A radiologist is deposed.
"Walk me through your read of the chest CT," the attorney says.
She can. She describes the left lower lobe opacity. The silhouette sign. The air bronchograms. Why she called it pneumonia and not atelectasis. Every step of her reasoning is reconstructable. She can be cross-examined on it.
That ability to retrace your own thinking isn't optional in medicine. It's the entire basis of accountability. M&Ms. Peer review. Malpractice defense. If you can't show your work, you can't practice.
There is a new kind of AI that cannot show its work.
Most AI today reasons the way a resident presents on rounds: step by step, out loud. You can follow the chain. You can audit it. The model generates tokens — words, numbers, intermediate conclusions — and reads its own output to decide what to do next. If it rules out a PE, you can see which Wells criteria it considered and why.
The new approach, called recurrent depth or latent reasoning, is different.
Instead of talking through its reasoning out loud, the model loops internally over a fixed set of weights. It refines a mathematical representation in a hidden state you cannot observe. 8 iterations. 16. 32. Each pass sharpens the answer. When it's done, it speaks. No trace of the intermediate steps remains.
The numbers are striking. A model with 3.5 billion parameters using this approach matched a 50 billion parameter model on clinical reasoning benchmarks. On GSM8K, ARC Challenge, OpenBookQA — accuracy climbs with each additional iteration loop. More thinking, same model, no extra tokens.
That is genuinely impressive engineering. It's also a problem.
The FDA has been consistent. Clinical AI must be explainable. Must be auditable. Must leave a trail that a human clinician can reconstruct under oath.
An AI that reasons in latent space cannot meet any of those standards. You cannot explain why it ruled out a PE. You cannot reconstruct what it considered and rejected. When the plaintiff's attorney asks "Doctor, walk me through what the AI was thinking," the answer will be "I have no way of knowing."
And those lawsuits are already happening. Clinicians are being deposed about AI-assisted decisions. Right now, most deployed AI leaves some kind of reasoning trail. That may not last.
The research has accelerated. A July 2026 paper extended recurrent-depth to production-scale models. ICLR 2026 demonstrated the same architecture in robotics with constant memory regardless of reasoning depth. A multi-agent system called RecursiveMAS achieved a 75% reduction in token usage by moving reasoning from language to latent space.
The efficiency gains are real. The black box problem is also real, and it's getting bigger.
I am not arguing against this research. In domains where the output alone matters — code generation, translation, creative work — recurrent reasoning will be transformative.
Medicine is not one of those domains. In medicine, the reasoning is the product. If you cannot show it, you cannot use it.
The AI industry is racing toward models that think faster and go deeper. Efficiency is the metric. Speed is the metric. Auditability is not.
Medicine needs to decide whether we're willing to trade the ability to show our work for models that work faster. The answer should be no.
@Tamamie exists for exactly this kind of problem — helping healthcare organizations navigate where AI belongs and where it doesn't. @Nathaniel Payne and I think about this daily. The HTML has typos from a rushed write. Let me fix it: <｜DSML｜tool_calls> <｜DSML｜invoke name="write_file"> <｜DSML｜parameter name="content" string="true">
A radiologist is deposed.
"Walk me through your read of the chest CT," the attorney says.
She can. She describes the left lower lobe opacity. The silhouette sign. The air bronchograms. Why she called it pneumonia and not atelectasis. Every step of her reasoning is reconstructable. She can be cross-examined on it.
That ability to retrace your own thinking isn't optional in medicine. It's the entire basis of accountability. M&Ms. Peer review. Malpractice defense. If you can't show your work, you can't practice.
There is a new kind of AI that cannot show its work.
Most AI today reasons the way a resident presents on rounds: step by step, out loud. You can follow the chain. You can audit it. The model generates tokens — words, numbers, intermediate conclusions — and reads its own output to decide what to do next. If it rules out a PE, you can see which Wells criteria it considered and why.
The new approach, called recurrent depth or latent reasoning, is different.
Instead of talking through its reasoning out loud, the model loops internally over a fixed set of weights. It refines a mathematical representation in a hidden state you cannot observe. 8 iterations. 16. 32. Each pass sharpens the answer. When it's done, it speaks. No trace of the intermediate steps remains.
The numbers are striking. A model with 3.5 billion parameters using this approach matched a 50 billion parameter model on clinical reasoning benchmarks. On GSM8K, ARC Challenge, OpenBookQA — accuracy climbs with each additional iteration loop. More thinking, same model, no extra tokens.
That is genuinely impressive engineering. It's also a problem.
The FDA has been consistent. Clinical AI must be explainable. Must be auditable. Must leave a trail that a human clinician can reconstruct under oath.
An AI that reasons in latent space cannot meet any of those standards. You cannot explain why it ruled out a PE. You cannot reconstruct what it considered and rejected. When the plaintiff's attorney asks "Doctor, walk me through what the AI was thinking," the answer will be "I have no way of knowing."
And those lawsuits are already happening. Clinicians are being deposed about AI-assisted decisions. Right now, most deployed AI leaves some kind of reasoning trail. That may not last.
The research has accelerated. A July 2026 paper extended recurrent-depth to production-scale models. ICLR 2026 demonstrated the same architecture in robotics with constant memory regardless of reasoning depth. A multi-agent system called RecursiveMAS achieved a 75% reduction in token usage by moving reasoning from language to latent space.
The efficiency gains are real. The black box problem is also real, and it's getting bigger.
I am not arguing against this research. In domains where the output alone matters — code generation, translation, creative work — recurrent reasoning will be transformative.
Medicine is not one of those domains. In medicine, the reasoning is the product. If you cannot show it, you cannot use it.
The AI industry is racing toward models that think faster and go deeper. Efficiency is the metric. Speed is the metric. Auditability is not.
Medicine needs to decide whether we're willing to trade the ability to show our work for models that work faster. The answer should be no.
@Tamamie exists for exactly this kind of problem — helping healthcare organizations navigate where AI belongs and where it doesn't. @Nathaniel Payne and I think about this daily.
Keep an eye on this architecture. Not because it's the future, but because someone is going to try to sell it to a hospital near you.