AI in Medicine 2026: What’s Proven vs What’s Only Promised

The Angle. AI in medicine is real, but narrower than the headlines. It is already very good at looking (scans, retinas, heart traces) and it cracked protein folding. It is unproven at deciding and still preclinical at operating. The gap that matters now is not between AI and doctors. It is between approval and proof.

The Trick

In January 2026, The Lancet published the final results of MASAI, a Swedish trial that randomised 105,915 women to have their mammograms read with or without AI support. Some of the coverage that followed said AI had cut missed cancers by 12%.

That is the trick. Here is the reveal. The trial did record 1.55 interval cancers† per 1,000 women in the AI arm, against 1.76 in the standard arm. But the confidence interval ran from a 35% reduction to an 18% increase. In plain words, the trial cannot tell us that AI reduced missed cancers.

What it did show is still remarkable. AI-supported reading was non-inferior to standard double reading and more sensitive (80.5% against 73.8%). An earlier analysis of the same trial found radiologists' reading workload fell by 44%.

8 reach strong evidence, 7 did not survive testing, and the rest sit somewhere in between.

The short version

Mojo's Law: Bayes' theorem

A test can be right almost all the time and still be wrong most of the times it raises the alarm. The reason is an 18th-century formula that every medical AI pitch should be forced to show.

\text{PPV} = \frac{\text{Se} \times p}{\text{Se} \times p + (1 - \text{Sp}) \times (1 - p)}

PPV is the positive predictive value: the chance that an alarm is real. Se is sensitivity, Sp is specificity, and p is how common the condition is in the people being tested.

Try it with round numbers (illustrative, not from any one product). A tool with 90% sensitivity and 95% specificity, used where 1 person in 100 has the disease, flags 90 true cases and 495 false ones in every 10,000 people. Its PPV is about 15%. Roughly 85 of every 100 alarms are false.

This is not a thought experiment. When Michigan Medicine tested Epic's widely used sepsis model on 38,455 hospital stays, its PPV was 12%. Accuracy on a tidy test set is not usefulness on a busy ward. That gap is why the rest of this piece cares less about benchmark scores and more about what happened to real patients.

How many AI tools are doctors actually allowed to use?

The US Food and Drug Administration's list of AI-enabled medical devices held 1,614 entries in its 4 September 2026 update, covering decisions up to 29 June 2026. In July 2025 it held 1,247. 2025 was the record year, with 335 authorisations.

Radiology dominates. About 76% of listed devices were reviewed by the radiology panel, with cardiovascular devices a distant second at about 9.5%. In every year since 2016, radiology's share has stayed above 59%.

Now the catch. About 96% of these devices (1,553 of 1,614) reached the market through the 510(k) route, which asks a maker to show its device is substantially equivalent to one already sold, rather than to prove a new patient benefit.

An audit of 1,357 FDA-listed AI devices, published in PLOS Digital Health in August 2026, found that 34 (2.5%) were linked to a registered prospective trial and just 3 (0.2%) had been evaluated on patient-centred outcomes such as death, illness or readmission.

The FDA itself says the list is not comprehensive and does not yet flag devices built on large language models; it says it is exploring how to tag them. So the honest summary is this: a clearance is a driving licence, not a safety record.

Where has AI already proved itself?

AI has proved itself where medicine is a matter of looking. Six of the eight claims on my ladder with strong evidence are about screening and imaging, and that is no accident: finding patterns in pixels is exactly what deep learning was built to do.

Breast screening. The full screening round of MASAI, published in The Lancet Digital Health in February 2025, covered about 106,000 women. AI-supported reading found 338 cancers against 262 with standard double reading, 29% more, with no significant rise in false positives and 44% less reading work.

In Germany, the PRAIM study followed 463,094 women across 12 sites and found detection rose from 5.7 to 6.7 per 1,000 (Nature Medicine, January 2025). PRAIM was not randomised and one co-author works for the software maker, so it supports MASAI rather than doubling it.

Colonoscopy. A 2024 meta-analysis of 44 randomised trials and 36,201 people found AI-assisted colonoscopy raised the adenoma detection rate‡ from 36.7% to 44.7% (Annals of Internal Medicine). That earns the top grade. The fine print: the extra finds are mostly small polyps, and advanced lesions per procedure were similar.

And there is a sting. In a 2025 Polish study of 19 experienced endoscopists, detection without AI fell from 28.4% to 22.4% after AI arrived in their clinics (Lancet Gastroenterology & Hepatology). It is observational, but it is the clearest warning yet that the assistant can quietly weaken the expert.

Eyes. The IDx-DR system became the first autonomous AI cleared by the FDA in 2018, after a pivotal study of 900 people with diabetes showed 87.2% sensitivity and 90.7% specificity for referable eye disease.

