Last week the Blue Cross Blue Shield Association published an analysis of its own claims data and arrived at a number built to travel: $942 million. That is what the association says its member plans paid between 2023 and 2025 because hospitals coded a larger share of inpatients as medically complex, and it attributes the shift to artificial intelligence tools, now in use at more than 60 percent of hospital systems, that scan laboratory results and physician notes for secondary diagnoses.1 The number traveled as intended. Within a day it had become a morality tale about greedy hospitals and greedier doctors, and the word "upcoding" was doing the work that "fraud" would do if anyone were prepared to defend it in court. I want to take the analysis seriously, because it deserves that, and then explain why the outrage is aimed at the wrong target. Hospitals aren't gaming the system. They're obeying it, and on the merits of the particular dispute they are in the right. The insurers who object to them are obeying the same system from the other side. The system itself is the offense, and it was written by neither party.
The system is Medicare's Inpatient Prospective Payment System, and almost none of the commentary I read last week mentioned it. Before 1983, Medicare paid hospitals their reported costs, and costs rose accordingly. In 1983 Congress amended the Social Security Act to pay instead by a classification scheme developed at Yale during the 1970s, the diagnosis related group, whose original purpose was monitoring quality and utilization rather than setting prices.2 Under the prospective system a hospital isn't paid for what it does to a patient. It is paid a lump sum for the admission, calculated by multiplying a base rate by a weight assigned to the patient's group, and the weight reflects how costly the government expects that kind of patient to be relative to all others.3 The government's expectation, in other words, is the price. As of 2024 the scheme contains 766 Medicare Severity DRGs, and most of the underlying 340 base groups are split into two or three payment tiers depending on whether the record contains a secondary diagnosis classified as a complication or comorbidity, or a major one.4 The tiers are the whole story. The same operation carries three prices, and what separates them is whether a second diagnosis appears in the chart.
That structure is itself a relatively recent government decision. Before fiscal year 2008 there were 538 groups. That year the Centers for Medicare and Medicaid Services revised the use of complicating conditions so extensively that it renamed the system, and the count rose to 745.5 The government tripled the number of doors a secondary diagnosis could open and then expressed surprise when hospitals began looking for keys. The surprise was theatrical. CMS predicted the effect in the rule that created it, and in 2012 Congress ordered the agency to recoup $11 billion in what it called documentation and coding overpayments from the first three years of the new system, which CMS did through across-the-board rate reductions beginning at 0.8 percent a year in fiscal 2014.6 Notice what that episode establishes. The federal government has already adjudicated, in its own rulemaking, the question of whether more thorough documentation raises payment without raising illness. It concluded that it does, and it took the money back from every hospital in the country, including the ones that had changed nothing. That is what the rules are: a lever for hospitals to pull and a ratchet for the government to pull back, on a schedule measured in years.
Now the association's evidence, at full strength. The share of Blue plan inpatient claims coded as medically complex rose from 37 percent at the start of 2023 to 40 percent at the end of 2025. Within major bowel procedures, claims at the highest complexity level rose from 10.2 percent to 22.7 percent. Roughly 70 percent of the added cost, $653 million, came from secondary diagnoses, at about $11,000 per additional complex case.7 The test the association applied is a fair one on its face. Its head of data science put it plainly: if patients are truly sicker, you'd expect more treatment, and instead the analysis found significantly more anemia diagnoses without a corresponding increase in transfusions.8 Earlier work by the same group found the same pattern in postpartum anemia, where diagnoses rose sharply at the fastest-growing hospitals while transfusion rates barely moved.9 Any honest physician has to concede that diagnoses climbing while treatment stays flat looks suspicious, and that the pattern deserves an explanation.
The explanation is medical, and it is the point at which the association's case comes apart. The 2023 international guidelines from the AABB, the body that governs transfusion practice, make a strong recommendation that hemodynamically stable hospitalized adults be transfused only when hemoglobin falls below 7 grams per deciliter.10 Anemia from blood loss after a colon resection is common and real, and in the great majority of cases it is managed with observation, iron, and time, precisely because the evidence says that transfusing above that threshold harms patients. Medicine has spent fifteen years teaching surgeons to put the blood bank order down. A flat transfusion rate is therefore strong evidence that hospitals practice modern medicine and no evidence at all that the anemia wasn't there. A hemoglobin that falls after an operation is anemia. The software didn't invent it. It found it in the same laboratory record the surgeon was already checking every morning.
