Book a Demo
All posts
Clinical Research6 min read

What is an AI second opinion?

An AI second opinion reads the same radiograph you do, but in a different way - and that difference is what makes it useful. How to think about it as a practising clinician or clinic owner - and why it works best on every image, not only the difficult ones.

Authored by
Annabel ParkmanRegulatory Lead
Anton JanérCEO & Co-Founder

An AI second opinion reads the same radiograph you do, and marks areas that may show a finding. It does not change the image, and it does not diagnose or decide treatment, but it reads the image differently from dental professionals. That difference is what makes it useful.

Even for experienced dentists, caries is often easy to miss. This post is about how to think about the second opinion tool, which could help find more pathologies when integrated into dental practice's day-to-day routine.

Human judgement is inevitably flawed

Dentists reading the same image do not always agree about the findings with each other,1,2 and even with their own readings two weeks later.2 Within a single image, readers fall prey to a well-documented effect in radiology called satisfaction of search - once they find one lesion, they become more likely to miss the next.3–7 Radiologists are also prone to inattentional blindness: in a study where a gorilla icon forty-eight times the size of an average nodule was placed in the last case of a familiar lung-nodule detection task, 83% of radiologists did not see the gorilla, even though most looked directly at it.8

Performance changing throughout the day further adds to the human error. Radiologists' accuracy in detecting fractures is significantly lower at the end of a working day,9 and cancer screening order rates fall significantly as a day progresses.10 More opioids and antibiotics are prescribed later in the clinic session, associated with fatigue.11,12 When in a rush, dentists reading bitewings detect significantly fewer findings.13

If those errors lead to inconsistent diagnosis, the pathologies may progress, leading to health deterioration and late, and potentially more invasive, treatment that could have been avoided in the first place.

AI also makes mistakes, but different ones

A second opinion tool can help here. Dentists and dental students using AI found significantly more early carious lesions,14–19 although AI may get it wrong in other ways.20 It may flag more pathologies that aren't present,15,19 underestimate lesion depth,21,22 and mark false positives on normal anatomy such as the pulp chamber and cervical burnout.23 This is often referred to as a 'different failure mode' in engineering and is a key concept in designing robust systems.

Treatment decisions depend on the diagnosis, so getting it right matters in every medical field. AI should be used as a complementary tool24,25 - in that way it can catch what you miss, and you can catch what it gets wrong, potentially resulting in more accurate diagnosis. A model that is fixed for each release reads the last radiograph on a Friday afternoon the same way as the first on a Monday. It makes mistakes in a consistent way, so they can be measured before release and do not grow when the schedule slips.

The confidence fallacy & car safety features

It is tempting to ask for a second opinion only on the images you feel unsure about. But uncertainty is an unreliable guide. In one study, 118 physicians worked through case vignettes. They diagnosed 55.3% of easier cases correctly, but only 5.8% of harder ones, and their confidence barely changed. Those who felt more confident also requested fewer additional tests.26 This is not an isolated finding. A comprehensive review describes overconfidence as a problem among physicians, which contributes to diagnostic error.27

The same paper suggests that overconfidence is partly intrinsic. Analogously - cars are equipped with different safety features to expect the unexpected. Nobody turns the safety system on only when they expect a crash, it's simply always on.

A missed finding is, by definition, one you did not notice. The images where a second opinion matters most can feel entirely routine. Like trying to switch on a collision warning system a second before you hit a deer is plainly not the safest approach - using the tool only when you feel unsure is a similar pattern that may leave you with a suboptimal analysis.

Too little trust, and too much

Under-use is a real risk. In one study, radiologists rated advice as lower quality when it appeared to come from AI.28 The risk here is that when radiologists are too confident in their own judgement and underweight AI suggestions, they simply do not realise the gains from AI tools.29

Over-reliance is just as real. Expectedly, diagnostic accuracy is significantly worse after inaccurate advice, whether labelled as coming from AI or human.28 It is shown that radiologists of all experience levels are prone to automation bias,30 and that mistakes in AI output can mislead even experts.31 If this bias leads observers to shift their detections towards the AI tool's pattern, they grow more confident in both correct and incorrect lesion staging32 and the optimal gain may not be attained. Thankfully, dentistry lends itself well to a clinical inspection to corroborate the image analysis - something that is not always the case in fields that may require invasive surgery to determine the accuracy.

How to use it

  • Use it on every radiograph you assess, not only the ones that feel difficult. There is often an auto-analysis setting available.
  • Form your own view first. If you auto-detect, hide the findings before you have made your own judgement. Where you and the tool disagree, look again.
  • Treat a suggestion as a reason to look, not a reason to treat. The decision stays yours.
  • It is one option among several. Longer appointments when radiographs are taken, or reviewing images again away from time pressure, may also reduce errors, though both cost time.13 A colleague's second read is another.

