ANALYSIS
Between early 2023 and late 2025, the share of inpatient hospital cases billed as "medically complex" rose from 37 percent to 40 percent across Blue Cross Blue Shield networks. Three percentage points. It sounds like rounding error. It cost patients and insurers $942 million.
That figure comes from a Blue Cross Blue Shield Association (BCBSA) analysis published Wednesday, and it arrives with a detail that transforms it from a billing dispute into something more fundamental: the additional diagnoses driving those costs were not, in most documented cases, accompanied by additional treatment. Hospitals were billing for sicker patients. The patients were not getting sicker care.
This is what an AI arms race looks like when it lands inside a fee-for-service healthcare system. Hospitals deploy AI coding tools that scan medical records, doctors' notes, and lab results for secondary diagnoses — conditions that, when appended to a primary claim, push the billing into a higher-reimbursement category. Insurers respond by deploying their own AI to contest those claims. Neither system is designed to improve care. Both are optimized to win a revenue dispute. The patient is the terrain on which the battle is fought.
Luke Chalker, BCBSA's senior vice president of product and data science and a co-author of the analysis, was direct about what the data showed. "Critically, what we found is underneath all of that data [was] no change in corresponding care for a more complex patient," Chalker told reporters. "We find no evidence of a corresponding change in care."
The BCBSA examined specific secondary diagnoses as test cases. One example: anemia following major bowel surgery. Anemia diagnoses at hospitals using AI coding tools rose significantly. Transfusion rates — the actual clinical response to anemia — barely moved. "If patients are truly sicker, we'd expect to see more treatment," Chalker said. "The disconnect between diagnoses and treatment suggests that AI is identifying more billable conditions, not sicker patients."
This finding echoes an earlier BCBSA study focused specifically on maternity care, which documented a dramatic increase in diagnoses of acute posthemorrhagic anemia at certain hospitals, again without a corresponding rise in transfusions. That single diagnosis, the earlier study estimated, added $22 million to maternity admission costs in one year. The new analysis extends that pattern across the full inpatient landscape — and across a two-year window that saw AI coding tools move from novelty to standard hospital infrastructure.
AI-powered coding systems scan patient records, physician notes, and laboratory results to identify secondary diagnoses. When a secondary diagnosis is appended to a primary claim, it can shift the entire case into a higher-reimbursement "diagnosis-related group" — even if the secondary condition was incidental and received no treatment. Ambient AI systems can also transcribe clinician-patient conversations and flag additional billable conditions in real time. Both capabilities have legitimate administrative uses. In a fee-for-service payment system, both also create direct financial incentives to document more, regardless of whether more was treated.
Here is the original argument the BCBSA data makes possible, but does not quite make: this is not a story about bad actors. It is a story about a system that has automated its own perverse incentives. The hospitals deploying AI coding tools are, in most cases, doing exactly what the fee-for-service structure rewards them for doing — documenting every possible diagnosis, maximizing every possible reimbursement category. The AI doesn't corrupt that incentive. It accelerates it. What took a skilled human coder hours of chart review now happens automatically, at scale, across thousands of patient records simultaneously.
The result is a billing apparatus that has, in effect, decoupled diagnosis from care. A condition can now exist — on paper, in a billing system, in an insurer's actuarial tables — without the patient ever receiving treatment for it. This is not fraud in the conventional sense. It is optimization. The AI found the legal maximum. The system paid it.
This matters beyond the $942 million figure. As Senate Republicans have already voted to allow algorithms to deny Medicare care, the same AI infrastructure now inflating hospital billing is being positioned on the insurer side to automate claim rejections. Two algorithmic systems, each designed to extract maximum advantage from the same fee-for-service architecture, are now in direct opposition — and the cost of that opposition is borne by the people whose health is nominally being managed.
The BCBSA analysis is careful to note that AI can have legitimate benefits: reducing paperwork, cutting administrative burden, catching genuine coding errors. These are real. They are also, in the current incentive environment, beside the point. BCBSA's own framing acknowledges that the financial incentives built into fee-for-service healthcare can redirect those legitimate capabilities toward profit maximization. The technology is neutral. The system it operates inside is not.
Progressive healthcare advocates argue that no amount of AI governance reform addresses this root problem. Advocates for a Medicare for All-type system, including Sen. Bernie Sanders (I-Vt.) and Rep. Pramila Jayapal (D-Wash.), argue that replacing the fee-for-service model with a universal public program eliminates the billing category arbitrage that AI tools are currently exploiting. A global budget system, where hospitals receive fixed annual funding rather than per-diagnosis reimbursement, removes the financial incentive to append secondary diagnoses that receive no treatment. There is no billing category to game if billing categories don't determine revenue.
That argument is worth taking seriously on its own terms — not because it is politically convenient, but because the BCBSA data makes a specific structural case for it. The $653 million attributable to secondary diagnoses pushing claims into higher-paying categories is not a technology problem. Replacing the AI with human coders would not eliminate it; it would only slow it. The incentive to find those secondary diagnoses exists because the payment system rewards finding them. AI made that incentive faster and cheaper to act on. It did not create it.
The healthcare AI conversation in the United States has been dominated by two competing narratives: the techno-optimist promise of efficiency and the techno-skeptic fear of algorithmic bias. The BCBSA analysis suggests a third possibility that neither narrative anticipated — not that AI will make healthcare better or that it will discriminate more efficiently, but that it will make the existing system's worst features faster, cheaper, and harder to audit. The opacity already built into AI systems compounds this: when a diagnosis emerges from a machine-learning scan of a patient record rather than a physician's clinical judgment, the chain of accountability between care and documentation becomes harder to trace and harder to contest.
There is a specific consequence that the BCBSA figure doesn't capture: the $942 million in additional costs doesn't disappear. It moves. It becomes higher premiums, higher deductibles, higher out-of-pocket costs for the patients whose records were scanned to generate it. The people paying for the AI billing arms race are not the hospitals deploying it or the insurers contesting it. They are the patients whose conditions — documented but untreated — became line items in a revenue dispute they never knew was happening.
