Can Large Language Models Provide Safe and Evidence-Based Nutritional Recommendations for Dental Implant Patients? A Comparative Benchmarking Study

When a dentist places a titanium screw into a jawbone to hold a false tooth, the body must treat that metal as a friend, not a foe. This process, where bone grows tightly around the implant to lock it in place, is a delicate biological dance that depends on more than just surgical skill. It relies heavily on the patient's overall health, particularly what they eat. Just as a house needs strong bricks and mortar to stand, the healing tissue and new bone require specific nutrients like protein, vitamins, and minerals to rebuild themselves after surgery. If these building blocks are missing, the implant might fail to integrate, or the surrounding gum tissue could become inflamed and sick. For decades, doctors have known that good nutrition helps, but the sheer volume of scientific advice on exactly what to eat, when to eat it, and which supplements work has become overwhelming. In this complex landscape, a new kind of digital assistant has emerged: large language models. These are powerful computer programs trained on vast amounts of text that can answer questions, summarize research, and offer advice in human language. As these tools become common in clinics, a critical question arises: can a computer, which has never seen a patient or performed surgery, safely and accurately tell a person what to eat to help their dental implant succeed?

A team of researchers set out to answer this by putting four of the most popular artificial intelligence systems to a rigorous test. They did not simply ask the computers general questions like "what is good for bones?" Instead, they created 120 specific, realistic scenarios that a dentist might face. These situations covered the entire journey of implant therapy, from checking a patient's diet before surgery to managing complications after the procedure, and even dealing with patients who have other serious health conditions. The scenarios were designed to be tricky, requiring the computer to weigh different pieces of evidence and provide a recommendation that was not just factually correct, but also safe and practical for a real person. To ensure the test was fair, the researchers submitted every single scenario to four different AI models—GPT-5, Claude, Gemini, and Copilot—using the exact same instructions. They asked each model to answer the same question three times to see if it would give the same advice every time, and then they collected a total of 1,440 responses to analyze.

To judge these answers, the researchers assembled a panel of five human experts, including top surgeons and nutrition scientists who have spent decades studying how the mouth heals. These experts were blinded, meaning they did not know which computer generated which answer. They compared every AI response against a strict, pre-written standard based on the best available medical guidelines and scientific studies. The experts looked for several things: did the advice match the science? Was it safe? Did the computer make up fake studies or references? And did it admit when the science was unclear? The results showed that while all the computers could talk about nutrition, their ability to give safe, accurate, and evidence-based advice varied significantly. One model, GPT-5, consistently outperformed the others, providing recommendations that aligned most closely with expert standards. It was the most accurate in its nutritional advice, the safest in its warnings, and the most honest about the limits of current scientific knowledge. It also made the fewest mistakes, such as inventing fake research papers to support its claims.

The other three models performed well but fell short in specific areas. The second-best model, Claude, was close to the leader but still made more errors. The other two, Gemini and Copilot, struggled more frequently with safety and accuracy. A major point of failure for all the models was when the questions involved commercial nutritional supplements. In these cases, where scientific evidence is often messy or conflicting, the computers tended to become overconfident, offering definitive advice even when the science did not support it. They also varied in their reliability; some models gave different answers to the exact same question when asked multiple times, while others remained consistent. The study found that even the best-performing computer still hallucinated, or made up, scientific references about 4.4 percent of the time, a rate that was low but still present. This means that while the AI can be a helpful tool for summarizing information, it cannot yet be trusted to make final decisions on its own.

The researchers concluded that these artificial intelligence systems are powerful assistants that can help dentists and patients navigate the complex world of nutrition for dental implants, but they are not ready to replace human judgment. The best models showed they could understand the biological needs of healing bone and offer sound general advice, but they still stumble when the evidence is weak or when specific commercial products are involved. The study suggests that the future of using these tools in dentistry lies in a partnership where the computer provides a quick, evidence-based summary, and the human expert verifies the details, checks for safety, and tailors the advice to the specific patient. Until the computers can consistently avoid making up facts and can handle uncertain evidence with the same caution a human doctor would, their role must remain supportive rather than decisive. The technology is advancing rapidly, but for now, the safest path for a patient's implant is to let the computer do the research and the human do the deciding.