In 2016, AI pioneer Geoffrey Hinton said people should stop training radiologists, because deep learning would outperform them within five years.33,34 Nearly a decade later, he said he had spoken too broadly and was wrong on the timing, and now expects AI and radiologists to work together, making radiologists 'a whole lot more efficient in addition to improving accuracy'.34 An AI second opinion is not a better dentist. It is a different reader, one that does not tire or rush and has its own blind spots. Used on every image and checked by your judgement, it can help make missed findings less likely.14

Sources

  1. Lam EWN, Law AS, Nguyen RHN, et al. Interexaminer Agreement in the Radiologic Identification of Apical Periodontitis/Rarefying Osteitis in the National Dental Practice-Based Research Network PREDICT Endodontic Study. J Endod. 2021;47(10):1575-1582. doi:10.1016/j.joen.2021.07.008
  2. Natto ZS, Olwi A, Abduljawad F. A comparison of the horizontal and vertical bitewing images in detecting approximal caries and interdental bone loss in posterior teeth: A diagnostic accuracy randomized cross over clinical trial. J Dent Sci. 2023;18(2):645-651. doi:10.1016/j.jds.2022.08.006
  3. Ashman CJ, Yu JS, Wolfman D. Satisfaction of Search in Osteoradiology. Am J Roentgenol. 2000;175(2):541-544. doi:10.2214/ajr.175.2.1750541
  4. Adamo SH, Gereke BJ, Shomstein S, Schmidt J. From "satisfaction of search" to "subsequent search misses": a review of multiple-target search errors across radiology and cognitive science. Cogn Res Princ Implic. 2021;6(1):59. doi:10.1186/s41235-021-00318-w
  5. Berbaum KS, Franken EA, Dorfman DD, et al. Satisfaction of Search in Diagnostic Radiology. Invest Radiol. 1990;25(2):133-139. doi:10.1097/00004424-199002000-00006
  6. Huynh JD, Rhodes SC, Hatton JF, Khademi JA. Satisfaction of Search in Periapical Radiograph Interpretation. J Endod. 2021;47(2):291-296. doi:10.1016/j.joen.2020.11.001
  7. Kim YW, Mansfield LT. Fool Me Twice: Delayed Diagnoses in Radiology With Emphasis on Perpetuated Errors. Am J Roentgenol. 2014;202(3):465-470. doi:10.2214/AJR.13.11493
  8. Drew T, Võ MLH, Wolfe JM. The Invisible Gorilla Strikes Again: Sustained Inattentional Blindness in Expert Observers. Psychol Sci. 2013;24(9):1848-1853. doi:10.1177/0956797613479386
  9. Krupinski EA, Berbaum KS, Caldwell RT, Schartz KM, Kim J. Long Radiology Workdays Reduce Detection and Accommodation Accuracy. J Am Coll Radiol JACR. 2010;7(9):698-704. doi:10.1016/j.jacr.2010.03.004
  10. Hsiang EY, Mehta SJ, Small DS, et al. Association of Primary Care Clinic Appointment Time With Clinician Ordering and Patient Completion of Breast and Colorectal Cancer Screening. JAMA Netw Open. 2019;2(5):e193403. doi:10.1001/jamanetworkopen.2019.3403
  11. Philpot LM, Khokhar BA, Roellinger DL, Ramar P, Ebbert JO. Time of Day is Associated with Opioid Prescribing for Low Back Pain in Primary Care. J Gen Intern Med. 2018;33(11):1828-1830. doi:10.1007/s11606-018-4521-8
  12. Linder JA, Doctor JN, Friedberg MW, et al. Time of Day and the Decision to Prescribe Antibiotics. JAMA Intern Med. 2014;174(12):2029-2031. doi:10.1001/jamainternmed.2014.5225
  13. Plessas A, Nasser M, Hanoch Y, O'Brien T, Bernardes Delgado M, Moles D. Impact of time pressure on dentists' diagnostic performance. J Dent. 2019;82:38-44. doi:10.1016/j.jdent.2019.01.011
  14. Mertens S, Krois J, Cantu AG, Arsiwala LT, Schwendicke F. Artificial intelligence for caries detection: Randomized trial. J Dent. 2021;115:103849. doi:10.1016/j.jdent.2021.103849
  15. Devlin H, Williams T, Graham J, Ashley M. The ADEPT study: a comparative study of dentists' ability to detect enamel-only proximal caries in bitewing radiographs with and without the use of AssistDent artificial intelligence software. Br Dent J. 2021;231(8):481-485. doi:10.1038/s41415-021-3526-6
  16. Gakenheimer DC. The efficacy of a computerized caries detector in intraoral digital radiography. J Am Dent Assoc. 2002;133(7):883-890. doi:10.14219/jada.archive.2002.0303