Nor can the software bill for it on its own, which is the part of the story that the word "upcoding" is meant to obscure. Under the federal coding rules a diagnosis can only be assigned on the strength of the treating physician's statement that the condition exists; a coder may not infer it from clinical criteria, and an abnormal laboratory value may not be coded unless the physician indicates that it is clinically significant.11 A secondary diagnosis becomes reportable when it required clinical evaluation, treatment, diagnostic procedures, a longer stay, or increased monitoring.12 Serial hemoglobins are monitoring. What the tool actually does, then, is read the record, notice that the surgeon has been following a falling hemoglobin without naming the condition, and ask him whether the patient has acute blood loss anemia. He says yes, because she does, and signs his name to it. Fraud requires a false statement. No one at the association has alleged that the laboratory values were fabricated or the diagnoses untrue. The complaint, stated precisely, is that hospitals have become better at documenting true things, and the association concedes as much in the fine print, acknowledging that its analysis relied on claims rather than clinical documentation, and allowing through its own vice president that there may be an element of correct coding in the numbers.13
Consider what the association is really describing. For years those patients were exactly this sick, and the hospitals were paid as though they were not, because the surgeon's note said "colectomy, doing well" and left the anemia in the lab column where the coder couldn't touch it. Every comorbidity that went undocumented was a discount, and the insurer collected it without ever having negotiated for it. What the association now calls $942 million in added cost is, from the other side of the table, the end of a subsidy that ran silently in its favor for two decades. The patients didn't get sicker in 2023. The paper finally caught up with them.
And the patients are, in fact, sicker, which makes the association's whole premise, that the trend must be artificial because it is large, exactly backward. The share of American adults with two or more chronic conditions rose from 47.3 percent in 2013 to 51.4 percent in 2023, and it rose fastest among the young.14 The share taking five or more prescription drugs nearly doubled between the turn of the century and 2012, from 8.2 percent to 15 percent.15 Severe obesity climbed from 7.7 percent of adults a decade ago to 9.7 percent.16 The first edition of Harrison's Principles of Internal Medicine, in 1950, ran to 1,590 pages; the twenty-first edition runs to 4,384, in two volumes.1718 Medicine is harder than it was, the patients carry more disease into the building than they used to, and over the same period the government's payment for treating them has fallen. Adjusted for the cost of running a practice, Medicare physician payment has declined 33 percent since 2001.19 Hospitals lose about 13 cents on every Medicare dollar, and 206 rural hospitals have closed or given up inpatient care since 2010, with 41 percent of the survivors operating at a loss.2021 The seven largest publicly traded insurers, meanwhile, reported a combined net income of roughly $34 billion for 2024.22 Whatever else is true of the hospitals that bought coding software, they weren't the ones with money to spare.
There is a deeper reason the paper lagged the patient for so long, and it is a reason I know from the inside. When a physician writes a note, he writes it for the next physician. My problem list on a critically ill patient names what I am actively fighting: the shock, the failing lungs, the kidneys shutting down. The dozen chronic conditions in the background, the ones that make every one of those fights harder to win, get a line if they get anything, because the note's purpose is to focus the attention of the team on what must be done in the next twelve hours. The skill a physician spends a career acquiring is precisely the skill of leaving things out. A medical student presents two hundred data points; a seasoned attending distills the same patient into two sentences that tell you where he is in his disease and what will decide his course. That compression is the most useful thing one doctor can hand another, and it is the opposite of what the prospective payment system rewards. The grouper doesn't want the synthesis. It wants the list, every comorbidity spelled out in its own reimbursable line, and the longer and more disjointed the list, the more impressed the formula is. To a physician, a note that reads like an inventory is a sign that the writer doesn't understand the patient. To the payment system it is the only kind of note that counts.
There is a matter of vantage as well. Every physician now in practice trained into the disease burden of his own era and took it as the baseline, the way a fish takes the water; nobody at the bedside today was there in 1960 to notice the difference. The profession's awareness that the man in bed four carries a load of chronic illness his grandfather never would have carried at the same age is dim and statistical, acquired from journals rather than from the work. And even a physician who held the fact clearly in mind wouldn't write it down, because a clinical note isn't a public health document. The rising burden of disease is a truth about populations. The note is about one man, on one morning, and what must be done for him before noon. Nobody records, twenty-five times a shift for thirty years, boy, these patients are sicker than they used to be, even when the patient in front of him is a perfect specimen of the trend. The comorbidities sit in the record for anyone who cares to look. They simply aren't the point of the encounter, and the note is a faithful record of what the point was.
