.png){kind=small}
AI in Orthodontics: How to Manage Hallucinations
Hallucinations in generative AI are essentially "nonsense responses," as described by a participant in one of our AI Sprints training sessions. These hallucinations refer to outputs that seem accurate but are, in fact, incorrect, irrelevant, or entirely fabricated. For example, when preparing material for a lecture, I prompted ChatGPT to provide information about a foundational AI finance model, "Bloomberg." Instead, the model responded with "Blomindalle," an incorrect name that points to deeper limitations, even in simple tasks.
While hallucination rates are improving, with current models showing rates between 3% and 10%, eliminating them entirely remains challenging due to AI's reliance on predictive algorithms rather than genuine understanding or real-time data verification. This is particularly critical in high-precision fields like orthodontics, where even a minor hallucination, such as confusing "Angle's malocclusion Class III" with "Class II", can lead to significant treatment implications.
Why Do Hallucinations Happen?
AI hallucinations arise from several factors, including:Training Data Quality and Diversity, models trained on vast, diverse datasets are prone to absorbing both accurate and inaccurate information. Prompt Complexity, how questions or requests are phrased can influence the likelihood of hallucinations. The more complex or ambiguous the prompt, the higher the chance for inaccuracies. And Model Design, AI tools are often optimized to generate plausible responses rather than strictly accurate ones, making hallucinations an inherent risk.
These issues can be particularly problematic in settings that require precision, such as legal, medical and dental fields. Inaccurate information could lead to misunderstandings or errors in diagnostic and treatment planning. Moreover, generative AI is not limited to text; some AI transcription tools that convert audio to text can also add hallucinated words to reports and documents. Research from the University of Michigan found hallucinations in eight out of every ten audio transcriptions of public meetings. A machine learning engineer found hallucinations in more than half of 100 hours of Whisper transcriptions, and a developer reported hallucinations in nearly all of the 26,000 transcriptions they created with Whisper. Such fabrications are problematic, as Whisper is used in various industries to translate and transcribe interviews, generate text, and create subtitles for videos.
Sometimes the issue lies not with the tool itself but with improper use. Utilizing an unvalidated AI tool just to boost productivity can be risky, leading to unexpected outcomes. It's crucial to build an AI team for your business that includes experts not only in generative AI but also in the specific domain of your industry.
How Can We Prevent Hallucinations?
Several strategies can help reduce hallucinations in AI. We explore this topic in depth during our AI Sprints training in orthodontics, but here are a few key strategies: Writing precise, clear prompts reduces the likelihood of hallucinations, especially in tasks requiring factual accuracy (prompt engineering). Training AI on orthodontics-specific data helps the model better understand context, reducing errors in diagnostics and treatment recommendations. Retrieval-Augmented Generation (RAG), this method combines generative models with a database of verified information, allowing AI to pull answers from reliable sources, enhancing both accuracy and relevance. Introducing checkpoints where AI outputs are reviewed by professionals or cross-referenced against known standards provides an additional layer of assurance, especially in clinical environments (Quality and Verification mechanisms). And most importantly, by fine-tuning models on orthodontic-specific data, you create an AI that is more adept at understanding the nuances of orthodontics knowledge.
Takeaway
While AI offers tremendous potential, its current limitations require a balanced approach. Like integrating any other digital tool, training a new employeer, ou launch a new product, AI systems must be created, tested and validate but professionals experts in your domain specific. These tools should be used as a supplement rather than a substitute for professional judgment and expertise. At Ortho.i®, our mission is to master Responsible AI use in orthodontics by educating our patients, colleagues and supporting organizations in developing and training AI systems with high-quality, accurate data to ensure trustworthy results.
Get in touch with us, https://www.orthoi.ai
EXPLORE OUR LATEST VIDEOS ON YOUTUBE
In this episode Prof Dr Adriano Araujo, PhD talk with Angela an AI assistant developed by Ortho.i Robotics. An exciting conversation about the Generative Ai limitations and how to prevent hallucinations and use responsible AI in Orthodontics.
Comments