In short. Outside HIPAA, a federal breach-notice rule, a few state laws and the holder’s own promises may guard a copied record. Those promises can change, and they can pass to a buyer. We’d change our view if HHS or a court said a copy sent to a consumer app at the patient’s direction stays under HIPAA, or if a published test found full histories clearly safer.
What this is. A report built mostly from primary sources: HHS guidance, court opinions, FTC releases, vendor notices and help pages, and our own count of HHS’s breach portal. Law-firm summaries, news reports and one preprint fill gaps.
Contents
Someone will hold the outside record
On 3 November 2026 Andreessen Horowitz (a16z) published a report saying companies are racing to become the consumer’s healthcare home. It describes that home as one app that knows your records, orders labs, reads your wearable and connects you to clinicians.[4] An appendix shows that a16z backs several of the contenders.
Whoever wins will hold what we call the outside record. This is the health data that gathers outside the clinic: chats, wearable streams, self-ordered lab panels and copied medical files.
KFF’s poll of 1,343 US adults in February and March 2026 shows that sharing medical information with AI is already common. Of those who used AI for their physical or mental health, 41 percent had uploaded personal medical information. That is 13 percent of all adults.[23]
The report leaves one thing out. What binds whoever holds that record when they are sold, breached or wrong? This essay answers by following one record from a clinic into a chatbot. Every legal reading here is ours, not a lawyer’s.
On our reading, the copy leaves HIPAA when it lands
HIPAA’s rules bind covered entities and their business associates.[14] Say a patient tells a covered entity to send records to an app that is neither. HHS says the data is then no longer protected by the HIPAA rules, and the covered entity isn’t liable under HIPAA for what the app does with it next.[6]
We read that guidance to mean that a record copied from a clinic into ChatGPT leaves HIPAA the moment it lands.[6] Records from inside HIPAA can already cross over. OpenAI lists US hospital records and One Medical among the sources a ChatGPT user can connect.[16]
Most of the AI vendors we read put their consumer health apps outside HIPAA. Here is what each one says on its own pages.
| Vendor | What the vendor says |
|---|---|
| OpenAI | Health isn’t meant for clinical or covered-entity use, and OpenAI doesn’t offer a business associate agreement for it.[17] |
| Medical records that US users move to the Google Health app, whether sent from a provider or uploaded, are no longer subject to HIPAA’s protections.[19] | |
| Microsoft | It calls Copilot Health a direct-to-consumer wellness product. Its page says that “HIPAA does not apply to most direct-to-consumer wellness products.”[20] |
| Anthropic | It launched test versions of health-record connectors for US Claude Pro and Max subscribers in January 2026. It says it doesn’t train models on users’ health data. A dated note added in February says its HIPAA-ready products are for healthcare providers and payers. The launch text had listed consumers among their users.[21] |
| Amazon | It began rolling out Health AI to US customers on 10 March 2026, with no Prime or One Medical membership needed, in what it calls a HIPAA-compliant setting.[22] |
Amazon is the telling exception. It says it trains Health AI models on abstracted patterns, without directly identifying data. It uses protected health information, it says, only for purposes HIPAA permits.[22] Staying inside HIPAA doesn’t mean staying out of training.