That is what clinical documentation specialists have been hired to correct since the day prospective payment began, and it is what the new software does, with more patience and better recall than any human coder. What ought to impress the public about the arrangement is how much friction it took to get there. For forty years the incentive has pointed one way and physicians have kept writing the other way, for each other rather than for the formula, to the point that hospitals must employ entire departments and now license language models to extract from the record what the doctor knew all along and declined to itemize. The association reads that as evidence of a scheme. I read it as a testament that the people closest to the patient have never once adjusted their thinking to the government's price list, to the point that the institutions around them have had to build expensive machinery to translate.
Artificial intelligence didn't invent any of this, either. Donald Simborg coined the phrase "DRG creep" in the New England Journal of Medicine in 1981, two years before Medicare mailed its first prospective payment.23 In 2021 the Inspector General of Health and Human Services reported that Medicare stays billed at the highest severity level rose almost 20 percent between fiscal 2014 and 2019, that 54 percent of them reached that level on the strength of a single qualifying diagnosis, that their average length of stay fell from 6.9 to 6.4 days, and that 5 percent of hospitals billed between 52 and 79 percent of their stays at the top tier.24 All of this predates the large language model. What software adds is speed and completeness. A tool that reads every note and every lab value doesn't invent the incentive. It removes the friction that had kept the incentive from being fully satisfied, which is what every tool does to every incentive it touches.
Now the other side of the ledger, because it's the same game played in the other direction. Insurers have machines too, and they had them first. In 2023 ProPublica reported that Cigna's medical directors had rejected more than 300,000 payment requests over two months using an automated system that matched procedures to diagnoses, spending an average of 1.2 seconds on each.25 In 2024 the Senate Permanent Subcommittee on Investigations found that UnitedHealthcare's denial rate for post-acute care in Medicare Advantage rose from 8.7 percent to 22.7 percent between 2019 and 2022, that the company used an algorithmic tool called nH Predict to manage those denials, and that CVS had deployed a program called Post-Acute Analytics whose projected savings escalated from roughly $10 million to $77 million.26 Commercial payers now run what the industry calls clinical validation programs, frequently through outside vendors, which strip secondary diagnoses from a claim after the fact and reassign the group downward, with sepsis, acute kidney injury, malnutrition, and encephalopathy as the favored targets.27 Reuters this spring described the result as AI fighting AI, and the association's own vice president conceded that when these tools operate independently they generate friction.28 Hospital executives have said their adoption of coding software was in part an answer to payers' willingness to deny or stall payment.29 Each side is arming against the other, and each side's armament is the other's justification.
Here the moral framing collapses entirely, because when the rules pay insurers for diagnoses rather than for disputing them, insurers hunt diagnoses with precisely the enthusiasm they now condemn. Medicare Advantage pays plans more for sicker enrollees, as measured by the diagnoses the plan reports. The Medicare Payment Advisory Commission estimates that coding intensity alone makes Medicare Advantage risk scores about 10 percent higher than those of comparable patients in traditional Medicare, and that between coding and favorable selection the program will pay plans roughly 20 percent more than it otherwise would, about $84 billion in 2025.30 In 2020 the Justice Department sued Anthem, the largest Blue Cross licensee in the country, alleging that it paid a vendor to comb medical records for diagnosis codes to add, made no effort to delete codes the same review failed to support, and regarded the program, which generated more than $100 million a year, as a cash cow.31 That is the corporate family whose trade association now warns the public about technology-enabled upcoding. And UnitedHealth Group, whose insurance arm has fought coding intensity as hard as anyone, owns Optum, which sells hospitals a coding and documentation platform advertised to improve case mix index through more complete and accurate code capture.32 One conglomerate supplies both sides of the trench.33 I don't raise this to call insurers hypocrites, though the word fits. I raise it because it settles the question of character. The same organizations behave one way when the rulebook rewards finding diagnoses and the opposite way when it rewards disputing them. Conduct follows the rules. The rules are written in Baltimore.
