# Artificial intelligence mammogram: 91% false positives at 1% prevalence — biopsy or watch?

Alex Rivera · October 25, 2026

> At 1% prevalence, 91% of AI-flagged mammograms are false positives; learn base-rate math and how posterior probability guides biopsy versus watch.

| Takeaway | Detail |
| --- | --- |
| At 1% prevalence, 91% of AI-flagged mammograms are false positives | Base-rate math with roughly 90% sensitivity and 91% specificity produces the headline figure. |
| Biopsy-versus-watch decisions should use posterior probability, not the raw flag | Convert sensitivity, specificity, and population prevalence with a 2x2 natural-frequency table. |
| In reviewed active-assisted-living literature, activity recognition was the leading AI application goal (34/64, 53%) | Daily monitoring was next at 10/64 (16%). |
| AI safety review includes identifying safety concerns and trustworthy-AI challenges | OpenClassrooms lists AI safety concerns, challenges in creating responsible and trustworthy AI, and workplace impact assessment. |

 This guide explains how base-rate math turns a 91% false-positive headline into a posterior probability for biopsy decisions.

 It shows how to apply a 2x2 natural-frequency table to any rare-condition AI alert.

## How the 91% Number Is Actually Made

 The “91% of AI-flagged mammograms are false positives at 1% prevalence” headline is not a performance review of an AI system; it is the ordinary arithmetic of a rare condition tested with a good but imperfect screen. The useful question is not whether the headline is alarming, but what posterior probability it implies for the person who received the flag.

 Work it through with natural frequencies. In 10,000 screens at 1% prevalence, 100 people have cancer and 9,900 do not. At about 90% sensitivity, the AI flags 90 of the 100 cancers and misses 10. At about 91% specificity, it also flags about 899 of the 9,900 healthy screens, leaving about 9,001 true negatives. That produces 989 total flags, of which 899 are false, so 899 ÷ 989 ≈ 91% of positive flags are false positives—exactly the headline, derived from prevalence and accuracy rather than from any vendor-specific failure.

| AI result | Cancer present | No cancer | Total |
| --- | --- | --- | --- |
| Flagged | 90 | 899 | 989 |
| Not flagged | 10 | 9,001 | 9,011 |
| Total | 100 | 9,900 | 10,000 |

 The arithmetic is stable: change prevalence or specificity and the false-positive share changes with them. A lower prevalence makes the share rise even if sensitivity and specificity stay the same; a higher specificity makes it fall. That is why the headline cannot be read as “the AI is wrong 91% of the time” in any individual case.

 The reason AI can feel worse than a human reader is not necessarily that its error rate is higher. A radiologist’s “flag” is layered judgment under uncertainty plus recall, so uncertainty is spread across callbacks, comparisons, and follow-up; a single binary AI alert compresses that same uncertainty into one alarming message.

 Before acting on any such alert, convert sensitivity, specificity, and your population’s prevalence into a posterior probability with this same 2×2 natural-frequency method: true positives divided by all positives. The resulting number is the probability that this flagged screen represents cancer; it is the input to the biopsy-versus-watch decision, not a substitute for it.

![How the 91% Number Is Actually Made — Artificial intelligence mammogram](https://static.mm-ais.com/article-images-ai/artificial-intelligence-mammogram-91-fal-ai-419f9620.jpg)

## The Evidence: What Grounding Actually Supports

 This essay's grounding contains no mammography-AI trial data: none of the available sources report a randomized or paired-reading mammography study, a product's sensitivity or specificity, or a screening prevalence for a defined population. The method for sourcing it is therefore a literature check rather than a citation, and every performance claim should be treated as unverified until it is traced to a primary study or a regulator's labeled summary.

 Source check one is a PubMed search. Search PubMed (pubmed.ncbi.nlm.nih.gov), the index used in this essay's own source base, for randomized or paired-reading trials of AI mammography, including MASAI-style designs. From each paper, record sensitivity, specificity, and screening prevalence directly from the methods and results; do not take them from press coverage, which can round, select, or omit the denominator.

| Check | Record from the primary source | Do not use |
| --- | --- | --- |
| PubMed trial | Sensitivity, specificity, and screening prevalence | Press coverage or rounded headlines |
| FDA clearance | Labeled performance and the operating threshold | Marketing claims or unverified summaries |

 Source check two is regulatory clearance data. For any named product, verify its FDA clearance summary and labeled performance, which states the operating threshold used in the cleared indication. Compare that threshold with the prevalence in the population where the tool will be used; a clearance summary is a regulatory record, not proof that the same operating point holds in a different screening population.

