Who Owns the Join

A FIELD GUIDE

The hidden toll booths in AI, healthcare, and the law, and the small builder trying to do right by the patient.

Jason Newell Writer, technologist, builder of local-first AI

A solo builder sits in front of a workstation at one in the morning, watching a number climb. Every question the system answers, every label it assigns, every small reshuffle of text trips a meter somewhere, and the meter is counted in tokens. The build works. The demo lands. And still the number climbs, and the builder feels the particular unease of having rented something they thought they owned.

This is the ordinary condition of building with AI right now. Not a dramatic failure. A slow leak. The system reaches for a language model the way an earlier generation of software reached for a database call, as the default way to get anything done, and most of what it asks the model to do never needed a model at all.

Part I. The bill is not the disease

Tokens are the meter, not the malady. The real fragility is reaching for a probabilistic text generator as a general purpose function. A great deal of what teams route through a model is classification, extraction, routing, deduplication. That work is deterministic, or close to it, and plain code or a small trained classifier does it cheaper and more reliably. The token meter just makes the waste legible.

I have done this myself, reached for the model because it was sitting right there, and let it answer a question a regular expression had already answered for twenty years. So when a company is said to rely too heavily on AI, the honest description is usually narrower and stranger. It is renting frozen, generic cognition for jobs that never needed cognition, and paying by the word for the privilege.

Part II. Routing is not the same as agreement

Watch what the popular single interface products actually do with a question. They route it. A meta layer reads the request, classifies it, and dispatches it to whichever model is judged best for that kind of task, sometimes splitting a complex job into pieces and handing each piece to a different model. That is real engineering and it is genuinely useful. It is also not integration, and it is certainly not consensus. Each piece still lands on one model, and nobody in the pipeline is assigned to check anyone else.

The committee layer does now exist. In early 2026 one of these products launched a mode that runs a question across three frontier models at once and uses a synthesizer to show where they agree and differ. On its face that is the committee a skeptic would ask for. But it is opt in, gated to the most expensive tier, and the product says plainly that running three models does not guarantee accuracy. The deeper problem is independence. Three frontier models trained on overlapping data are not independent voters. They tend to agree on the same plausible wrong answer, so convergence buys a confidence that was never earned. Agreement among similar minds is not a second opinion.

The thing that ever felt like real collaboration was built differently. It was adversarial. A code reviewer whose whole job is to find fault in code it did not write, with a different objective from the system that produced the code, catches what the author was blind to precisely because it is not the author. The principle has a name now in the field, separation of concerns, and the cleanest statement of it is blunt. The model that wrote the code should not be the one that reviews it, because it carries the same blind spots in both directions.

There is an irony worth recording. The clearest consumer example of that adversarial pattern, a free code review that ran on every pull request, is being retired this very month. Its consumer version shuts down in mid July of 2026, and the broader command line and editor tools for individuals have been folded into a new agent platform. The capability is not dying. It is moving behind an enterprise paywall and a rebrand. The most collaborative thing in the toolbox is being repriced.

The other tools each have a hard edge in the same place. One vendor runs nine review agents in parallel to catch cross file regressions a single pass misses. A widely used reviewer looks only at the diff, not at how the change interacts with the rest of the codebase, misses roughly two in five real vulnerabilities on one benchmark, and returns different findings on a second run of the same review. A 3D tool will rig a character in under a minute and hand back a generic skeleton and a shelf of canned clips, then leave you to finish the weight and the momentum by hand in another program, because physicality has to know the body it is moving. Another model is quick and cheap and good at a few things and unreliable at several others. The boundary is always the same shape. The tool does the bounded core, and a human, or a second genuinely independent system, has to carry the seam.

So verification is not a matter of head count. It is a matter of independence, and of the fact that checking an answer is usually easier than producing one. The value is the asymmetry, not the number of models in the room.

Part III. Own the loop

If consensus and curiosity both cost more rather than less, neither can be the cure for the bill. The cure is ownership of the learning loop. A smaller model trained on your data, run locally, that learns the structure of your domain can beat a giant generic one inside that domain, and it grows cheaper over time instead of billing you forever. The strongest case for local is not privacy. It is that you own the loop that improves.

