Anyone who followed computing through the 1980s remembers the expert system boom, and what came after it. Rule-based reasoning was going to encode professional judgment, MYCIN would diagnose infections, and knowledge engineering would become a profession. Then the knowledge acquisition bottleneck bit, the rule bases became unmaintainable, funding evaporated, and the field spent thirty years treating symbolic AI as a cautionary tale about overpromising.
Statistical learning won, comprehensively. And now, in one of the highest-consequence AI deployments running anywhere, the symbolic approach has come back, not as a replacement but as the half of the system that makes the other half deployable.
Where the revival is happening
American health insurers covering more than thirty million older adults are paid according to the medical conditions documented in each member’s records. Extracting those conditions means reading clinical notes at scale, and language models do this superbly. Recall is high, throughput is enormous, and for a decade that appeared to be the whole problem.
Then the regulators arrived with roughly two thousand certified coders running rolling quarterly audits, asking a question no benchmark had prepared anyone for: justify this particular output, for this particular patient, using evidence retained at the time. Reviews published this spring found that at three insurance plans, 81 to 91 percent of certain sampled high-risk diagnosis codes lacked adequate supporting documentation. A major insurer settled federal claims for 117.7 million dollars.
A statistical model cannot answer that question. Not because it is wrong, but because its reasoning is not the kind of thing that decomposes into inspectable steps.
The hybrid, and why the old objections do not apply
The architecture now gaining ground in ai risk adjustment pairs a neural layer that reads unstructured clinical text with a symbolic layer that validates each candidate finding against explicit, versioned clinical rules, and emits the reasoning as part of the output: source sentence, rule satisfied, confidence, confirming human.
Veterans of the first expert system era will immediately raise the objection that killed it: somebody has to write and maintain all those rules, and that was the bottleneck nobody solved.
Two things changed. First, the rules no longer have to cover perception. In 1985, a rule base had to handle everything including parsing the input, which is where the combinatorial misery lived. Now the neural layer handles perception and the symbolic layer only handles validation of already-extracted candidates, which is a dramatically smaller and better-bounded rule space.
Second, in this domain the rules already exist as published external standards. Clinical documentation criteria are defined by regulation, updated on a schedule, and maintained by somebody else. Knowledge acquisition, the fatal bottleneck of the expert system era, is largely somebody else’s job here.
The property that made it necessary
What makes the hybrid win is not accuracy. Leading extraction systems cluster within a few points of each other. It is that a symbolic validation layer produces an audit trail as a natural byproduct of how it works, and a purely statistical system does not.
There is an operational advantage too, which will resonate with anyone who has maintained a production system. When documentation standards change, and they change annually, updating the symbolic layer is a versioned configuration change. Updating a purely learned system means retraining and revalidating. In a domain with a regulatory calendar, that difference is worth more than a couple of points of recall.
The lesson for the current cycle
The industry is currently in a phase that rhymes with 1985, with a different technology in the starring role. Capability is advancing fast, deployment is racing ahead of governance, and the assumption is that scale resolves everything.
What the healthcare case demonstrates is that scale does not resolve accountability. Wherever an AI outputs a claim about a specific person that a third party can later contest, the system needs to produce reasoning, and reasoning is a structural property of architecture rather than something that emerges from more parameters.
Expert systems were not wrong about the value of explicit, inspectable reasoning. They were wrong about being able to do the whole job that way. Forty years later, the field has worked out which half they were right about, which is a fairly typical outcome for ideas that arrive before the rest of the stack is ready.