 If a paper or clearance summary lacks one of the three numbers, do not fill the gap with a headline figure or an average from another study. Record the missing input, use the available two numbers only for a bounded range, and mark the posterior as provisional. The decision rule can still be stated in method terms: posterior probability is true positives divided by all positive flags, and the inputs must come from the same population and threshold.

 Keep the source trail visible in the final note: name the PubMed record or FDA summary, the population, the threshold, and the date checked. That makes the evidence auditable without turning a regulatory label or a single trial into a universal performance claim.

## Alarm vs. Posterior: Three Responses Compared

 The action after an AI mammogram flag should follow the posterior probability, not the alarm. Convert sensitivity, specificity, and your population's prevalence into a natural-frequency posterior first, then choose the response with the lower expected harm.

 Option A, biopsy on every flag, maximizes action sensitivity: it is least likely to leave a flagged cancer unbiopsied. But at a ~9% posterior—about 90 true cancers among roughly 989 flagged in 10,000 screened—it sends about 10 false-positive patients to an invasive procedure for every true cancer found. That cost and harm per cancer found is why it loses when the posterior is low.

 Option B, ignore the AI entirely, restores the pre-AI baseline and avoids AI-driven false positives. It loses whenever the flag carries validated information beyond the radiologist's read, because it forfeits that sensitivity uplift.

 Option C, the threshold rule, compares the two actions directly: watch and retest when the posterior is low, and biopsy only when it is high enough. The threshold is the posterior at which the expected harm of an unnecessary biopsy equals the expected harm of delaying a cancer diagnosis. Below it, watch-and-retest wins by avoiding invasive procedures for patients whose probability of cancer remains low; above it, biopsy wins because cancer is likely enough to justify the procedure.

| Posterior band | Preferred action | Why it wins |
| --- | --- | --- |
| Below the harm-balance threshold | Watch and retest | Avoids invasive procedures when most flagged patients do not have cancer |
| Above the harm-balance threshold | Biopsy | Cancer is likely enough to justify the procedure's cost and harm |
| Every flag, regardless of posterior | Biopsy all | Catches the most cancers but imposes the highest false-positive harm per cancer found |

 Thus the choice is not between trusting and distrusting the AI; it is a posterior-driven choice among three responses, with the threshold set by the relative harms you elicit rather than by the headline's false-positive rate.

## The Costs That Decide Biopsy vs. Watch

 Biopsy versus watch is a cost comparison, not a reflex. The two numbers that should decide it are the cost of delaying a cancer through the watch interval and the cost of an unnecessary biopsy; both must be elicited from local sources rather than borrowed from an article.

| Error type | Cost to quantify | Source to use | Source to reject |
| --- | --- | --- | --- |
| False positive: unnecessary biopsy | Procedural cost, complication rate, anxiety, overdiagnosis risk | Provider's published fee schedule and complication statistics | Generic figure from an article |
| False negative: delayed diagnosis | Stage-shift risk over a 3–6 month watch period | Follow-up data for the same population and pathway | National average or anecdote |

 For a false positive, the relevant cost is not a headline price. It is the full burden of an unnecessary biopsy: procedural cost, anxiety, and overdiagnosis risk. Get the procedural cost and complication rate from the provider's own published fee schedule and complication statistics; do not accept a generic figure from an article. If a provider will not publish both, treat the missing number as uncertainty in the decision, not as zero.

 For a false negative, the relevant cost is the change in treatment options and survival from a delayed diagnosis. Quantify it as the stage-shift risk of a 3–6 month watch period, sourced from follow-up data for the same population and screening pathway. Do not substitute a national average or an anecdote for that local follow-up estimate.

 Then compare the two on the same scale: expected harm per flagged mammogram. A watch period is favored when the expected harm of an unnecessary biopsy exceeds the expected harm of the delay at the posterior probability in hand; biopsy is favored when the reverse is true. The crossover point is a local number, because both costs are local.

 Keep the two errors separate in the record. Anxiety and overdiagnosis are not captured by a fee schedule; stage shift is not captured by a single survival percentage. Elicit each as its own quantity, then combine them explicitly rather than letting one dominate by default.

## What the Evidence Does Not Establish

 The 1%-prevalence rule is a conditional screening heuristic, not a universal prior. Before acting on an AI flag, verify four things: the population is average-risk; prevalence is stated for that exact cohort; sensitivity and specificity are published for the deployed threshold; and the decision is screening rather than diagnosis. If any check fails, the headline's posterior is not your posterior.