And you stop baking the moving parts into frozen weights. A brain does not recompute seven times eight from first principles every time. It cached the answer, which is exactly right, because that frees the expensive machinery for the things that actually vary. The rule is tiered. Compute what changes. Look up what is authoritative and moving. Cache only the stable, high frequency primitives.

This is a rule the healthcare world already wrote without naming it. Treat the record as a structured content store with interchangeable heads. The chart does not copy the scan into itself. It holds a pointer, and the pointer resolves to the archive when someone needs the pixels. The same discipline carries straight into an AI system. Compute what varies, look up what is authoritative, cache the stable primitives, and keep the curiosity in the training loop where surprise is useful rather than in the serving loop where you want the system bounded and dull.

Part IV. When the data is not yours to memorize

Here is the wall almost nobody plans for. In a real domain, some of what looks like data to memorize is not free knowledge at all. It is licensed property, and healthcare is the clearest case in the country.

The procedure codes, the ones a clinician uses to bill for what was done, are created and owned by the American Medical Association, a private association, not any government body. They are copyrighted and have never entered the public domain, even though federal rules require their use, which means you pay to use the codes the law obliges you to use. Even the government operates under a license for them. The association reported hundreds of millions of dollars in royalty and credentialing revenue in a recent year and does not break out what the codes alone bring in. A code you are legally required to use to get paid for treating a patient, that even the government rents from a private body. That is the baseline.

The diagnosis codes are the clean counterexample. They are maintained free of charge by the federal health agencies, downloadable by anyone, updated on a public calendar. You can train a model on the entire diagnosis set and ship it. So the field is split down the middle. One mandatory code set is private and metered. The other is public and free.

Mental health stacks both private walls on top of the free road, and it is worth walking through because it is the sharpest instance of the whole pattern. The manual that defines the disorders is owned and copyrighted by the psychiatric association, another private body, and you buy the book. But there are no separate codes in it. The codes printed alongside each disorder are the free diagnosis codes. The manual is a paywalled criteria text that reprints the public codes, and its criteria are deliberately more specific than the code can carry, so the clinician needs the paywalled book to satisfy a documentation standard the free code cannot express on its own.

Then the service itself is billed with the private procedure codes, and those are timed to the minute. Sixteen to thirty seven minutes is one code. Thirty eight to fifty two is another. Fifty three and beyond is a third. One minute moves the code, the longer one pays more, and the longer one is audited harder, against thresholds that are not published. So a clinician is bracketed by two private toll booths, the manual that defines the diagnosis and the codes that price the session, with a stretch of free public road in between.

The pieces are joined, and the joins are where the quiet damage lives. The manual maps to the diagnosis code, but the mapping loses specificity, so the chart ends up saying less than the clinician knew. The diagnosis has to justify the service code, a medical necessity pairing the payer checks, so a valid diagnosis and a valid procedure can still be refused if the pairing is wrong. The crosswalk software that automates all of it is itself built on the paywalled tables. Each join has a different failure mode, and every one of them can be broken while every individual code looks correct.

For an in house AI the consequence is concrete. You can train on the free diagnosis codes and redistribute the result. You cannot reproduce the private procedure codes or the diagnostic manual and hand them to a backend without infringing. The domain data you were told to memorize is, in part, someone else's copyright, and the rules that grade whether your coding is correct live in payer policies that vary and audit thresholds that nobody posts.

Part V. The same machine, other floors

None of this is a healthcare quirk. It recurs anywhere a private body gets to mint the unit everyone is required to use, and the two cleanest parallels are sitting in the open in law and finance.

In law it is the most naked version of all. Private organizations write technical standards, the codes for how a building is wired or how steel is poured. A legislature then adopts a standard by citing it, and at that moment the standard is the law. But the organization keeps the copyright and sells access, and one such body earns most of its revenue from selling the standards it writes. So you can be required to obey a binding rule that you have to pay a private organization to read. The courts are fighting about exactly this right now. A 2026 appeals ruling held that publishing standards once they have been written into law is likely fair use, a 2020 Supreme Court decision kept a state from copyrighting its own official annotated code, and the deeper question, whether the copyright survives at all once a standard becomes law, has still not been settled, only worked around case by case. One judge put the stakes in a single sentence. People should not have to pay to read the laws that bind them.

