AI in Health Research: Real Progress and Its Limits

The strongest cases for AI in health research are also the most specific. In 2020, researchers at MIT used a deep learning model to identify halicin, an antibiotic effective against a wide range of bacteria, including resistant strains. BenevolentAI flagged baricitinib, an existing rheumatoid arthritis drug, as a COVID-19 candidate within days, and it was later authorized for emergency use. AI-assisted mammography is improving early detection. Those advances help real patients, and they share one structural feature: each result can be checked quickly, cheaply and objectively. That is why the progress is genuine — and why it stops where fast verification stops.

The advances that hold up

A deep learning model is software that learns patterns from large datasets instead of following hand-written rules. Pointed at AI drug discovery, it compresses a search space no human team could read at all. IBM Watson's natural language processing, which extracts meaning from written text, sifted millions of papers and trial reports and surfaced six promising oncology molecules in weeks — a task the same source estimates would have taken human researchers months or years. Atomwise screened existing medicines against Ebola in a single day in 2016, and BenevolentAI's call on baricitinib came within days (Allied Academies).

The common thread is leverage, not magic. Repurposing an existing drug starts with safety data already on file. Screening known compounds starts with a bounded list. When a suggestion can be tested in a day, being wrong is cheap, so the process can iterate many times. Iteration — not one brilliant prediction — is where the speed comes from.

Diagnostics: the help that reaches patients

AI medical imaging is the most mature category, and the reason is structural: the target is comparatively stable and the model's role can be bounded. StateViewer reads FDG-PET scans — imaging that maps the brain's glucose consumption — compares each case against a large library of prior studies, proposes the most compatible neurodegenerative pattern, and shows which brain regions drove the answer (UNIA). AI-assisted mammography has improved early detection and radiologists' precision (APD).

The gains are often distributive rather than scientific. In remote parts of Turkmenistan, ultraportable X-ray units running CAD4TB software, with support from the Global Fund, are widening access to specialist-level tuberculosis screening where no specialist sits (UNDP). Chile's Centro de Modelamiento Matemático uses algorithms to classify patients and prioritize care, and the country's first National AI Policy, published in 2021, underpins the National AI Center (CENIA), the first AI doctorate in Chile and Latin America, and an Ethical Algorithms project (NOA).

Deployment plus guardrails is what makes that help durable instead of a demonstration.

The pattern behind the progress: fast verification

Every strong case shares one feature: a cheap, fast, objective way to check whether the machine got it right — a binding assay, a protein structure, a labeled scan, a docking score. Fast loops let models iterate, and iteration is where the advantage lives.

Where verification is slow, contested or impossible to automate — long-term clinical outcomes, causal claims about whole populations, the economics of running trials — the advantage shrinks sharply. The medical literature is sober about this: most subfields of AI in medicine are still in a research phase, with many trials under way simply trying to establish utility. The same literature frames AI as an assistant that processes clinical data and returns a concrete aid, dependent on the use case, the type of data — image, lab result, signal, clinical text — and the clinical context, not as a substitute for professional judgment (PMC).

"AI accelerates science" is too coarse a claim. It accelerates the part of science that can be scored quickly.

The counterpoint: what the progress does not prove

The counterpoint is not that the advances are fake. It is that the figures quoted for AI in health research measure early identification, not medicines that reached patients. Between a promising molecule and a prescription stand data quality, regulatory approval and ethical considerations, and none of those improves because a model improves (Allied Academies). Screening a compound is not evidence that it will survive trials. Faster candidate generation compresses the cheapest stage of the pipeline; if validation capacity does not expand alongside it, the result is congestion downstream rather than relief.

There is also a disclosure problem. Corporate case studies and partnership announcements rarely end at a patient outcome, and they are published by the party that benefits when the result is believed. Many flagship examples are retrospective by construction — repurposing an existing drug, rescreening known compounds — which demonstrates pattern-finding on data that already exists, not prospective predictive power on data that does not.

The strongest objection, and the response

The fair objection is that this framework understates a real shift. The sample is small, but halicin and baricitinib were considered practically impossible before, which suggests the sample understates the trend rather than overstating it. Slower adoption may reflect regulatory caution and fragmented data, not the limits of the technology. Holding AI to a stricter evidentiary bar than conventional research tools looks inconsistent, since those tools are imperfect too. And dismissing the results as hypothesis generation ignores deployments in Chile and Turkmenistan that already change patient routing and access.

The response is not to deny any of it, but to sort the evidence by what it can bear. Halicin and baricitinib are real. They relocate: from proof that AI is changing medicine to proof that AI is changing the front end of research. Both statements can be true at once; only one is supported today. The deployments that change routing are exactly the ones bounded as support for a professional — the modest framing the evidence actually licenses.

Who gains, who waits

Models learn from data that exists, and data exists where research money is. Concentration is therefore the structural risk: institutions holding large datasets, compute providers and large pharmaceutical partners capture the early value, while research systems without data infrastructure or supercomputing access wait — and health systems that inherit tools without the staff to validate them carry the risk without the benefit. Public policy is one of the few visible counterweights, and Chile's strategy, research center, doctoral pipeline and ethical-algorithms work were built on purpose (NOA).

What the progress is worth

The evidence supports a narrower and more useful conclusion than the promotional version. AI in health research is delivering documented advances in drug repurposing, molecule screening, medical imaging and access to screening, and those advances help because they land where a result can be verified fast. The same rule explains the ceiling: nothing about a better model shortens a long trial or repairs fragmented data.

Three takeaways follow:

1. Fund what works. Repurposing, screening and bounded diagnostic support return value today; treat them as infrastructure, not as a bet on a future breakthrough.
2. Read every new claim for its verification loop and for where it ends — at a candidate, a screen or a structure, or at a patient outcome. The first three are the easy part.
3. Pair models with validation capacity, data infrastructure and staff who can check the output. Otherwise faster discovery becomes a longer queue.

The technology is doing real work at the front end of research, where verification is cheap. Everything beyond that is a hypothesis, and it should be labeled as one.