The obvious objection is that Blue Cross plans are private companies, that Medicare's grouper doesn't bind them, and that if they chose to pay hospitals by DRG the consequences are their own. The objection is correct, and it is the most revealing fact in the dispute. Commercial insurers adopted the federal classification because the entire documentation, coding, and billing apparatus of American hospitals was built to the federal specification, and because a payer covering one in three Americans cannot practically negotiate a different unit of payment with several thousand hospitals. The government didn't merely set Medicare's prices. It set the grammar in which every price in the sector is expressed. When that grammar consists of 766 groups, a diagnosis vocabulary that grew by 487 codes in last year's annual update alone,34 and a tier structure that turns a single hemoglobin value into an $11,000 swing, private parties don't escape the incentive by signing a private contract. They inherit it, along with the arms race it guarantees.
A century ago the sociologist Franz Oppenheimer distinguished two ways of acquiring wealth. One's own labor, and the equivalent exchange of it for the labor of others, he called the economic means. The unrequited appropriation of the labor of others he called the political means.35 He had taxation and conquest in mind, but there is a subtler variant that describes American hospital finance exactly. When the state sets the terms of exchange in sufficient detail, the participants in the market stop competing to serve the customer and start competing to satisfy the state. Under prospective payment neither the hospital nor the insurer earns a dollar by the patient's judgment of the care he received. The hospital earns by documenting to the grouper. The insurer earns by disputing to the grouper. Each lobbies, every year, over the contents of the next annual rule, and then each competes to play that rule better than the other. The patient isn't the customer of either party. He is the occasion for a transaction between them, and his hemoglobin is a token in it.
I'll grant the association one thing. Hospitals answer the insurers with their own aggregate data, a case mix index up about 5 percent from 2019 to 2024, and insist the patients are simply sicker,36 and while I've argued that they are, no one can prove it from a claim, which is the association's own admission about its own analysis. That inability is the real indictment. A payment system in which the payer, the payee, and the regulator cannot distinguish a sicker patient from a better-documented one, after forty years and billions of dollars in audits, has failed at the one thing a price exists to do, which is to carry information. And that a dispute over a federal payment formula became, within hours, a public referendum on the greed of physicians says something about how completely the profession has lost the argument for itself, a subject that deserves its own essay.
Consider what a market price would look like here. It would be agreed before the service, between the party providing it and the party paying for it, with the patient or an agent he chose as the counterparty. Bowel resections and deliveries, the two procedures the association studied, are overwhelmingly scheduled, so the standard objection that no one can shop in an emergency doesn't reach them. It does reach some of what I do for a living, and I'll concede it, while noting that emergencies are a minority of admissions and that even for them a hospital and an insurer contracting freely would arrive at terms that owe nothing to a federal severity table. When a price is a promise made in advance rather than a classification imposed afterward, there is nothing to upcode and nothing to downcode. The software on both sides would have to find other work. Some of it might even find its way to the bedside.
The association's $942 million is the sound of the game being played faster, and it will get faster still, because both sides are now building machines that read the rulebook better than any human ever did. Every dollar spent on that contest is a dollar spent on compliance rather than care, and the escalation is financed entirely by premiums and taxes. The terminal state of a system built on the political means is two armies of algorithms arguing over a form while the man whose name is at the top of the form is never consulted. If that offends you, and it should, direct the offense at the people who wrote the form. Don't hate the player. Hate the game, and then ask who made it.
