Assessing the Quality of AI-Generated Health Information: A Comparative Study of Large Language Models in Patient Education on Dental Veneers

The Digital Dentist: A Tale of AI, Veneers, and Reading Scores

Imagine you're standing in a vast, noisy library where millions of books are being written every second by invisible robots. This is the internet, and specifically, the world of Artificial Intelligence (AI). These AI "robots" are called Large Language Models. Think of them as super-powered parrots that have read almost everything ever written on the internet. They can chat, write stories, and answer questions, but they don't actually know things the way humans do; they just predict what words should come next based on patterns they've seen before.

Now, picture a specific corner of this library dedicated to your smile: dentistry. One of the most popular topics here is "laminate veneers." These are like tiny, custom-made porcelain stickers that dentists glue onto the front of your teeth to make them look perfect, straight, and white. Because people care so much about their smiles, they often turn to the internet first to ask questions like, "Will it hurt?" or "How long will it last?" But here's the catch: the internet is full of noise. Some information is great, some is wrong, and some is written in a language so complicated that only a scientist could understand it. This is where the question of the day comes in: If you ask a super-smart AI robot about veneers, will it give you the right answer, and will you actually be able to understand it?

The Great Chatbot Showdown

In this study, four different AI chatbots decided to step into the ring for a friendly (but serious) competition. The contestants were ChatGPT-5.3 mini, Gemini-3.5 Flash, DeepSeek-V4, and Microsoft Copilot. The researchers wanted to see which of these digital brains was the best at answering 20 of the most common questions people ask about dental veneers.

To judge the contestants, the researchers didn't just ask, "Did you like the answer?" Instead, they used a set of strict scoring rules, like a panel of three expert dentists acting as judges. They looked at four main things:

1. Accuracy: Did the robot tell the truth? (Imagine a judge checking if the robot said "veneers are made of chocolate" vs. "veneers are made of porcelain.")
2. Reliability: Could you trust the source? Did the robot explain why it knew the answer, or did it just guess?
3. Quality: Was the answer helpful and complete, or was it a vague, half-baked response?
4. Readability: Was the answer written in plain English that a regular person could understand, or was it a confusing mess of big words?

The Results: Who Won the Crown?

The competition was fierce, and the results showed that not all AI robots are created equal. The judges found statistically significant differences between the four models, meaning the winners weren't just lucky; they were genuinely better at their jobs.

The Accuracy and Reliability Champion:
ChatGPT-5.3 mini took home the gold medal for being the most accurate and reliable. It scored the highest on the accuracy scale, averaging 4.400 out of 5. It also won the reliability contest with a score of 4.517 and the overall quality contest with 4.400. In simple terms, if you asked ChatGPT-5.3 mini about veneers, it was the most likely to give you the correct, trustworthy, and high-quality information.

The "Easy to Read" Champion:
However, being the smartest doesn't always mean being the easiest to understand. That title went to Gemini-3.5 Flash. While it was a close second in accuracy, it won the "Readability" race with a Flesch Reading Ease Score (FRES) of 38.92.
Here's a quick analogy for that score: The FRES scale goes from 0 (impossible to read) to 100 (as easy as a nursery rhyme). A score of 38.92 is still a bit tricky—it's like reading a high school textbook—but it was the easiest to read among the four. The other models were much harder to digest.

The Struggling Contestant:
DeepSeek-V4 had a tough day. It scored the lowest in almost every category. Its accuracy was 4.150, its reliability was 3.517, and its quality was 4.033. But the real struggle was in readability. DeepSeek-V4 had a FRES score of 25.33, which means its answers were written in a very dense, complex style that would be hard for most people to understand without a dictionary handy.

The Middle Ground:
Microsoft Copilot landed somewhere in the middle. It did well, but it didn't beat ChatGPT-5.3 mini in accuracy or reliability, nor did it beat Gemini in readability.

What This Means for You

The study concluded that while AI chatbots are becoming powerful tools for learning about dental treatments like veneers, they are not all the same. ChatGPT-5.3 mini proved to be the most trustworthy source of facts, while Gemini-3.5 Flash was the best at speaking "human."

However, there is a big "but." The researchers emphasized that even the best AI robot can make mistakes or miss the mark. Just because a robot gives an answer doesn't mean it's perfect. The study suggests that while these tools are great for getting a head start on information, you should never let them replace a real dentist. A human expert is still needed to double-check the facts, because in the world of your smile, getting it right matters more than just getting an answer quickly.

In the end, the paper rejects the idea that all AI models perform the same. Instead, it shows that the quality of the information you get depends entirely on which robot you ask. So, if you're curious about veneers, you might want to ask the smartest robot for the facts, but maybe ask the easiest-to-read robot to help you understand them—just make sure a real dentist gives the final thumbs-up!