Outside HIPAA, the FTC’s breach rule only says who to tell
Once a copy leaves HIPAA, the FTC rule that may bind its holder is the Health Breach Notification Rule. It is about notice. If health data leaves an app the rule covers and the user didn’t permit it, the company must tell users, the FTC and sometimes the media. As we read it, the rule sets no limit on what an app may collect, keep or analyze.[24]
The rule dates from the 2009 Recovery Act, passed partly because HIPAA didn’t cover some holders of personal health records.[24] In January 2021, as the FTC proposed a settlement with the period-tracking app Flo Health, two commissioners wrote that the agency had never brought an action under the rule.[25]
The first case came in February 2023. GoodRx agreed to a $1.5 million civil penalty for not telling users, the FTC and the media that it had shared their health information with Facebook, Google and others.[26]
A 2024 update, adopted over two commissioners’ dissent, took effect on 29 July 2024. It says the rule reaches apps and connected devices that track things such as fitness, fertility, sleep and mental health. It also counts an unauthorized disclosure as a breach, as well as a hack.[24] In September 2026 the FTC withdrew its 2021 health-app policy statement as obsolete, citing that update. The release withdraws guidance only. It doesn’t restate the rule’s status.[27]
The latest FTC health-privacy case we found, against Hims & Hers, rests on other laws, and an appendix gives them.[28]
Washington’s My Health My Data Act binds holders more tightly. It covers health information a company derives from non-health data, including by machine learning. Its main duties began on 31 March 2024, or 30 June 2024 for small businesses. A company needs consent to collect such data and separate consent to share it, unless a requested service needs the step. Selling it takes a signed authorization.[29]
Two bills would add more, a revised New York health-privacy bill and Senator Cassidy’s Health Information Privacy Reform Act. Neither had become law by 3 November, and an appendix gives their status.
That leaves company promises doing much of the work. The next two sections show how a promise can move, and where it can go.
OpenAI’s health promises moved between January and September
OpenAI’s health product sits on HIPAA’s edge. It launched Health on 7 January, calling it a separate space.[15] On 23 July it began rolling out Health to US users, and said more than 300 million people a week bring health questions to ChatGPT.[16] Between January and September, three promises changed.
| Promise | Earlier | Later |
|---|---|---|
| Health data in other chats | 7 January: Health contents and memories would never flow into other chats.[15] | July: users can let ChatGPT draw on their connected health data across chats.[16] |
| Training | 7 January: Health chats wouldn’t train OpenAI’s foundation models.[15] | July post and help center: connected records, Apple Health data and the chats that use them stay out of training and ad targeting. Other health chats follow the user’s normal training setting.[16][17] |
| Asking before use | July: ChatGPT would ask by default before using connected records to shape a reply.[16] | Since September: new health connections follow the user’s global plugin setting, which by default allows low-risk actions.[17] |
The training promise narrowed. The July post and the help center now draw the training line around connected data rather than around health as a subject.[16][17]
The promises also sit in four documents, and not every promise appears in the privacy notice. OpenAI’s Health Privacy Notice, updated on 29 June 2026, cites Washington’s and Nevada’s consumer health data laws. By default, it says, OpenAI doesn’t use Health data to improve its foundation models. A limited number of staff and service providers may still see Health data to improve model safety, unless the user opts out.[18]
The notice says nothing about ads.[18] We found the no-ads promise only in the July post and the help center.[16][17] OpenAI’s main privacy policy says personal data may pass to a successor in a bankruptcy or other deal.[65]
That last clause is where the next section begins.
A sale can carry the promises to a buyer
When a holder is sold, its promises can go with the data. At 23andMe that happened in public, and the customers’ written privacy promise allowed it.