In finance the twin is the identifier on every North American security. It is owned by a bankers association, operated by a private firm, and you pay to issue one and pay to use them, with the association keeping a cut. Antitrust suits argue that the fees exist not because the nine digit string is valuable but because it was designated the standard, and they seek more than a billion dollars in recovered charges. Captivity by designation, in the plainest terms.

The purest version is the credit rating, a private classification referenced into regulation, with a concealed method and a conflicted pay model, that substitutes one private verdict for any independent assessment of risk. In 2008 the verdict was wrong and everyone was bound to it, which is the whole danger in one example. When the unit is singular, owned, and mandated, you cannot form a second opinion against it. Monopoly by designation forecloses consensus by construction. There is one such code, one such identifier, one incorporated standard. The monopoly is not on the answer. It is on the question, on the right to define the unit everyone else has to denominate their claims in.

Part VI. Collaboration and its twin

So is the future more cooperation or more land grabbing. Both, on different layers, often inside the same companies, and the relationship between the two is the entire story.

In AI the model layer is diverging. Models are specializing rather than converging, and on at least one large platform no single model handled more than a quarter of enterprise tasks by late 2025, while the big labs wall their own ecosystems. The connective layer is converging, toward open protocols and shared standards. But watch where the value is expected to settle. The bet across the industry is that the orchestration layer, not the model, is where the money accrues, the way the abstraction layers above commodity infrastructure captured more value than the infrastructure underneath. The standard opens, and the chokepoint climbs a floor.

Healthcare is living this in real time, and the cooperative half is real. A federal health technology effort announced at the White House in the middle of 2025 grew from sixty organizations to more than seven hundred within months. Networks are being pushed to expose modern interfaces by the middle of 2026, information blocking is being enforced more aggressively, a national provider directory is being stood up, and a patient facing campaign is trying to kill the clipboard. A nationwide exchange had already moved more than half a billion records. The rhetoric throughout is the patient holding their own records and choosing their own apps.

And at the same time the same names are consolidating the chokepoints. The payer that also owns the clinics and the data analytics captures a margin at every stage of the patient journey, and that model is now under a federal antitrust offensive built on a foreclosure theory, that whoever controls the bottleneck can foreclose rivals by withholding it or charging them dearly. The affordability evidence so far cuts against the consolidation story. After one such firm bought a chain of surgery centers, the prices it charged other insurers rose. The integrated case deserves a fair hearing, since coordinated care can in principle lower the total cost. But the same scale that could coordinate care is the scale that raises prices and forecloses competitors, and the jury, in this instance literally, is still out.

So the question of who is best positioned splits cleanly in two. Best positioned to capture is whoever owns the join. Best positioned to benefit is the patient, in whose name the entire apparatus is built and who holds the least leverage of anyone at the table. Those are not the same actor, and the distance between them is the answer to the question.

The lockout is a race, because the door is being pried open and walled at the same moment. It is pried open by free interfaces, by enforcement, by a deliberately permissive pledge, by an app directory for patients. It is walled by the gates on that directory, the identity verification partners and the evaluations and the review, by the standing toll of needing a sponsoring customer and a months long security review, by the data foreclosure the regulators are litigating, and underneath all of it by the code paywalls that never went anywhere. A small builder still has to pass the gates, find a sponsor, avoid being foreclosed by whoever owns the data, and pay for the vocabulary the claim is written in.

And here is the tell that closes the loop. At the gathering where all these companies demonstrated their bright new interoperability, the hardest problem in the room was not the data format. It was matching a patient accurately to their own data. Identity. The load bearing wall. It did not dissolve at national scale. It got promoted. Whoever ends up owning patient identity reconciliation owns the next toll booth, and that is exactly the layer the consolidated payer and the orchestration platform are both moving to own.

What stays open

So what does the future hold. Not a clean win for openness, and not a clean victory for enclosure. A standard that keeps getting freer while the chokepoint keeps climbing to the next floor. More collaboration and more consolidation from the same hands, because they were never opposites.

The clever workarounds are not a phase we pass through on the way to something settled. The startup threading the free interfaces, the nonprofit posting the building code, the engineer wiring two models to argue with each other, they are the permanent friction of people trying to make their own data mean more than the platform intended, against a structure that keeps installing a meter on the part that makes data mean anything at all.