Blue Cross Blue Shield Association. Analysis examines AI in hospital billing as spike in complex patients adds nearly $1 billion in extra costs. September 24, 2026. https://www.bcbs.com/about-us/association-news/bcbsa-analysis-ai-coding-tools-affects-healthcare-costs
Centers for Medicare & Medicaid Services. Design and development of the Diagnosis Related Groups (DRGs). October 2019. https://www.cms.gov/icd10m/version37-fullcode-cms/fullcode_cms/Design_and_development_of_the_Diagnosis_Related_Group_(DRGs).pdf
Medicare Payment Advisory Commission. Hospital acute inpatient services payment system. Payment Basics. https://www.medpac.gov/wp-content/uploads/2022/10/MedPAC_Payment_Basics_23_hospital_FINAL_SEC.pdf
Medicare Payment Advisory Commission. Hospital acute inpatient services payment system. Payment Basics. https://www.medpac.gov/wp-content/uploads/2022/10/MedPAC_Payment_Basics_23_hospital_FINAL_SEC.pdf
Centers for Medicare & Medicaid Services. Design and development of the Diagnosis Related Groups (DRGs). October 2019. https://www.cms.gov/icd10m/version37-fullcode-cms/fullcode_cms/Design_and_development_of_the_Diagnosis_Related_Group_(DRGs).pdf
Centers for Medicare & Medicaid Services, Office of the Actuary. Estimated financial effects of the documentation and coding adjustments required by section 631 of the American Taxpayer Relief Act of 2012. https://www.cms.gov/medicare/medicare-fee-for-service-payment/acuteinpatientpps/downloads/fy2017-nprm-actuary-estimate-of-medicare-documentation.pdf
Muoio D. Hospitals' use of AI coding tools cost BCBSA plans $942M more for similar care: analysis. Fierce Healthcare. September 24, 2026. https://www.fiercehealthcare.com/finance/hospitals-use-ai-coding-tools-cost-bcbsa-plans-942m-more-similar-care-analysis
Blue Cross Blue Shield Association. Analysis examines AI in hospital billing as spike in complex patients adds nearly $1 billion in extra costs. September 24, 2026. https://www.bcbs.com/about-us/association-news/bcbsa-analysis-ai-coding-tools-affects-healthcare-costs
Blue Cross Blue Shield Association. Studies show how AI is shaping hospital billing trends. March 2026. https://www.bcbs.com/news-and-insights/report/ai-boosting-hospital-billing
Carson JL, Stanworth SJ, Guyatt G, et al. Red blood cell transfusion: 2023 AABB international guidelines. JAMA. 2023;330(19):1892-1902. https://doi.org/10.1001/jama.2023.12914
Centers for Medicare & Medicaid Services and National Center for Health Statistics. ICD-10-CM Official Guidelines for Coding and Reporting, FY 2026. Sections I.A.19 and III. https://www.cms.gov/files/document/fy-2026-icd-10-cm-coding-guidelines.pdf
Centers for Medicare & Medicaid Services and National Center for Health Statistics. ICD-10-CM Official Guidelines for Coding and Reporting, FY 2026. Sections I.A.19 and III. https://www.cms.gov/files/document/fy-2026-icd-10-cm-coding-guidelines.pdf
Muoio D. Hospitals' use of AI coding tools cost BCBSA plans $942M more for similar care: analysis. Fierce Healthcare. September 24, 2026. https://www.fiercehealthcare.com/finance/hospitals-use-ai-coding-tools-cost-bcbsa-plans-942m-more-similar-care-analysis
Watson KB, Wiltz JL, Nhim K, et al. Trends in multiple chronic conditions among US adults, by life stage, Behavioral Risk Factor Surveillance System, 2013-2023. Prev Chronic Dis. 2025;22:240539. https://www.cdc.gov/pcd/issues/2025/24_0539.htm
Kantor ED, Rehm CD, Haas JS, Chan AT, Giovannucci EL. Trends in prescription drug use among adults in the United States from 1999-2012. JAMA. 2015;314(17):1818-1831. https://doi.org/10.1001/jama.2015.13766
Emmerich SD, Fryar CD, Stierman B, Ogden CL. Obesity and severe obesity prevalence in adults: United States, August 2021-August 2023. NCHS Data Brief No. 508. September 2024. https://www.cdc.gov/nchs/products/databriefs/db508.htm
Harrison TR, ed. Principles of Internal Medicine. 1st ed. Philadelphia: Blakiston; 1950. 1,590 pages. Collation as described by Patrick's Rare Books. https://www.patricksrarebooks.com/product-page/harrison-principals-of-internal-medicine-1950-first-edition
Loscalzo J, Fauci AS, Kasper DL, Hauser SL, Longo DL, Jameson JL, eds. Harrison's Principles of Internal Medicine. 21st ed. New York: McGraw Hill; 2022. 4,384 pages. https://books.google.com/books/about/Harrison_s_Principles_of_Internal_Medici.html?id=lfC7zgEACAAJ