Since 2007, the bankruptcy court found, every version of 23andMe’s privacy statement had allowed a sale of personal data in a merger, acquisition or asset sale. Its terms would then bind the buyer. In 2022 the company added bankruptcy to that list. That December it dropped a bold line from its California notice saying it didn’t sell personal data.[2]
On 10 June 2025 23andMe’s interim chief executive, Joseph Selsavage, testified before the House Oversight Committee about the bankruptcy sale. Since the bankruptcy was announced, he said, about 1.9 million customers, roughly 15 percent, had asked the company to delete their data.[1]
Seventeen days later a bankruptcy judge in St. Louis approved the sale to TTAM Research Institute, a nonprofit headed by co-founder Anne Wojcicki, for $305 million. The ruling required no customer opt-in, though it said that in the abstract “a company’s sale of genetic data is a scary proposition.”[2]
The consumer privacy ombudsman, who advises the court on customers’ privacy, recommended separate, express consent from each customer before their data moved. The court declined, for practical reasons and in deference to lawmakers.[2]
The deal moved the data into a new subsidiary and sold that subsidiary’s shares. The court called that a sale of customer data only in a technical sense. An opt-in rule, it said, created a significant risk that the sale wouldn’t close. Lawmakers could forbid such sales, it added, but hadn’t taken so firm a stance.[2]
After the sale, the 23andMe Research Institute’s privacy statement, updated 19 May 2026, still says personal data may be accessed, sold or transferred in a bankruptcy, merger or asset sale. Genetic data may go only to a US nonprofit research institution or a qualifying company, and the buyer must adopt the statement. Even after an account is deleted, the company or its lab keeps genetic data, date of birth and sex, citing lab rules.[33]
A bill would add a consent step. Senator Cornyn’s Don’t Sell My DNA Act would add genetic information to the Bankruptcy Code’s rules on personal data. A court then couldn’t approve its use, sale or lease under those rules unless every affected person, including non-parties, agreed in writing after the case began. The federal bill-status file, updated 26 September 2026, lists nothing since its introduction and referral to committee on 22 May 2025.[36]
Our essay on cheap inference argues that every stored copy of personal data gets easier to mine as machine reading gets cheaper. Health data fits that argument closely, because a person can’t swap out a genome or a diagnosis the way they reset a password. An appendix covers 23andMe’s breach settlements and a privacy verdict against Meta over the Flo app.
A breach shows how much is at stake
Even holders bound by HIPAA’s rules get breached at scale. Change Healthcare, part of UnitedHealth Group’s Optum, moves data, claims and payments between doctors, pharmacists, health plans and governments.[40] On 31 July 2025 it told HHS that about 192.7 million people had been affected.[41] An appendix sets out how the attack happened.
HHS’s portal lists Change as a business associate, and its entry is the largest there. By our tally of the portal, large-breach reports submitted in 2024 listed 290 million affected people, the highest of any year since 2009. The tally counts reports, so some people appear more than once. Without Change’s entry, 2024 would rank below 2023’s 183 million.[42]
A healthcare home would hold the outside record for hundreds of millions of people. We think that record is more personal than what Change held. If it sits outside HIPAA, HIPAA’s security rules wouldn’t apply to it.[6]
When the holder is wrong
Holders also turn the record into readings and advice, and few checks reach their accuracy. Device law decides whether FDA reviews a health product’s accuracy, and it decides by what the product is for.
On 14 July 2025 FDA warned WHOOP that its blood-pressure feature was a device sold without the clearance or approval it needed.[45] In January 2026 FDA guidance said some non-invasive blood-pressure estimators may count as low-risk wellness products. To qualify, their values must be validated, and they must not be meant to diagnose or treat disease, among other conditions.[46]
In June 2026 FDA told WHOOP it didn’t plan to enforce device rules for the changed feature, citing the company’s changes and the new guidance.[47] The wellness line doesn’t decide who keeps the reading, which sits in a company’s database either way.
Chatbot advice faces a similar gap. In a randomized, preregistered trial published in Nature Medicine in February 2026, 1,298 participants worked through ten written medical scenarios. Each was randomly assigned to one of three 2024 models or to a control group free to use any source.[48]
Tested alone, the models found relevant conditions in 94.9 percent of cases. People using them found relevant conditions in fewer than 34.5 percent and chose the right next step in fewer than 44.2 percent, no better than the control group. Standard tests of medical knowledge, and simulated patients, didn’t predict these failures, the authors write. They recommend testing with human users before public release.[48]
An appendix covers a triage study of ChatGPT Health. It also covers one inspector-general study of how often Medicare Advantage insurers’ skilled-nursing denials were appealed and overturned in June 2024.