 The rule breaks in high-risk surveillance. Where prevalence is several-fold higher than the average-risk baseline, the same sensitivity and specificity produce a higher posterior for each positive flag, and a watch-first default may under-treat. A higher prior can make biopsy the better default even when test performance is unchanged. Recompute prevalence for that cohort before applying the rule; do not import the average-risk posterior.

 The rule also breaks in diagnostic imaging. A symptomatic patient is not a screening population, so the AI flag operates on a different prior entirely. Screening base-rate math does not transfer, and the flag should be weighed with clinical findings and the diagnostic prior rather than with screening prevalence.

 The rule still wins in average-risk population screening at roughly 1% prevalence, and in any setting where the vendor's threshold, sensitivity, and specificity are published. There, convert those three inputs into a posterior probability with a 2x2 natural-frequency table before choosing biopsy or watchful waiting. If the metrics are unpublished, treat the flag as unquantified and fall back to standard care instead of to the headline's false-positive rate.

 Use the table as a routing check: a "breaks" setting means stop and recompute; a "still wins" setting means proceed to the posterior calculation.

| Setting | Effect on the 1%-prevalence rule | Required action |
| --- | --- | --- |
| High-risk surveillance (prevalence several-fold higher) | Breaks: posterior per flag rises; watch-first may under-treat | Recompute prevalence before applying the rule |
| Diagnostic/symptomatic imaging | Breaks: different prior; screening base-rate math does not transfer | Use diagnostic prior and clinical findings, not screening prevalence |
| Average-risk screening at ~1% prevalence | Still wins | Compute posterior from sensitivity, specificity, and prevalence |
| Any setting with published vendor threshold, sensitivity, and specificity | Still wins | Convert the three inputs into a posterior probability using a 2x2 natural-frequency table |

 Also worth reading: **What is a pre-mortem? CATCS 2030 planning turns uncertainty into decisions**: [What is a pre-mortem? CATCS](/what-is-a-pre-mortem-catcs-2030-planning-turns-uncertainty-into-decisions/) · **The tribal mind and institutional trust breakdown**: [tribal mind and institutional trust](/the-tribal-mind-and-institutional-trust-breakdown/) · **Why managers trust automation: $11 review vs $18,000 error**: [Why managers trust automation: $11](/why-managers-trust-automation-11-review-vs-18000-error/)

## And Your One-Page Worksheet

 Copy this five-line natural-frequency worksheet and use it before acting on any rare-condition AI alert. It converts three inputs—prevalence, sensitivity, and specificity—into counts and then into a posterior probability, so the next step is based on the chance that this particular flag is true rather than on the alarm itself.

## Frequently Asked Questions

 **What does the 91% false-positive figure actually describe?**

 It is the ordinary arithmetic of a rare condition tested with a good but imperfect screen, not a performance review of an AI system.

 **What should guide a biopsy-versus-watch decision after an AI flag?**

 Posterior probability, not the raw flag.

 **How can sensitivity, specificity, and prevalence be converted for a rare-condition alert?**

 With a 2x2 natural-frequency table.

 **Which AI application goal was most common in the reviewed active-assisted-living literature?**

 Activity recognition, at 34 of 64 studies.

 **What does the AI safety review cover?**

 OpenClassrooms lists AI safety concerns, challenges in creating responsible and trustworthy AI, and workplace impact assessment.

 **What is the more useful question than whether the headline is alarming?**

 What posterior probability the flag implies for the person who received it.

## Quick answers

| What does the takeaway say about AI-flagged mammograms at 1% prevalence? | The takeaway says that at 1% prevalence, the headline figure concerns false positives among AI-flagged mammograms. |
| --- | --- |
| What produces the headline figure? | Base-rate math with roughly the stated sensitivity and specificity produces the headline figure. |
| What should biopsy-versus-watch decisions use instead of the raw flag? | Biopsy-versus-watch decisions should use posterior probability, not the raw flag. |
| What tool converts sensitivity, specificity, and population prevalence? | A 2x2 natural-frequency table converts sensitivity, specificity, and population prevalence. |
| What does the guide explain and show? | The guide explains how base-rate math turns the headline figure into a posterior probability for biopsy decisions and shows how to apply a 2x2 natural-frequency table to any rare-condition AI alert. |

 Sources: [Com](https://www.businesstimes.com.sg/companies-markets/increasing-prevalence-artificial-intelligence)

Canonical: https://www.judgmentcallpodcast.com/2026/10/artificial-intelligence-mammogram-91-false-positives-at-1-prevalence-biopsy-or-watch/
Markdown: https://www.judgmentcallpodcast.com/2026/10/artificial-intelligence-mammogram-91-false-positives-at-1-prevalence-biopsy-or-watch/index.md