What the study tested
The paper is "Assessing the reliability of artificial intelligence tools for laypersons seeking basic orthodontic knowledge: focus on retainers, mini-implants, and lateral cephalograms." Sadia Rizwan and colleagues at Dow University of Health Sciences in Karachi wrote it for BMC Oral Health. The journal page shows it was received on 23 June 2026, accepted on 6 November 2026 and published online on 14 November 2026. It is an early-release accepted version, so the publisher says it may still be edited before the final Version of Record. Bioengineer.org covered it on November 19.
The design was a cross-sectional comparison:
- Questions: 61 orthodontic questions drawn from frequently asked questions in the Google search bar.
- Topics: retainers, mini-implants and lateral cephalograms. A lateral cephalogram is a standardized side-view skull X-ray that orthodontists use for measurements.
- Chatbots: the free tiers of ChatGPT (GPT-4o), Gemini (1.5 Flash) and Microsoft Copilot. The abstract names no Copilot model version.
- Prompting: identical prompts written for a lay audience, entered between May 10 and May 15, 2026.
- Raters: five board-certified orthodontists scored every answer independently on a five-point rubric. A 1 meant completely inaccurate or poor, and a 5 meant highly accurate or excellent.
The authors say they looked at four dimensions: reliability, accuracy, completeness and clinical applicability.
The results
Overall differences were statistically significant. The Friedman test gave χ²(2) = 48.32, p < 0.001, with a Kendall's W of 0.396.
| Chatbot | Mean score | 95% CI |
|---|---|---|
| ChatGPT (GPT-4o) | 4.31 | 4.20–4.42 |
| Gemini (1.5 Flash) | 3.94 | 3.82–4.06 |
| Microsoft Copilot | 3.67 | 3.53–3.81 |
Pairwise Wilcoxon tests, with a Bonferroni-adjusted threshold of α = 0.017, found ChatGPT significantly ahead of both Gemini and Copilot (p < 0.001 each). The abstract does not report a test of Gemini against Copilot, so the gap between those two shouldn't be called significant. Agreement among the five raters was excellent, with ICC (2,1) = 0.87 (95% CI 0.81–0.92).
By topic, cephalometric questions produced the most consistent answers across the models. Mini-implant biomechanics produced the most variability. The abstract gives no per-topic scores and doesn't say which chatbot won each category.
How far the findings go
The results are tied to a version and a date. The tests ran in May 2026 on named free-tier builds. The authors themselves flag dynamic model updates as a limitation. Chatbots change often, so this isn't a ranking of today's products.
Copilot's model isn't identified. Copilot is a Microsoft front end that has used different underlying models over time. Without a stated version, the Copilot result can't be reproduced or compared like-for-like with GPT-4o.
The scores are expert opinion. The numbers show how five orthodontists rated the answers. They don't measure patient comprehension, clinical outcomes or the chance that any single answer is right. The authors also name evaluator subjectivity as a limitation.
Other studies disagree. Earlier orthodontic work in the same journal found different orders:
- A 2024 study on clear aligners rated Copilot's answers most reliable, ahead of Gemini and ChatGPT-4.
- A 2025 retainer study that tested four models put Gemini 2.5 ahead of ChatGPT-4.1 and Copilot GPT-4.1 on accuracy.
- An October 2026 assessment of ChatGPT, Gemini and Copilot concluded that all three gave similar and highly accurate answers in orthodontic patient consultation.
Taken together, the rankings depend heavily on model version, question set and rubric. Copilot placing last here doesn't settle anything about it in general.
What this means for Windows and Microsoft 365 users
Copilot now sits in Windows, Edge and Microsoft 365, so people will ask it health and dental questions in passing. A few practical points follow:
- Treat chatbot answers on clinical topics as general orientation, not advice.
- Check anything involving appliances, treatment timing or pain with a clinician.
- Expect different answers from different assistants, and from the same assistant after a model update.
- In a workplace, don't assume the Copilot your tenant runs matches the free consumer version tested here.
The authors reach a similar conclusion. They see conversational AI as a promising supplementary educational tool, but they say professional supervision is still essential. In their words, AI output should complement professional consultations and not replace them.
Bottom line
In this study, ChatGPT (GPT-4o) scored higher than Gemini 1.5 Flash and Microsoft Copilot on 61 layperson orthodontic questions, as rated by five orthodontists. That is a useful data point for comparing AI assistants. It is not a general ranking, and it doesn't show that any of the three is safe to follow without professional review.
References
- Study Finds ChatGPT Outperforms Gemini and Copilot on Basic Orthodontic Questions - Bioengineer.org Bioengineer.org · 2026-11-19T13:52:27+00:00
- Exploring LLM-based chatbot effectiveness in answering questions related to the risks and benefits of orthognathic treatment: a cross-sectional study - PMC pmc.ncbi.nlm.nih.gov
- Assessing the reliability of artificial intelligence tools for laypersons seeking basic orthodontic knowledge: focus on retainers, mini-implants, and lateral cephalograms | BMC Oral Health | Springer Nature Link link.springer.com