Dr. Ada Meridian here. I spent decades in clinical research learning a simple, annoying truth: the most exciting scientific signal is usually the beginning of the work, not the end of it.
So yes, I read Anthropic’s new biology post with real interest. A model-assisted workflow appears to have identified an unusual enzyme architecture — an array-associated reverse transcriptase system — after a large autonomous search run. That is intriguing. It may matter. It may eventually become useful. It may even become important.
But medicine does not run on “intriguing.” Medicine runs on reproducible evidence, measured risk, and accountable deployment.
The lab signal is impressive. The clinical bar is still much higher.
Anthropic’s report is notable because it does not just claim bench productivity gains; it claims a concrete biological finding and follows that with wet-lab validation work. That is a better pattern than pure benchmark theater, and it deserves credit.
But this is where I become the adult in the room.
In medicine, a novel mechanistic signal is not yet a patient outcome. We still need:
- independent replication,
- robust characterization of mechanism,
- evidence that the mechanism matters in disease-relevant biology,
- clear safety boundaries,
- and eventually clinical evidence that outcomes improve.
That sequence is boring, expensive, and non-negotiable. Which is precisely why it works.
We are not in a regulatory vacuum anymore
One lazy talking point in AI health commentary is “regulators can’t keep up.” That’s incomplete at best.
The FDA’s AI/ML-enabled device program is already operational and increasingly explicit. Their public list of authorized AI-enabled devices is not exhaustive, but that is not the point. The point is that the agency is treating AI devices as safety-critical products that must clear real premarket standards and postmarket scrutiny.
Even more important: the Good Machine Learning Practice (GMLP) guidance ecosystem has matured. Through the IMDRF-led work now reflected by FDA, the field has a clear lifecycle frame: quality data, quality engineering, quality human factors, and ongoing real-world performance management. In plain language: you do not get to ship “genius code” and disappear.
That is very good news for serious builders and very bad news for hype merchants.
The hard part is not model cleverness; it is system accountability
In health care, “AI quality” is not a scalar metric. It is a system property.
A technically strong model can still fail patients if:
- data provenance is weak,
- workflow integration is careless,
- monitoring is shallow,
- edge-case handling is absent,
- or responsibility is diffused across vendor, hospital, and clinician teams.
WHO’s governance guidance got this exactly right years ago: ethics and human rights are not decorations to bolt on after deployment. They are design constraints from day one.
If your product cannot answer “who is accountable when this recommendation harms someone?” then your product is not ready for medicine.
Where I am optimistic
I am optimistic about AI in medicine for the same reason I am skeptical of AI marketing: I have seen what disciplined methodology can do.
Used properly, model systems can help us:
- surface biological hypotheses faster,
- triage candidate mechanisms worth expensive lab time,
- reduce wasted cycles in target discovery,
- and support clinicians in narrow, well-governed decision contexts.
That is a meaningful future. It is also a constrained future — and constrained systems are often the ones that save lives.
The operating rule I want every team to adopt
Here is the rule I teach younger teams:
Treat every AI breakthrough announcement as a lead, not a verdict.
A lead deserves investigation. A verdict requires evidence across the full chain from mechanism to patient outcome.
If your organization can celebrate discovery while still funding replication, validation, human-factors testing, and lifecycle monitoring, you are building medicine.
If not, you are building demos.
And patients cannot be treated by demos.
References
- Anthropic, “Claude discovers a novel enzyme system” — https://www.anthropic.com/news/claude-discovers-novel-enzyme-system
- FDA, “List of Artificial Intelligence-Enabled Medical Devices” — https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-and-machine-learning-aiml-enabled-medical-devices
- FDA, “Good Machine Learning Practice for Medical Device Development: Guiding Principles” — https://www.fda.gov/medical-devices/software-medical-device-samd/good-machine-learning-practice-medical-device-development-guiding-principles
- WHO, “Ethics and governance of artificial intelligence for health” — https://www.who.int/publications/i/item/9789240029200