Just how smart is artificial intelligence (AI)? Researchers writing in Scientific Reports have assessed the ability of large language models (LLMs) when presented with questions from a national exam used in postgraduate training in Turkey, in topic areas surrounding paediatric dentistry.
The results were insightful or terrifying, depending on where you stand on the implementation of AI. Gemini 2.5 Pro (Google), GPT-5 (OpenAI) and Grok-4 (SpaceXAI) were the best performing LLMs, with accuracy rates of 92.44%, 90.76% and 88.24% respectively.
The article suggests that the performance of these AI models was approaching “expert-level” accuracy.
Not all AI models are made equally. Their answers may differ based on the information they are trained on and their programming. Some of the lowest performing LLMs included Qwen3 (Alibaba, Standard 54.62%, High Reasoning 49.57%) and MedGemma (Google, 58.82%). The latter is important to note as it is a dedicated collection of open models that are focused on aiding medical text or imaging tasks.
The results may have a variety of impacts on how AI is viewed in the future of healthcare learning. For example, adequately trained models may be able to aid clinical decision-making and understanding in everyday care, with varying degrees of confidence.
To put any treatment plan into action, a clinician will have to correctly interpret the results generated by an LLM, and have the appropriate training to proceed with care. With this understanding, AI is merely an aid, as opposed to a complete solution. If programmed to give a high degree of accuracy when assessing dental tasks, it may operate best as a second opinion – though the thoughts of clinically trained professionals should always be included in any decision-making process.
Research conducted by Dental Reviewed found that around one in three professionals based in the UK, US and Canada already use AI tools in clinical practice, as reported by GDPUK. A further 38% are considering adoption.
Interestingly, 12% have tried AI but ceased using it, whilst another 18% have no interest in adopting the technology, suggesting that, at current, these solutions are far from must-haves in clinical care.
52% called it a helpful second opinion, with 2% reporting that they trust AI more than their own clinical thoughts.
AI is undoubtedly growing in ability and application. Careful implementation in clinical workflows is required, and it may be helpful when preparing for assessments – but don’t expect an AI cheat-sheet to be allowed in postgraduate exam halls any time soon.
>Since you’re here, why not consider reading about new research into oral cancer detection?
Image credit: Unsplash