In a 2024 randomised trial among 164 young people with diabetes, offering the AI exam at the clinic visit raised completed eye screening within six months from 22% to 100% (Nature Communications). That second result matters more than the first: the best test is the one people actually take.

Heart. In the EAGLE trial, a pragmatic cluster RCT of 22,641 adults, an AI reading of a routine ECG raised new diagnoses of a weak heart pump from 1.6% to 2.1% within 90 days (Nature Medicine, 2021). Small in absolute terms, large for the patients it found.

TB and pancreas. Since 2021 the WHO has allowed computer-aided detection to replace human readers of chest X-rays for TB screening in people aged 15 and over, a conditional recommendation on low-certainty evidence. For countries with a heavy TB burden, India among them, that is a big practical door.

And PANDA, a model that looks for pancreatic cancer on ordinary non-contrast CT scans, reached 92.9% sensitivity and 99.9% specificity in a real-world test on 20,530 patients (Nature Medicine, 2023). It has not yet been tested on whether it saves lives.

Can AI diagnose better than a doctor?

On paper, increasingly yes. In a clinic, we do not know yet, and the first real test came back flat.

The paper wins. Microsoft's MAI-DxO system, which works through a case like a panel of doctors ordering tests, was published in Science on 22 October 2026. According to reports of the paper, it solved 83.2% of 304 hard New England Journal of Medicine case records, while physicians allowed to use reference material solved 36.1% of a held-out set.

All of the authors work for Microsoft, and the cases are simulated. Google's AMIE, tested in a randomised, double-blind study published in Nature in 2025, was rated better than 20 primary-care doctors on 30 of 32 axes by specialists, across 159 role-played consultations by text chat.

The awkward middle. A 2024 randomised trial in JAMA Network Open gave 50 physicians either GPT-4 or their usual resources for diagnostic vignettes. Doctors with the model scored 76%, against 74% without it, a difference that could be chance. The model on its own scored 16 points higher than the doctors without it. The tool helped the machine more than it helped the human, which says as much about how we design the partnership as about the model.

The first real test. In June 2026, Nature Medicine published a pragmatic cluster-randomised trial of a generative-AI decision tool across 16 primary-care facilities in Kenya, covering 9,691 patients. Treatment failure at 14 days was 2.2% with the tool and 2.0% without, with no significant difference (adjusted odds ratio 0.77, 95% CI 0.55 to 1.08).

The tool was safe; it just did not measurably help. An earlier company-reported study with Penda Health in Kenya found 16% fewer diagnostic errors and 13% fewer treatment errors, but again no difference in how patients felt a week later.

Sepsis, the cautionary tale. In 2022, a prospective study of the TREWS early-warning system across five hospitals found in-hospital deaths 18.7% lower (relative) when clinicians confirmed its alert within three hours. That study was not randomised. A year earlier, an external check of Epic's proprietary sepsis model at Michigan Medicine found an AUC of 0.63, against the 0.76 to 0.83 claimed, and the model missed 67% of sepsis cases while alerting on 18% of all patients.

Paperwork, the quiet win. Ambient AI scribes, which listen to a consultation and draft the note, have now been through randomised trials. In a three-arm UCLA trial of 238 physicians, one product cut time in notes by 9.5% and the other made no significant difference (NEJM AI, 2025). A separate stepped-wedge trial of 66 clinicians found work exhaustion fell modestly. Physicians also reported occasional clinically significant errors in the drafts.

My read:

Is AI discovering new cures?

AI has changed the start of drug discovery and has not yet delivered at the finish line. As of November 2026, no AI-discovered drug has full approval from a major regulator.

The genuine breakthrough. AlphaFold predicted the structures of around 200 million proteins and won Demis Hassabis and John Jumper half of the 2024 Nobel Prize in Chemistry. AlphaFold 3, published in Nature in May 2024, extends the method to proteins bound to DNA, RNA and drug-like molecules. This is as close to proven as science gets. But a protein's shape is one step in a long chain from target to medicine.

The pipeline numbers. A 2024 analysis in Drug Discovery Today found AI-discovered molecules passed Phase I trials 80 to 90% of the time, against a historical 40 to 65%. At Phase II, where a drug must show it works, the success rate was about 40%, in line with history. Phase I mostly tests safety; Phase II is where hope meets biology.

The front-runner. Insilico Medicine's rentosertib, a drug for the lung-scarring disease idiopathic pulmonary fibrosis, was designed with generative AI. In a Phase 2a trial of 71 patients in China, its primary safety goal was met, and lung capacity (FVC) rose 98.4 mL over 12 weeks on the 60 mg dose against a 20.3 mL fall on placebo, an exploratory result (Nature Medicine, June 2025). A Phase III trial of 320 patients at 47 Chinese centres began in July 2026. The company says the drug is not approved anywhere.

