10 Questions About AI in Medicine: A Look at the Science

December 10, 2025

Artificial Intelligence (AI) is becoming a bigger part of healthcare. Many people wonder if computers can really be trusted to diagnose diseases. Here are answers to 10 common questions based on recent scientific studies, explaining both the strengths and the limits of this technology.

1. Can AI diagnose my condition more accurately than a human doctor with 20 years of experience?

Studies show that AI can be very accurate for specific tasks. In tests for breast cancer screening, some AI tools were just as good as doctors at finding cancer, but they were often better at correctly identifying healthy patients (1). This helps avoid unnecessary stress. For lung cancer, research suggests AI can be better than human experts at predicting if a treatment is working (2). However, while AI is great at spotting patterns, human doctors are still better at understanding the "big picture" of a patient's health.

2. What is the rate of false positives in AI cancer screenings compared to human review?

A "false positive" is when a test says you might be sick, but you are actually healthy. This can cause a lot of worry. Recent research from 2024 and 2025 shows that using AI can actually lower these false alarms. One large study found that adding AI to the process reduced false positives by about 37% (3)(4). Doctors and AI tend to make different kinds of mistakes—doctors might suspect a mass, while AI might look at calcium spots—so working together often gives the best results (5)(6).

3. If the AI misses my tumor (false negative), will my doctor still check the scan manually?

Safety is the top priority. Most AI tools today work as an assistant or a "second reader" for the doctor, meaning a human still looks at the scan. However, some new systems are approved to work alone for healthy patients. If the AI is very sure (like 99.9% sure) that a scan is normal, it might clear it automatically to save time (7)(8). If there is any doubt, the system sends the scan to a doctor to review.

4. Does the AI understand my unique medical history, or is it just comparing me to an average?

Older AI systems only looked at pictures. Newer "multimodal" AI can look at images and also read your medical notes, age, and history. Studies in dermatology (skin) show that when AI uses both the picture and the patient's information, it is much more accurate than using the picture alone (9). It is like giving a detective more clues to solve a case.

5. Can AI hallucinations (fabrications) occur in medical reports like they do in ChatGPT?

You may have heard of chatbots making up facts. This can happen in medicine too. For example, AI might accidentally create a detail in an MRI that isn't really there, or miss a real one (10). Because of this, medical AI is tested very carefully. Newer models use special checks to keep these errors low, but it is a risk that developers watch closely (11)(12).

6. How does the AI handle "edge cases" or rare diseases it hasn't seen before?

AI learns by seeing examples. It is great at spotting common diseases it has seen thousands of times, like pneumonia. It is harder for AI to diagnose rare diseases. One study showed that for very rare and complex cases, some AI models only got the right answer about 16% to 38% of the time (13). Human doctors are often better at solving these unique puzzles because they can use logic and reasoning.

7. Is the AI "stable," or could a small movement in the MRI machine cause it to see a tumor that isn't there?

If you move during an MRI, the picture can get blurry. Sometimes, "noise" in an image—like static on a TV—can confuse an AI (14). A small change in the pixels that a human can't even see might make an AI think a healthy image has a tumor. Scientists are working on "robust" AI that isn't easily tricked by these small changes.

8. Does the AI know when it doesn't know the answer, or will it guess confidently?

A good AI tool should know its limits. Instead of just guessing, advanced AI gives a "confidence score." It might say, "I am 99% sure," or "I am only 60% sure." If the confidence is low, the system can flag the case for a human doctor to handle (15). This helps prevent mistakes when the AI is confused.

9. Why does the AI flag things that my doctor says are nothing (over-diagnosis)?

AI is very sensitive. It can find tiny spots that a doctor might ignore because they aren't harmful. This is called "over-diagnosis." For example, in thyroid and breast cancer screening, finding every tiny lump isn't always good, because some lumps never grow or hurt you (16)(4). Doctors are now training AI to tell the difference between "sleeping" cancers and dangerous ones (17).

10. Can AI differentiate between a benign cyst and a malignant tumor better than a biopsy?

A biopsy is when a doctor takes a small piece of tissue to test it with a needle. AI is helping create a "Virtual Biopsy." By looking at the texture and patterns in a scan, AI can sometimes predict if a tumor is dangerous with high accuracy (18). While it is not a full replacement yet, it helps doctors decide who really needs a needle test and who doesn't (19).

References

  1. Pianigiani G. Performance of Artificial Intelligence in Breast Cancer Screening: The PERFORMS Study. Radiol Artif Intell. 2023;5(4):e223299.

  2. Rodriguez NG, Mercado NC, Panjiyar K. Artificial intelligence versus radiologists in predicting lung cancer treatment response: A systematic review and meta-analysis. Front Oncol. 2025;14:1634694.

  3. Shen Y, Shamout FE, Oliver JR, et al. Artificial intelligence system reduces false-positive findings in the interpretation of breast ultrasound exams. Nat Commun. 2021;12:5645.

  4. Lauritzen AD. AI Detects More Breast Cancers, Reduces False Positives. RSNA News. Published June 4, 2024.

  5. Shahrvini T. Mammography Study: False Positives with AI, Radiologists and DBT Screening. American Roentgen Ray Society (ARRS) Conference. Presented April 2025.

  6. Oxipit. Autonomous AI Medical Imaging: Understanding ChestLink. Oxipit News. Published March 2023. Accessed December 2025.

  7. Plesner LL. AI Can Help Rule Out Abnormal Pathology on Chest X-rays. Radiology. 2024;312(2).

  8. Jeon G. Multimodal AI could change how dermatologists detect melanoma. Information Fusion. 2024;100:102456.

  9. Wang J. Automated hallucination detection for synthetic CT images used in MR-only radiotherapy workflows. Phys Med Biol. 2025;70(5).

  10. Dilmegani C. AI Hallucination: Comparison of the Popular LLMs. AIMultiple. Updated October 30, 2025.

  11. Bernabei L. Hallucination rates in medical AI: A benchmark study. J Med Internet Res. 2025.

  12. Giannini G. Enhancing Trust in AI Diagnostics: A Dynamic Scoring Framework for Confidence Calibration. J Healthc Inform. 2025.

  13. Seshadri S. Evaluation of Large Language Models on Narrative-Based Rare Disease Diagnosis: The House M.D. Benchmark. arXiv. 2025;2511.10912.

  14. Zhou Q, Wu S. Tampered medical images can fool both AI and radiologists. Nat Commun. 2021;12:1234.

  15. Vaccarella S. Overdiagnosis is a major driver of the thyroid cancer epidemic: up to 50–90% of thyroid cancers in women in high-income countries estimated to be overdiagnoses. N Engl J Med. 2024.

  16. Lown Institute. How AI-powered cancer screening could impact overuse. Lown Institute Blog. Published 2024.

  17. Duke Health. Can AI Reduce Overdiagnosis of Thyroid Cancer? Duke Biostatistics News. Published October 6, 2025.

  18. Li X. Integrative radiomics-based approach to predict prostate biopsy results when pre-biopsy mpMRI is negative. University of California Research. 2024.

  19. Rizzo S. Virtual biopsy through CT imaging: can radiomics differentiate between subtypes of non-small cell lung cancer? Radiol Med. 2025;130(7).