than manual review alone.
Key Statistics
A 2026 study of 10 multimodal AI models found that the best-performing systems accurately extracted 90% of tabular data from pharmacology research papers, with a modified relative error rate of just 0.05.
According to research reviewed by Gram Research, AI systems extracted only 45% of data from figures and images in scientific papers, with image resolution and information density identified as primary barriers to better performance.
The 2026 analysis compared 3 open-source and 7 closed-source AI models for visual information extraction from pharmacology literature, revealing that both types had distinct strengths and weaknesses for different data formats.
The Quick Take
- What they studied: Whether artificial intelligence can automatically read and understand safety information about herbs and supplements from scientific research papers and images.
- Who participated: Researchers tested 10 different AI computer programs (3 free versions and 7 paid versions) on their ability to extract data from tables and figures in pharmacology research.
- Key finding: The best AI systems correctly extracted 90% of information from data tables and 45% from charts and graphs, with an error rate of 0.05—meaning they’re fairly reliable but not perfect.
- What it means for you: AI could help safety experts work faster to identify which herb-drug combinations might be dangerous, potentially protecting people who take supplements with prescription medications. However, the technology still needs improvement for reading complex images.
The Research Details
Researchers compared how well different artificial intelligence systems could read and understand information from scientific papers about herbs and drugs. They tested 10 different AI models—some free and some paid—by giving them tables and images from pharmacology studies and checking if the AI extracted the correct information. This is similar to asking different students to read a chart and report what they see, then grading how accurate each student was.
The AI systems being tested are called “multimodal models,” which means they can understand both text and images. The researchers specifically focused on visual information—tables and figures—because these contain important safety data that’s hard for computers to read automatically. They measured success by checking how many data points the AI got right and how far off the wrong answers were.
Right now, experts have to manually read thousands of scientific papers to find information about herb-drug interactions. This is slow and expensive. If AI can do this work automatically and accurately, experts could monitor many more natural products for safety problems. The current system only tracks about 33 natural products carefully, but thousands are sold in stores. Better automation could help protect millions of people who take supplements.
This study tested real AI systems on actual research data, which makes the results practical and relevant. The researchers tested multiple AI models to see which ones work best, rather than just testing one. However, the study doesn’t specify exactly how many tables and images were tested, which would help readers understand how thoroughly the systems were evaluated. The fact that the best models achieved 90% accuracy on tables is strong evidence they work reasonably well for that task.
What the Results Show
The best-performing AI systems successfully extracted 90% of the information from data tables in scientific papers. This is quite good—it means if a table had 100 pieces of information, the AI would get about 90 of them right. For more complex images and figures, the AI did less well, correctly extracting only 45% of the data. The error rate (how wrong the mistakes were) was very small at 0.05, meaning when the AI made mistakes, they weren’t huge errors.
The researchers found that two main problems limited how well the AI worked: image quality and how crowded the information was. When images were low resolution (blurry or pixelated) or packed with lots of information close together, the AI struggled more. This makes sense—if a human had to read a blurry chart with tiny text, they’d struggle too.
The study compared open-source AI models (free versions anyone can use) with closed-source models (paid versions from companies). Both types had strengths and weaknesses, suggesting that different AI systems might be better for different tasks.
The research showed that image resolution was one of the biggest problems preventing better accuracy. When images were clearer and easier to read, the AI performed better. Similarly, when information was spread out and not too crowded, the AI could extract it more reliably. This suggests that improving image quality in scientific papers could help AI systems work better. The study also found that different AI models had different strengths—some were better at tables, others better at figures—suggesting that using multiple AI systems together might give better results than relying on just one.
This research builds on earlier work where scientists created a ‘knowledge graph’ (a database of connections) for 33 natural products and their safety information. That earlier project showed the value of organizing herb-drug interaction data, but it was done by hand, which is slow. This new study tackles the scaling problem—how to do this work for thousands of products instead of just 33. The 90% accuracy rate for tables suggests AI is ready to help with this bigger task, at least for the easier part of the work.
The study doesn’t clearly state how many tables and images were tested, making it hard to know if the results are based on a small sample or a large one. The 45% accuracy for figures is much lower than for tables, and the study doesn’t explain exactly why or how to fix it. The research also doesn’t test whether the AI’s mistakes would actually cause safety problems in real-world use—a 90% accuracy rate might be fine for some purposes but not others. Finally, the study doesn’t compare AI performance to human experts, so we don’t know if AI is better, worse, or equal to people doing this work manually.
The Bottom Line
AI systems show promise for helping experts monitor herb-drug interactions, but they’re not ready to work completely alone yet. The 90% accuracy on tables suggests they could handle that part of the work with human review of the 10% they miss. For figures and images, the 45% accuracy means humans would need to do most of the work. Confidence level: Moderate. The technology is useful as a helper tool but not as a replacement for expert review.
This research matters most to pharmacists, doctors, and public health officials who need to track supplement safety. It also matters to people who take both prescription drugs and herbal supplements—this technology could eventually help identify dangerous combinations they should avoid. Researchers and companies developing AI systems should care about these results because they show where improvements are needed. People who only take prescription drugs or only take supplements don’t need to worry about this specific research.
If AI systems are improved based on these findings, it could take 1-3 years before they’re reliable enough to use in real safety monitoring. The immediate next step is improving how AI reads images and figures. Once that’s solved, experts could start using AI to help monitor more natural products within a few years.
Frequently Asked Questions
Can AI read scientific papers about herb and drug interactions?
Yes, AI can read tables from scientific papers with 90% accuracy, but struggles more with images and charts at 45% accuracy. Image quality and crowded information limit performance, so human experts still need to review AI results.
How accurate is AI at extracting data from research tables?
The best AI systems achieved 90% accuracy on tables with a very small error rate of 0.05, meaning they correctly read most information but still miss about 1 in 10 data points and need human verification.
Why is AI having trouble reading figures and charts?
Low image resolution (blurry pictures) and crowded information make it hard for AI to read figures accurately. When images are clearer and less crowded, AI performance improves significantly.
Could this AI help identify dangerous supplement interactions?
Potentially yes. If AI can automatically extract safety data from thousands of research papers, experts could monitor more natural products for dangerous drug interactions much faster than doing it manually.
Is AI ready to replace human experts reviewing supplement safety?
Not yet. While AI shows promise with 90% table accuracy, the 45% figure accuracy and lack of comparison to human performance means AI should work alongside experts, not replace them, for now.
Want to Apply This Research?
- Users taking both prescription medications and herbal supplements should track which supplements they’re taking, the dosage, and any new symptoms or side effects that appear. This creates a personal safety record that can be shared with doctors.
- Before starting any new herbal supplement, users should check with their pharmacist or doctor about potential interactions with their current medications. The app could send reminders to do this check before purchasing new supplements.
- Set up monthly check-ins to review all current supplements and medications with a healthcare provider. The app could maintain a timeline of when supplements were started and any health changes that occurred, helping identify patterns.
This research describes AI technology for extracting information from scientific papers and does not provide medical advice about specific herb-drug interactions. Anyone taking herbal supplements with prescription medications should consult their pharmacist or doctor before starting new supplements or making changes to their regimen. The AI systems described in this study are research tools and are not yet approved for independent clinical decision-making. Always rely on qualified healthcare professionals for personalized medical guidance about supplement safety.
This research translation is published by Gram Research, the science division of Gram, an AI-powered nutrition tracking app.