American Medical Association. Medicare physician pay has plummeted since 2001. Find out why. April 21, 2025. https://www.ama-assn.org/practice-management/medicare-medicaid/medicare-physician-pay-has-plummeted-2001-find-out-why
Medicare Payment Advisory Commission. Hospital inpatient and outpatient services. In: Report to the Congress: Medicare Payment Policy. March 2025:chap 3. https://www.medpac.gov/wp-content/uploads/2025/03/Mar25_Ch3_MedPAC_Report_To_Congress_SEC.pdf
Chartis. 2026 rural health state of the state. February 10, 2026. https://www.chartis.com/insights/2026-rural-health-state-state
Becker's Payer Issues. Payers ranked by 2024 profits. February 12, 2025. https://www.beckerspayer.com/payer/payers-ranked-by-2024-profits/
Simborg DW. DRG creep: a new hospital-acquired disease. N Engl J Med. 1981;304(26):1602-1604. https://www.nejm.org/doi/abs/10.1056/NEJM198106253042611
Office of Inspector General, US Department of Health and Human Services. Trend toward more expensive inpatient hospital stays in Medicare emerged before COVID-19 and warrants further scrutiny. OEI-02-18-00380. February 2021. https://oig.hhs.gov/oei/reports/OEI-02-18-00380.pdf
Rucker P, Miller M, Armstrong D. How Cigna saves millions by having its doctors reject claims without reading them. ProPublica. March 25, 2023. https://www.propublica.org/article/cigna-pxdx-medical-health-insurance-rejection-claims
Vogel S. Senate report slams Medicare Advantage insurers for using predictive technology to deny claims. Healthcare Dive. October 21, 2024. https://www.healthcaredive.com/news/medicare-advantage-AI-denials-cvs-humana-unitedhealthcare-senate-report/730383/
Murphy LC, Anishchenko M. Protecting hospital revenue in 2026: why DRG downgrades and clinical validation denials demand different strategies. Davis Wright Tremaine. January 27, 2026. https://www.dwt.com/insights/2026/01/2026-hospital-revenue-clinical-validation-denials
Roy S, S K S. US insurers and hospitals turn to new AI for age-old battle over charges vs payments. Reuters. March 12, 2026. https://www.thestar.com.my/tech/tech-news/2026/03/12/us-insurers-and-hospitals-turn-to-new-ai-for-age-old-battle-over-charges-vs-payments
Muoio D. Hospitals' use of AI coding tools cost BCBSA plans $942M more for similar care: analysis. Fierce Healthcare. September 24, 2026. https://www.fiercehealthcare.com/finance/hospitals-use-ai-coding-tools-cost-bcbsa-plans-942m-more-similar-care-analysis
Medicare Payment Advisory Commission. MedPAC releases March 2025 report to the Congress: Medicare payment policy. Press release. March 13, 2025. https://www.medpac.gov/wp-content/uploads/2025/03/March_2025_MedPAC_Report_Press_Release_SEC.pdf
US Attorney's Office, Southern District of New York. Manhattan U.S. Attorney files civil fraud suit against Anthem, Inc., for falsely certifying the accuracy of its diagnosis data. March 27, 2020. https://www.justice.gov/usao-sdny/pr/manhattan-us-attorney-files-civil-fraud-suit-against-anthem-inc-falsely-certifying
Optum. Transform CDI, coding and auditing using one clinically intelligent platform: Enterprise CAC and CDI 3D. https://cdn-aem.optum.com/content/dam/optum3/optum/en/resources/brochures/enterprise-cac-cdi-brochure.pdf
Healthcare Dive. Optum a step ahead in vertical integration frenzy. April 12, 2018. https://www.healthcaredive.com/news/optum-unitedhealth-vertical-integration-walmart/520410/
ACDIS. CMS releases FY 2026 ICD-10-CM code update. July 10, 2025. https://acdis.org/articles/news-cms-releases-fy-2026-icd-10-cm-code-update
Oppenheimer F. The State: Its History and Development Viewed Sociologically. Gitterman JM, trans. 1922. Excerpted at Online Library of Liberty. https://oll.libertyfund.org/quotes/franz-oppenheimer-argues-that-there-are-two-fundamentally-opposed-ways-of-acquiring-wealth-the-political-means-through-coercion-and-the-economic-means-through-peaceful-trade-1922
American Hospital Association. Fact sheet: artificial intelligence and coding intensity. July 31, 2026. https://www.aha.org/fact-sheets/2026-07-31-fact-sheet-artificial-intelligence-and-coding-intensity