What companies, regulators, lawmakers and patients can do
Companies that hold the record can make their promises easy to check. Every promise about training, ads, retention and sale belongs in the privacy notice itself, dated, with a public log of each change. OpenAI’s notice covers training but not ads.[18]
The notice should also commit that a merger, sale or bankruptcy can’t widen how the record is used without fresh, opt-in consent from each person. 23andMe’s ombudsman urged similar consent before any customer’s data moved.[2]
The assistant should hold only the part of the record each task needs, for as long as the task needs it. A healthcare home that books, orders and prescribes will need a person before each decision, an appeal people use after it, and a record of both. Before the assistant’s public release, someone should test it with human users.[48]
Regulators and lawmakers have tools on the table. The FTC could keep enforcing the breach rule, whose text counts an unauthorized disclosure as a breach.[24] A bankruptcy rule like the Don’t Sell My DNA Act should cover all health data, including genomes.[36] Insurers that use prediction tools, like the one an appendix describes, could publish how often their denials are appealed and overturned, so anyone could measure the human check.[54]
One finding would weigh against keeping the record small. That would be a health assistant that gives clearly safer answers with a person’s full history than with only what each task needs. We haven’t found a published measurement of that for any health assistant. We’d want one before any company holds the outside record at that scale.
Much of what binds a holder is its own promises. So before you connect a hospital record to a chatbot, read every place the company makes them. At OpenAI that means four documents: the Health Privacy Notice, the July post, the help center and the main privacy policy. And deleting your account may not end every copy, as 23andMe’s statement shows.[33]
Scope and limits
- What we did
- We read the five vendors’ health pages, the 23andMe court record and FTC releases. We tallied the export of HHS’s breach portal on 3 November 2026 by year of report. HHS’s reports to Congress count by year of occurrence, so the totals differ. We checked the forecast against market data and the authors against a16z’s portfolio.
- What we didn’t do
- Our reading of HIPAA, the FTC rule and the state statutes is our own, not a lawyer’s. We ran no experiment and tested no product. We read several company and HHS pages through Internet Archive copies rather than live pages, and help-center text can change without notice.
- What this can’t tell you
- Whether named products mishandle data; a missing promise is not evidence of practice. Whether automation caused UnitedHealthcare’s rise in denials, or how many skilled-nursing denials in the inspector general’s data were wrong. How chatbots do with real patients and today’s models. This isn’t medical advice, and we don’t recommend tools.
- Confidence
- High that a record sent to a consumer app at the patient’s direction leaves HIPAA. HHS and Google say so, though the reading is ours, not a lawyer’s. Moderate on the 2024 breach total, which counts some people more than once. Lower on the ChatGPT Health triage figure in the appendix, which rests on four vignettes.
- What would change our view
- A published measurement showing that a health assistant gives clearly safer answers with a person’s full history than with only what each task needs. On the law, an HHS statement or a court ruling that a record sent to a consumer app at the patient’s direction stays under HIPAA.
- Corrections
- We log every correction, with its date, on How we work. Send one through a research enquiry.
Appendix
The report that prompted this essay, and the contenders its firm backs
On 3 November 2026 Daisy Wolf of a16z announced its State of Longevity report, written with Eva Steinman.[3] That day the health-tech writer Mansi (@mansihukmani_) shared it with a one-line forecast of her own: “The biggest company in the world will be a healthcare company.”[5] Mansi’s forecast isn’t new to a16z. In 2022 Daisy Wolf and Vijay Pande wrote that a consumer health tech company would be the world’s biggest.[10]
The firm that named the race for the healthcare home backs several contenders. Its portfolio page lists OpenAI and four of the six younger firms the report names: Function Health, General Medicine, Counsel Health and Ferry Health. Oura and WHOOP aren’t on it.[11] In October 2026 a16z led General Medicine’s $120 million Series B, with Eli Lilly investing alongside.[13] In January 2025 a16z announced a biotech fund that will deploy up to $500 million, with Lilly contributing all the capital.[12]
