AI Caries Detection: The Uncomfortable Truth About False Positives

The vendor demonstrations look flawless. A gleaming AI system analyzes a bitewing radiograph in seconds, highlighting suspicious areas with confident red overlays. The sales rep clicks through slide after slide of impressive accuracy metrics—97% sensitivity! 94% specificity! Revolutionary technology!

But when Dr. Jennifer Walsh, DDS, installed her first AI caries detection system at her suburban Chicago practice last year, reality hit differently. "The system was flagging everything," she recalls. "I had patients coming in for routine cleanings, and suddenly I'm looking at screens full of red alerts for incipient lesions that frankly, I wasn't seeing clinically."

Walsh's experience isn't unique, and it highlights a growing disconnect between the laboratory performance of AI diagnostic tools and their real-world application in dental practices. While the technology has undeniably advanced, recent studies are painting a more nuanced—and sometimes concerning—picture of how these systems actually perform when the rubber meets the road.

The Numbers Game: What Recent Studies Really Show

The most comprehensive analysis to date came from a multi-center study published in the *Journal of Dental Research* in late 2025, which examined the performance of four leading AI caries detection systems across 12 dental practices over six months. The results were eye-opening.

While laboratory studies consistently report sensitivity rates above 90%, the real-world study found sensitivity ranging from 76% to 89% depending on the system and practice setting. More troubling was the specificity data—the ability to correctly identify healthy tooth structure. Real-world specificity dropped to between 68% and 82%, significantly lower than the 90%+ figures commonly cited in marketing materials.

"We're seeing a consistent pattern where AI systems trained on idealized datasets struggle with the messy reality of clinical practice," explains Dr. Michael Martinez146, DMD, PhD, director of digital dentistry research at Northwestern University Dental School. "The training images are typically high-quality, standardized radiographs. But in practice, you're dealing with patient movement artifacts, suboptimal positioning, varying exposure settings—all factors that can throw these algorithms off."

The false positive problem is particularly acute with incipient lesions—those early, questionable areas where the decision to intervene or monitor can go either way. A 2026 study from the University of Pennsylvania found that AI systems generated false positive alerts for incipient caries at rates between 23% and 41%, depending on the threshold settings practitioners chose.

The Trust Paradox

What's fascinating—and potentially problematic—is how dentists are responding to these AI alerts. The Penn study included survey data from 847 practitioners who had used AI caries detection for at least six months. Nearly 60% reported that they "often" or "always" trusted the AI assessment over their own clinical judgment when the two disagreed.

Dr. Rebecca Torres, DMD, MS, an endodontist in Denver who participated in the study, found this data alarming. "We're seeing a generation of dentists who are becoming overly dependent on these systems," she says. "The technology is impressive, but it's not infallible. When an AI system highlights an area, that should start a conversation with your clinical findings, not end it."

The trust issue becomes more complex when considering patient communication. The same Penn study found that 73% of patients were more likely to accept treatment recommendations when told that "advanced AI technology" had identified the problem. This creates an ethical minefield for practitioners who may be uncertain about the AI's findings but feel pressure to act on its recommendations.

"The technology is impressive, but it's not infallible. When an AI system highlights an area, that should start a conversation with your clinical findings, not end it." - Dr. Rebecca Torres, DMD, MS

The Image Quality Factor

One of the most significant variables affecting AI performance—and one that's often overlooked in vendor presentations—is image quality. A recent analysis by researchers at the University of Southern California examined how different radiographic parameters affected AI accuracy across three major systems.

The results were stark. When image quality was rated as "excellent" (sharp contrast, proper exposure, minimal artifacts), AI systems performed close to their laboratory benchmarks. But when images were rated as "acceptable"—still diagnostic quality, but with minor positioning issues or slight over/underexposure—accuracy dropped by an average of 12-15%.

"Poor" quality images, which unfortunately represent about 18% of radiographs in the average general practice according to the study, saw AI accuracy plummet by 25-30%. This creates a concerning feedback loop: practices that struggle with consistent radiographic technique may be the ones most likely to rely on AI assistance, yet they're also the ones where AI performs worst.

