Large language models have become so prominent that they have taken over the word “AI” in everyday conversation, and their fluent, confident answers are a major reason. That confidence can be misleading, because a model predicts the next likely token, but fact-checking is unreliable. In healthcare and life sciences, where answers must be verified, explained, and attributed to someone accountable, that output needs independent scrutiny before anyone acts on it.
Generative AI produces fluent, confident answers, and that confidence can be misleading. A model predicts the next likely token, but fact-checking is unreliable, so in healthcare and life sciences, where answers must be verified, explained, and attributed to someone accountable, its output requires independent scrutiny before anyone relies on it.
Neurosymbolic AI combines a language model that generates answers with a layer of explicit, logical rules that checks them, so every conclusion can be traced and explained.
Where AI has delivered so far
The field itself has produced real results in healthcare for years, particularly in medical imaging and scientific research, and those successes did not depend on language models but on AI/ML algorithms that have been here for a long time. Language models add value to conversational tools and to speeding up literature review, serving as a starting point for human readers rather than a substitute for them.
In drug development, their influence has remained largely confined to the preclinical phase and Phase I, leaving the later phases, which account for the vast majority of costs, largely unaffected.
What “probabilistic” means in practice
A language model generates answers by predicting what is probabilistically most likely to come next. It can be fluent and useful, yet it has no built-in way to confirm that an answer is true. A language model can give different answers to the same question, because chance is built into how it chooses each word. Engineers can remove that variation, but a consistent model can still be wrong in the same way every time, while one that varies can be wrong differently each time. When an output can harm a patient, neither is acceptable.
The neurosymbolic approach
The neural part is the language model, which recognizes patterns in large volumes of data and generalizes across imperfect input. The symbolic part resembles classic programming, a set of explicit “if this, then that” rules that behave identically on every run and can show exactly how they reached a conclusion.
Each part compensates for the other’s limits. Rules written for every situation would grow into an unmanageable tangle of code, and a neural model on its own cannot give the certainty that critical decisions demand. Language models keep their place, with the symbolic layer checking their output before anyone relies on it. The design question for each use case is how much weight each part carries, and answering it well takes engineers who understand the clinical and regulatory setting as well as the code.
Where each approach belongs
The deciding factor is how much harm an error could do to a patient. Since that risk runs through most of healthcare and life sciences, the symbolic layer belongs in most of it. Language models, on their own, are well-suited to supervised, lower-risk settings such as chatbots, wellness tools, and hospital front desks. Verifying research hypotheses, selecting thousands of participants for a clinical trial, checking drug interactions, post-market monitoring, and any diagnostic or therapeutic support all call for explicit rules, defined on a case-by-case basis with close attention to the boundaries of each task.
The need for the symbolic layer follows from what the neural part does. Because it identifies statistical patterns, it can detect correlations between two factors but cannot establish causality. Another example can be found in drug development, where a drug candidate may change a biological marker as intended yet still fail to improve patient outcomes in late-stage trials, even after hundreds of millions of dollars have been invested. The symbolic layer checks a proposed mechanism against established biological knowledge early in the process, before those costs are incurred.
Regulated documentation follows the same logic. Patient information leaflets, for example, draw on source material compiled over a product’s lifecycle, and a language model can assemble that material well. A human reviewer remains responsible for the final result, and without an explanation of what the model did, that review means reading every statement from start to finish. A verification layer that checks the text against its sources and shows what it checked can shorten that review.
Trust and accountability
Medicine operates on liability. Physicians carry legal and professional responsibility for their decisions, so they can adopt only those recommendations whose reasoning is based on established clinical quidelines thus understandable for physicians. Diagnostic and therapeutic AI will therefore gain clinical acceptance to the extent that it can explain its output and that someone is accountable for it.
The sector still lacks a comprehensive strategy for managing AI-related risk in healthcare and life sciences. Efforts to build such frameworks are under way around the world, yet the technology evolves so quickly that putting them into practice is considerably harder than it was for earlier technologies, because a framework can start to lag behind the systems it governs before it is fully operational.
What this means for the industry
Capable language models are multiplying and becoming commodities, so the model itself is unlikely to be the differentiator. Because the verification layer has to be defined on a case-by-case basis, it requires engineers who understand the clinical and regulatory context as well as the code. Writing code has lost much of its scarcity value now that AI has removed the bottleneck of producing it, and engineers increasingly work one level higher, on architecture and on the specifics of the industry where a system is deployed. That is why HTEC pairs strong engineers with subject matter experts, and why the combination is well suited to this problem.
Generative AI will remain part of healthcare’s toolkit, and pairing it with deliberate verification is what allows it to be used where the stakes are highest. The question of who is responsible when an AI system is wrong is the subject of our companion piece, “Should AI Need a Medical License?”