  17. Tracy KD, Dykstra BA, Gakenheimer DC, et al. Utility and effectiveness of computer-aided diagnosis of dental caries. Gen Dent. 2011;59(2):136-144.
  18. Hegde S, Gao J, Cox S, Nanayakkara S, Logothetis R, Vasa R. Machine learning algorithms enhance the accuracy of radiographic diagnosis of dental caries: a comparative study. Dentomaxillofacial Radiol. 2025;54(8):632-641. doi:10.1093/dmfr/twaf053
  19. Lee S, Oh S il, Jo J, Kang S, Shin Y, Park J won. Deep learning for early dental caries detection in bitewing radiographs. Sci Rep. 2021;11(1):16807. doi:10.1038/s41598-021-96368-7
  20. Dvijotham K (Dj), Winkens J, Barsbey M, et al. Enhancing the reliability and accuracy of AI-enabled diagnosis via complementarity-driven deferral to clinicians. Nat Med. 2023;29(7):1814-1820. doi:10.1038/s41591-023-02437-x
  21. Komtui P, Panyarak W, Suttapak W, Nirunsittirat A, Wantanajittikul K, Wattanarat O. Deep learning versus general dentists: a clinical evaluation of caries diagnostic accuracy on bitewing radiographs. BMC Oral Health. 2026;26:1794. doi:10.1186/s12903-026-09088-7
  22. Heiden G, Pereira P, Yoshida ML, Oliveira D, Rocha M, Ribeiro APD. Comparative evaluation of proximal caries detection methods using human, artificial intelligence, and micro-CT assessments. J Dent. 2026;173:106835. doi:10.1016/j.jdent.2026.106835
  23. Pornprasertsuk-Damrongsri S, Vachmanus S, Papasratorn D, et al. Clinical application of deep learning for enhanced multistage caries detection in panoramic radiographs. Sci Rep. 2025;15:33491. doi:10.1038/s41598-025-16591-4
  24. Ganss C, Jung K, Schilling L, Sonderegger S, Neuhaus KW. Human and artificial intelligence performance in radiographic caries detection: ex vivo tooth section-referenced evaluation and implications for clinical decision-making. J Dent. 2026;168:106602. doi:10.1016/j.jdent.2026.106602
  25. Lai J, Guo S, Wang K, Li X, Yu X, Wang X. Artificial intelligence for binary dental caries diagnosis using intraoral images and dental radiographs: a systematic review and meta-analysis. BMC Oral Health. 2026;26:437. doi:10.1186/s12903-026-07770-4
  26. Meyer AND, Payne VL, Meeks DW, Rao R, Singh H. Physicians' Diagnostic Accuracy, Confidence, and Resource Requests: A Vignette Study. JAMA Intern Med. 2013;173(21):1952-1958. doi:10.1001/jamainternmed.2013.10081
  27. Berner ES, Graber ML. Overconfidence as a Cause of Diagnostic Error in Medicine. Am J Med. 2008;121(5):S2-S23. doi:10.1016/j.amjmed.2008.01.001
  28. Gaube S, Suresh H, Raue M, et al. Do as AI say: susceptibility in deployment of clinical decision-aids. NPJ Digit Med. 2021;4:31. doi:10.1038/s41746-021-00385-9
  29. Agarwal N, Moehring A, Rajpurkar P, Salz T. Combining Human Expertise with Artificial Intelligence: Experimental Evidence from Radiology. National Bureau of Economic Research; 2023. doi:10.3386/w31422
  30. Dratsch T, Chen X, Rezazade Mehrizi M, et al. Automation Bias in Mammography: The Impact of Artificial Intelligence BI-RADS Suggestions on Reader Performance. Radiology. 2023;307(4):e222176. doi:10.1148/radiol.222176
  31. Tschandl P, Rinner C, Apalla Z, et al. Human–computer collaboration for skin cancer recognition. Nat Med. 2020;26(8):1229-1234. doi:10.1038/s41591-020-0942-0
  32. Gonzalez-Valenzuela RE, Vu QTD, Mettes P, Loos BG, Marquering H, Berkhout E. Human-AI Collaboration in Proximal Early Caries Lesion Detection: Potential Benefits for Novices and Potential Risks for Experts. Caries Res. Published online September 10, 2026. doi:10.1159/cre/abuag009
  33. Gans J. Radiologists are the canaries in a coal mine. Joshua Gans' Newsletter. September 26, 2025. Accessed October 29, 2026. https://joshuagans.substack.com/p/radiologists-are-the-canaries-in
  34. Lohr S. Your A.I. Radiologist Will Not Be With You Soon. N Y Times. May 14, 2025. Accessed October 29, 2026. https://www.nytimes.com/2025/05/14/technology/ai-jobs-radiologists-mayo-clinic.html