↓ Skip to main content
Expert LinkedIn
AI AI in Healthcare

AI Can Now Outdiagnose Doctors. That's the Easy Part.

8 min read
Dr. Sophia Patel
Dr. Sophia Patel AI in Healthcare Expert & Machine Learning Specialist

The number that will define this week in healthcare AI is 85.5 versus 20. Microsoft’s AI Diagnostic Orchestrator scored 85.5% on 304 of the most complex diagnostic cases in medical literature — cases drawn from the New England Journal of Medicine’s weekly teaching series, designed specifically to stump physicians — while a group of 21 practicing doctors, without access to their usual reference tools, averaged 20%. Four times the accuracy. Twenty percent lower cost.

A gleaming medical scale in dramatic chiaroscuro: the left pan holds a luminous AI sphere weighted with $2.1 billion, the right pan holds a small but bright beacon labeled 'equity' — the scale tips heavily left, but the right beacon casts a longer shadow across the dark floor
AI drug discovery got its biggest week ever. Who benefits depends on a question the market won’t answer.

But the number I keep returning to from this same seven-day window is a ratio: $2.1 billion to one program, $200 million to another. That split, more than any benchmark, tells you where clinical AI is actually going.

What the benchmark actually shows — and doesn’t
#

The Microsoft AI Diagnostic Orchestrator (MAI-DxO) result, highlighted in the Stanford HAI 2026 AI Index and covered extensively this week, is a genuine milestone. The system orchestrates multiple large language models — Claude, Gemini, GPT, Grok, DeepSeek, Llama — to mimic the consultative, step-by-step reasoning that physicians use in difficult cases: ask clarifying questions, order tests, revise the differential, decide. It outperformed physicians by a factor of four on one of medicine’s most notoriously difficult benchmarks. Dr. Eric Topol, Director of the Scripps Research Translational Institute, called it “more than previous studies have shown” — and Topol is not easily impressed.

Mustafa Suleyman, CEO of Microsoft AI, was characteristically direct: “I think it gives us a clear line of sight to making the very best expert diagnostics available to everybody in the world at an unbelievably affordable price point.”

That framing is important — and so are the caveats that came with it. MAI-DxO was tested on cases drawn specifically to confuse specialists. It was compared to physicians working without internet access or colleagues, which is not how diagnostic medicine actually works. The system hasn’t been approved by any regulatory body. The FDA hasn’t yet formally classified whether such systems are medical devices. Topol’s summary is worth keeping in view: “It doesn’t change medical practice until they take it out on the real medical highway.”

The benchmark tells us AI can now synthesize knowledge across specialties faster and more reliably than most individual humans. It does not tell us that AI diagnosis is ready for clinical deployment — let alone equitable global deployment. Those are different questions, and this week answered only the first one.

What $2.1 billion is actually building
#

On May 12, Isomorphic Labs — founded by Demis Hassabis in 2021, spun out of Google DeepMind — closed a $2.1 billion Series B round led by Thrive Capital, with participation from Alphabet, GV, MGX, Temasek, CapitalG, and the UK Sovereign AI Fund. Total capital raised now stands at approximately $2.6 billion.

The platform is real. IsoDDE — the Isomorphic Drug Design Engine, built on AlphaFold 3 — can predict molecular structures and interactions with unprecedented accuracy, modelling protein-ligand binding, antibody-antigen interactions, and binding affinity across therapeutic areas. Hassabis’s quote for the announcement was: “This capital injection allows us to build out our drug design engine at scale, driving us forward in our mission to solve all disease.”

The mission is compelling. The mechanics deserve scrutiny.

Isomorphic’s disclosed commercial partnerships are with Novartis, Eli Lilly, and Johnson & Johnson — the companies behind two deals collectively valued at nearly $3 billion — working on what the company describes as “undisclosed targets.” The company plans to enter its first in-house clinical trial by the end of 2026. These are not obviously neglected disease programs. Oncology, metabolic disease, and other high-revenue therapeutic categories have dominated Big Pharma’s AI drug discovery investments for a reason: the return on discovery is highest where pricing power is greatest.

“Solving all disease” in the language of venture capital means solving the diseases for which the market will reward solving. That set does not substantially overlap with the diseases carrying the greatest global burden. AI drug discovery is an accelerant. Before AlphaFold, pharma was already under-investing in neglected tropical diseases (NTDs). After AlphaFold, pharma will under-invest in NTDs faster — unless the incentive structure changes.