The definitional fight to come. Takeda's psoriasis pill zasocitinib, which originated at Nimbus Therapeutics using computational design, had its new drug application accepted with Priority Review on 14 September 2026, with a decision due in the first quarter of 2027. Nimbus has credited algorithms with speeding its discovery; Takeda's announcement does not mention AI. If it is approved, expect a loud argument about whether it is “the first AI drug”.

Antibiotics and slippage. AI-guided screens found halicin (Cell, 2020) and abaucin (2023), and in August 2025 MIT researchers used generative AI to design compounds, NG1 and DN1, that cleared drug-resistant gonorrhoea and MRSA in mice (Cell). All three remain preclinical.

Isomorphic Labs, the Alphabet drug spin-out, had no confirmed first human dosing by the time of writing, after Hassabis moved the expected first trials from end-2025 to end-2026. Recursion ended two clinical programmes, REC-994 and REC-2282, in May 2025.

The law that matters here is Eroom's law, Moore's law spelt backwards: since 1950 the number of new drugs approved per billion dollars of R&D has halved roughly every nine years (Scannell and colleagues, Nature Reviews Drug Discovery, 2012). AI's real test is not a dazzling molecule. It is whether that curve finally bends.

Is AI doing operations yet?

Not on people. Robot-assisted surgery is routine, but a human surgeon is at the controls. Fully autonomous surgery remains in the lab.

In 2022, Johns Hopkins' STAR robot stitched pig intestine better than human surgeons on the same task, using marked tissue and a preset plan. In July 2025, the team's SRT-H system, trained by watching surgical video and taking spoken corrections, removed gallbladders from eight pig organs outside the body with a 100% success rate, more slowly than a surgeon (Science Robotics).

A May 2026 review in Nature Reviews Urology described soft-tissue autonomy as still confined to ex vivo and animal models. I found no verified autonomous soft-tissue operation on a human patient.

The Mojo Evidence Ladder: what is proven and what is promised?

I took 31 of the most repeated claims about AI in medicine and graded each against its best primary evidence on the Mojo Evidence Scale, from MES 5 (proven by replicated randomised trials) to MES 0 (tested and did not hold up). Only 8 of the 31 reach strong evidence. Six of those eight are in screening and imaging.

Three things jump out. Diagnosis and decisions, the area that gets the most headlines, has no claim above MES 3 and three that failed when tested. Drug discovery's only top-grade claim is AlphaFold, a research tool rather than a medicine. And in safety, bias and oversight, the red bar is the reassuring claims that broke: that algorithms are fair by default, that transcription is verbatim, and that FDA clearance means proven benefit.

Claim Check

Claim	Who makes it	MES	Mojo's read
AI mammography finds more cancers with less work	MASAI and PRAIM teams	4 Strong	Real gains; one randomised trial in one country so far
AI cut missed breast cancers by 12%	Coverage of MASAI, 2026	2 Early	Non-inferior, but the reduction is not significant
AI colonoscopy raises adenoma detection	44-trial meta-analysis	5 Proven	Mostly small polyps; watch for deskilling
AI beats doctors at diagnosis	Microsoft, Google	3 Promising	True on simulated cases, not yet on patients
AI decision tools improve patient outcomes	Vendors, pilots	0 Didn't hold up	First outcome RCT was null; more trials needed
Protein structure prediction is solved	DeepMind, Nobel committee	5 Proven	Structure is one step of many in drug design
AI drugs are reaching patients	AI-first biotechs	1 Claim	No full approval yet; first decision due Q1 2027
Robots now operate on their own	Research labs	2 Early	Ex vivo and animal work only
FDA-cleared means proven to help	Marketing copy	0 Didn't hold up	0.2% of audited devices have outcome evidence

The Counter-Angle

Here is the strongest case that I am being too harsh.

All fair. That is why every grade on the ladder is dated and will be revised each quarter. But medicine has outcome trials for a reason: confident tools that looked good on paper have harmed patients before. The burden of proof belongs with the people selling the tool, not with the patient.

Mojo's Forecasts

1. The FDA approves zasocitinib by 31 March 2027. Probability: 75%. If it does, I will grade the “first AI-designed drug” claim separately, because the definition matters.
2. No peer-reviewed report of an autonomous robot completing a soft-tissue operation on a human patient by 31 December 2027. Probability: 80%.
3. An FDA list update published by 31 December 2027 shows more than 2,000 AI-enabled devices. Probability: 70%.

All three go into the Mojo Forecast Ledger and will be scored in public.

Where this is debated next

FAQ

Margin notes

† Interval cancers are cancers diagnosed between screening rounds after a normal screen. They are the best available proxy for cancers a screen missed.

‡ Adenoma detection rate is the share of colonoscopies in which at least one adenoma, a polyp that can turn cancerous, is found.

The magic in medical AI is real. It just has to keep showing its working.