From second readers to a restored voice: what medical AI has shown

A medical AI result becomes much more informative when we can answer a few ordinary questions. Who used it? What happened to patients? What would have happened without it? And how much uncertainty remains?

An accurate prediction on a test dataset is one piece of evidence. A randomized screening trial, an observational hospital evaluation and an experimental device used by one person answer different questions. Treating them as interchangeable makes promising work harder to judge.

The examples below show several ways AI can become useful, with very different levels of clinical evidence.

Breast screening: stronger evidence, specific conclusions

Sweden's MASAI trial randomly assigned 105,934 women to standard double reading or AI-supported screening. AI triaged examinations to one or two radiologists and highlighted suspicious findings. The 2025 screening-performance analysis reported 29% higher cancer detection and 44.2% fewer screen readings, with no statistically significant increase in false positives. That workload measure counts readings; it does not mean an equivalent reduction in total staffing or costs.

The 2026 primary-outcome report found interval-cancer rates of 1.55 versus 1.76 per 1,000 participants. The ratio was 0.88, with a 95% confidence interval of 0.65–1.18. AI met the trial's non-inferiority criterion; the apparent 12% reduction did not establish superiority. Sensitivity was 80.5% versus 73.8%, while specificity was 98.5% in both groups. These are screening results, not evidence of lower breast-cancer mortality.

Germany's observational PRAIM study provides a different kind of evidence. It covered women aged 50–69 at 12 sites. Its adjusted estimates were 6.7 versus 5.7 cancers detected per 1,000, a 17.6% relative difference, without a statistically significant change in recall rates. Radiologists chose whether to use AI, and the analysis adjusted for observed selection bias. Residual confounding remains possible. The AI vendor, Vara, funded and participated in the study.

Research references: [1] [2] [3]

Stroke: promising associations inside a care network

An English evaluation used national audit data from 107 hospitals and examined Brainomix implementation at 26 sites. Thrombectomy rates rose from 2.3% to 4.6% at evaluation sites, versus 1.6% to 2.6% elsewhere. Among patients transferred from primary to comprehensive stroke centres, AI use was associated with a 64-minute shorter median door-in-door-out interval. This measures time at the referring hospital, not total time to treatment. ,

Because this was observational, concurrent changes in stroke services and differences between patients could contribute to the results. The comparison hospitals also increasingly used AI. NHS funding and analysis independent of the manufacturer are relevant, as are several authors' disclosed Brainomix interests. The findings support further evaluation of an integrated care pathway; they do not isolate a software-only causal effect.

Research references: [4] [5]

Tuberculosis: useful screening needs a diagnostic pathway

WHO has recommended computer-aided detection for chest-X-ray TB screening since 2021. Its June 2025 assessment found six products meeting its performance standards for people aged 15 and older. That assessment did not extend the recommendation to people younger than 15.

CAD can substitute for a human reader for this defined screening task. Programs still need locally calibrated thresholds, ongoing monitoring and follow-up diagnostic testing. A screening score does not establish a TB diagnosis. This example also shows why describing every medical AI as a “second reader” is too narrow. ,

Research references: [6] [7] [8]

Discovery: useful science, unfinished clinical development

The 2024 chemistry Nobel recognized David Baker's computational protein design and Demis Hassabis and John Jumper's protein-structure prediction. At the award date, the Academy described AlphaFold2 predictions covering roughly 200 million known proteins and use by more than two million people in 190 countries. These are research-scale measures.

AlphaFold 3 predicts structures of complexes involving proteins, DNA, RNA and other molecules. Such predictions can help researchers choose experiments, but do not by themselves demonstrate that a treatment works.

Rentosertib, developed using AI-assisted target identification and molecule design, reached a randomized phase 2a trial involving 71 adults with idiopathic pulmonary fibrosis in China. Safety was the primary endpoint. Over 12 weeks, the highest-dose group had a mean forced vital capacity change of +98.4 mL, versus −20.3 mL with placebo. This secondary endpoint is an early efficacy signal. Seven participants discontinued because of liver injury or dysfunction; acute exacerbations occurred in three highest-dose patients and one placebo patient. The small, short study cannot establish durable benefit or long-term safety.

There has been progress since that publication: on September 10, 2026, developer Insilico announced first dosing in a phase 3 trial. That is a company-reported development milestone, not a new efficacy result or marketing approval.

Antibiotic work remains earlier. A 2025 Cell study synthesized 24 AI-designed compounds, seven of which showed selective antibacterial activity. Two leads reduced bacterial burden in mouse models of gonorrhoea and MRSA infection. Those findings warrant further development; they do not establish human safety, treatment success or a cure.

Research references: [9] [10] [11] [12] [13]

A restored channel for communication

For Casey Harrell, a BrainGate2 participant living with ALS, an implanted interface has supported daily communication. A 2024 study reported sustained 97.5% word accuracy with further training over 8.4 months.

A June 2026 paper reported more than 3,800 hours of home use, close to two million words and an average 56 words per minute. He could operate it without researchers present, while trained care partners helped fit the hardware and start the software. It remains a single-participant investigational result involving surgery and wired equipment. Its value to one person is substantial; broader applicability still requires evidence. ,

Research references: [14] [15] [16]

The practical lesson

My reading of these cases is that useful medical AI needs a clearly bounded job, a workable human support system and an evaluation designed around the people affected. Sometimes the job is assisting an expert. Sometimes it automates one screening task or gives a person another way to communicate.

For builders and health systems, the next questions should be concrete: Does the result hold in this population? What happens after an alert? Who handles errors? What does the whole service cost? Are benefits sustained, and are harms being measured?

These selected studies do not establish that medical AI generally reduces mortality or that results transfer unchanged between settings. They do provide reasons to keep testing specific tools carefully, and to describe progress at exactly the scale the evidence supports.

Next for Humanity follows how technology can support human life, with evidence and limitations made visible.