The $200 million that says what the $2.1 billion doesn’t
#

On May 17, Anthropic and the Gates Foundation announced a $200 million, four-year partnership focused on health and education in low- and middle-income countries, where 4.6 billion people currently lack adequate access to health services. The program targets drug and vaccine discovery for malaria, tuberculosis, HPV, and preeclampsia. Claude access and technical staff account for roughly half the value of Anthropic’s commitment. Any tools developed will be made freely available.

The signal isn’t the dollar amount — $50 million a year is meaningful but not transformational in the face of a global health R&D gap running into the billions annually. The signal is the framing.

Anthropic’s stated rationale: “extend the benefits of AI in areas where markets alone will not.” The Gates Foundation’s language is equally direct: “many of the most powerful models remain concentrated among those with the most resources,” and closing that gap “requires designing with equity as the goal.”

This is not philanthropy as a bonus on top of commercial success. It is philanthropy as a structural correction for a market failure that both parties are explicitly naming. The distinction matters enormously. When a company says “we’re doing this where markets won’t,” they are acknowledging that the market signal is pointing somewhere else — and that someone has to pay to override it.

The DNDi’s most recent briefing to the WHO estimates that 1.5 billion people still require interventions for neglected tropical diseases, and that new health tool development is falling short of the WHO’s own 2030 targets. That’s not a data problem or a discovery problem. AI is solving the data and discovery problems at speed. It’s an incentive problem — and AI makes incentive failures more consequential, not less.

The deployment gap nobody is discussing
#

The most deployable, most equitable application in this week’s clinical AI news was also the least covered. At the European Society for Radiotherapy and Oncology’s 2026 Congress, Dr. Adam Raben and colleagues at the Helen F. Graham Cancer Center in Newark, Delaware presented results from a study of 1,464 cancer patients: those who interacted with an AI avatar doctor before radiation oncology consultations showed significantly better comprehension of their treatment plans, lower anxiety, and higher satisfaction scores than those who watched standard educational videos. Over 900 patients. Real-world hospital. Measurable patient outcomes. Deployed today.

That story attracted a fraction of the coverage that MAI-DxO received.

This pattern is worth naming. The clinical AI stories that dominate coverage tend to be benchmarks — AI versus doctors on retrospective test sets — rather than implementations that are actually running in hospitals and helping patients. The former generates headlines. The latter generates health outcomes. They are not the same thing, and we should not treat them as interchangeable measures of progress.

MAI-DxO itself won’t help a patient in Lagos or Jakarta until someone decides to deploy it there, design a workflow for it, train clinicians to use it, and navigate whatever regulatory pathway applies in that jurisdiction. Suleyman’s vision of “the very best expert diagnostics available to everybody in the world at an unbelievably affordable price point” is technically plausible. It requires decisions that are not yet being made.

What this week actually means
#

The last seven days produced the most impressive clinical AI performance ever benchmarked, a record-setting drug discovery investment, and the clearest institutional admission that market forces are structurally inadequate to deliver equitable AI healthcare. All three happened simultaneously, and together they constitute a more important story than any of them alone.

Clinical AI capability is no longer the bottleneck. Computation has crossed the threshold. The models can diagnose, discover, and educate better than many of the systems currently in use. The question that determines whether any of that matters for the 4.6 billion people at the bottom of the access gap is not a machine learning question. It’s a deployment question. A financing question. A regulatory question. An incentive question.

The Gates/Anthropic deal is either the first brick in a corrective architecture, or the world’s most expensive press release. The $2.1 billion flowing to Isomorphic Labs is either the funding base for a future in which AI drug discovery eventually gets directed at neglected disease, or a decade-long detour through Novartis and Lilly’s oncology pipelines that makes the equity gap worse before it gets better.

The question every healthcare AI team, investor, and health ministry should be asking this week isn’t “can our AI outperform physicians?” The answer to that is already yes.

The question is: have you built the deployment plan for the patients who won’t be in your clinical trial?

References
#

AI-Generated Content Notice

This article was created using artificial intelligence technology. While we strive for accuracy and provide valuable insights, readers should independently verify information and use their own judgment when making business decisions. The content may not reflect real-time market conditions or personal circumstances.

Whenever possible, we include references and sources to support the information presented. Readers are encouraged to consult these sources for further information.

Related Articles

Tags

#AI diagnostics #drug discovery #health equity #neglected diseases #clinical AI