And the patient is still the reason given for all of it and the owner of none of it. The scan that is theirs, the diagnosis set down in a manual they cannot read, the bill coded in a language they have to pay someone else to speak, and their own name, still the one thing in the system that nobody can reliably match to the person standing in the room. The future is not whether the data moves. It is who is standing at the seam when it does, and whether the person it describes is ever the one holding the map.

Sources and references

Primary and official sources are listed first within each area. Market, litigation, and product details current as of mid 2026 and subject to change.

• American Medical Association. CPT licensing frequently asked questions. https://www.ama-assn.org/practice-management/cpt/cpt-licensing-frequently-asked-questions-faqs
• Centers for Medicare and Medicaid Services. License for use of Current Procedural Terminology (AMA). https://www.cms.gov/license/ama
• Centers for Disease Control and Prevention, NCHS. ICD-10-CM files (free public download). https://www.cdc.gov/nchs/icd/icd-10-cm/files.html
• Centers for Medicare and Medicaid Services. ICD-10 codes and updates. https://www.cms.gov/medicare/coding-billing/icd-10-codes
• American Psychiatric Association. DSM frequently asked questions (there are no DSM codes). https://www.psychiatry.org/psychiatrists/practice/dsm/frequently-asked-questions
• American Psychological Association Services. Diagnostic coding and ICD-10-CM for billing. https://www.apaservices.org/practice/reimbursement/icd-diagnostic/dsm-5
• Centers for Medicare and Medicaid Services. Billing and coding: psychiatric diagnostic evaluation and psychotherapy services (time bands). https://www.cms.gov/medicare-coverage-database/view/article.aspx?articleId=57520
• American Psychological Association Services. Psychotherapy codes for psychologists. https://www.apaservices.org/practice/reimbursement/health-codes/psychotherapy
• Google for Developers. Sunset of the consumer version of Gemini Code Assist on GitHub. https://developers.google.com/gemini-code-assist/docs/deprecations/consumer-code-review
• Google Cloud Documentation. Gemini Code Assist release notes and transition to Antigravity. https://docs.cloud.google.com/gemini/docs/codeassist/release-notes
• Perplexity. Introducing Model Council. https://www.perplexity.ai/hub/blog/introducing-model-council
• Congressional Research Service. Copyright in standards incorporated by reference into law and the Pro Codes Act (R47656). https://www.congress.gov/crs-product/R47656
• Electronic Frontier Foundation. Another court rules copyright cannot stop people reading and speaking the law (ASTM v. UpCodes). https://www.eff.org/deeplinks/2026/04/another-court-rules-copyright-cant-stop-people-reading-and-speaking-law
• WatersTechnology. Class action lawsuit takes aim at CUSIP, S&P, FactSet and ABA. https://www.waterstechnology.com/regulation/7936086/class-action-lawsuit-takes-aim-at-cusip-sp-factset-aba
• Centers for Medicare and Medicaid Services. Health technology ecosystem: interoperability framework. https://www.cms.gov/health-technology-ecosystem/interoperability-framework
• ONC, HealthIT.gov. Trusted Exchange Framework and Common Agreement (TEFCA). https://www.healthit.gov/topic/interoperability/policy/trusted-exchange-framework-and-common-agreement-tefca
• Fierce Healthcare. HIMSS26: HHS officials offer updates on interoperability and information blocking enforcement. https://www.fiercehealthcare.com/health-tech/himss26-hhs-officials-offer-updates-interoperability-efforts-information-blocking
• TechTarget. Data standardization, the Health Tech Ecosystem, and TEFCA interoperability goals. https://www.techtarget.com/searchhealthit/feature/Data-standardization-key-to-Health-Tech-Ecosystem-TEFCAs-interoperability-goals
• Ropes and Gray. HHS announces new and renewed efforts to promote health information sharing. https://www.ropesgray.com/en/insights/alerts/2025/09/hhs-announces-new-and-renewed-efforts-to-promote-health-information-sharing-and-interoperability
• Healthcare Brew. After Optum bought ambulatory surgical centers, prices went up (Health Affairs study). https://www.healthcare-brew.com/stories/2026/02/13/optum-bought-ambulatory-surgical-centers
• Wedbush MarketMinute. DOJ intensifies structural challenge against UnitedHealth's vertical integration. https://investor.wedbush.com/wedbush/article/marketminute-2026-4-8-the-end-of-the-flywheel-doj-intensifies-structural-challenge-against-unitedhealths-vertical-empire