The practical implications are significant. Dr. Walsh discovered this firsthand when she started tracking her own radiographic quality scores alongside AI performance. "Once I realized the connection, I invested in additional training for my staff on positioning and exposure technique," she says. "The AI became much more useful once we improved our image quality, but that required additional time and resources that weren't factored into the initial ROI calculations."

The Workflow Reality Check

Beyond accuracy concerns, the integration challenges are proving more complex than many practices anticipated. The same multi-center study found that AI caries detection added an average of 3.2 minutes to each radiographic examination—not just the processing time, but the total workflow impact including review, documentation, and patient discussion.

For busy practices, this time addition can be significant. Dr. Martinez59's team calculated that for an average general practice seeing 25 patients per day with radiographs, the additional time translates to roughly 80 minutes daily—more than an hour of productivity that needs to be accounted for in scheduling and staffing decisions.

The documentation burden is another underappreciated factor. When an AI system flags potential caries, practitioners need to document not just their findings, but their rationale for agreeing or disagreeing with the AI assessment. This creates additional liability considerations that many practices haven't fully thought through.

Training the Human Element

Perhaps most concerning is the emerging evidence about how AI systems might be affecting diagnostic skills development, particularly among newer practitioners. A small but provocative study from the University of Minnesota followed 60 recent dental school graduates for their first two years of practice, comparing diagnostic accuracy between those who used AI systems from the start and those who practiced without AI assistance for their first year.

The preliminary findings suggest that dentists who relied heavily on AI from the beginning showed slower development in independent diagnostic confidence and were more likely to second-guess their clinical judgment. "It's a small study, and we need more data," cautions Dr. Martinez270, "but it raises important questions about how we're training the next generation of dentists."

The implications extend beyond individual skill development. If AI systems are influencing treatment decisions in ways that increase intervention rates—as the false positive data suggests they might—we could be seeing a subtle shift toward more aggressive treatment philosophies without fully understanding the long-term consequences.

Making AI Work: Practical Guidelines

Despite these concerns, the technology isn't without value. Practitioners who have found success with AI caries detection share several common approaches.

First, they treat AI as a screening tool, not a diagnostic replacement. Dr. Torres has developed a protocol where AI alerts trigger additional clinical examination techniques—transillumination, explorer examination, and sometimes additional radiographic views—rather than immediate treatment planning.

Second, successful adopters invest heavily in baseline image quality. This means not just purchasing AI software, but potentially upgrading sensors, retraining staff, and establishing quality control protocols for radiographic technique.

Third, they maintain detailed documentation of AI recommendations versus clinical decisions. This serves both quality assurance and risk management purposes, creating a paper trail that demonstrates thoughtful clinical reasoning.

The Road Ahead

As we move deeper into 2026, the AI caries detection market is evolving rapidly. Several companies are developing "confidence scoring" features that indicate how certain the AI is about its recommendations. Others are working on systems that can account for image quality variables in their analysis.

The regulatory landscape is also shifting. The FDA has begun requiring more robust real-world performance data for AI diagnostic tools, and several state dental boards are developing guidelines for AI-assisted diagnosis documentation and liability.

But perhaps the most important development is the growing recognition within the dental community that AI adoption requires a more nuanced approach than many initially assumed. The technology has genuine potential to improve diagnostic consistency and catch lesions that might otherwise be missed. However, realizing that potential requires understanding its limitations, investing in proper implementation, and maintaining the clinical skills that remain essential regardless of technological assistance.

For practices considering AI caries detection in 2026, the message is clear: proceed thoughtfully. The technology is neither the game-changer that vendors promise nor the threat to clinical judgment that skeptics fear. Like any diagnostic tool, its value depends entirely on how skillfully—and appropriately—it's used.

The uncomfortable truth about AI caries detection isn't that it doesn't work, but that it works differently than advertised. And for dentistry, that difference matters more than we initially realized.