The report’s text is silent on these ties. Its website and newsletter add a standard disclaimer that the companies named aren’t representative of all a16z investments, with a link to its investment list. The website’s authors’ panel also links their post on the General Medicine investment. Both versions describe the survey behind the report only as 500 randomly selected US consumers. They give no panel provider, field dates, weighting or margin of error.[4][67]
The forecast doesn’t match today’s market. At the close on 3 November 2026, CompaniesMarketCap ranked the chipmaker NVIDIA first, at about $4.99 trillion, and Lilly 12th, the top health company.[7] In our reading, Lilly is a drug business whose value doesn’t rest on holding the outside record. Demand for tirzepatide drove it to $1 trillion in 2025, CNBC reported.[8] The drug brought in most of Lilly’s revenue in the first nine months of 2026.[9]
More on the breaches, cases and bills
How the Change attack happened. UnitedHealth’s chief executive, Andrew Witty, gave Congress this account in May 2024. On 12 February 2024 criminals used compromised login details to get into one of Change’s Citrix remote-access portals, which had no multi-factor authentication. On 21 February a group calling itself ALPHV or BlackCat set off ransomware inside its systems.[40]
What breaches cost. IBM’s yearly breach-cost study named healthcare the most expensive industry for the 14th year running in its 2024 report.[43] Its 2025 report found the same, at an average of $7.42 million.[44]
23andMe’s settlements. At 23andMe, bankruptcy also shrank what a breach cost the holder. In July 2026, 42 attorneys general settled their claims over 23andMe’s 2023 breach with the bankruptcy trustee. The settlement was for $150 million in allowed claims, which are debts the estate accepts. Because the estate’s funds are finite and other claims compete for them, recovery is capped at $18 million.[34]
California separately settled its breach claims and its appeal over the sale in October, for $6 million, pending the bankruptcy court’s approval.[35]
Flo Health. A record can also move sideways while the app holding it is still running, as Flo Health’s case shows. On 1 August 2025 a federal jury found that Meta had broken California’s Invasion of Privacy Act.[37][38] Meta had obtained ovulation and menstrual information that women entered into the Flo app. In September the court held that the law’s $5,000 per violation applies whatever the actual harm.[38]
Google, Flo and the analytics firm Flurry agreed to pay a combined $59.5 million. A final approval hearing was set for 29 October 2026, and we haven’t confirmed the outcome.[37] On 20 August 2026 the class’s lawyers still described a judgment against Meta as expected but not yet entered. We found no report of one by 3 November.[39]
Hims & Hers. In July 2026 the FTC, Utah and California sued Hims & Hers, alleging that it shared customers’ health information with Meta, Snap and others. The claims rest on the FTC Act, a subscription-billing law and state consumer-protection law.[28]
New York’s bill. New York’s legislature passed a broad health-privacy bill in January 2025, and the governor vetoed it on 19 December 2025.[30] A revised bill passed both houses on 3 and 4 June 2026. As of 3 November, the Senate’s record shows it hasn’t gone to the governor.[31]
Cassidy’s bill. In Congress, Senator Cassidy’s Health Information Privacy Reform Act would have HHS set privacy, security and breach standards for health information held outside HIPAA. It would also have HHS write minimum-necessary rules for health data used to train or operate AI models. It was reported out of committee on 4 August 2026 and sits on the Senate calendar.[32]
Triage errors and insurers’ care denials
Triage. Ramaswamy and colleagues found that ChatGPT Health under-triaged 52 percent of true emergencies in clinician-written vignettes. That means it sent them to less urgent care than they needed. It handled stroke and anaphylaxis correctly.[49] That figure rests on four vignettes. A later preprint ran those vignettes through six other leading models, not ChatGPT Health. Pooled under-triage ranged from 17 to 50 percent, depending on wording and answer format.[50]
Appeals. Insurers hold claims records inside HIPAA and decide what care they pay for. They show whether people can still say no. HHS’s inspector general found that in Medicare Advantage, the private Medicare plans, insurers overturned 95 percent of appealed skilled-nursing denials in June 2024. They overturned 97 percent of those issued by UnitedHealth’s unit naviHealth.[54]
Yet only 18 percent of such denials were appealed. The 95 percent leans heavily on UnitedHealth Group, which overturned nearly all of its appealed denials, while plans with fewer appeals overturned a far smaller share. We read that as a human check that works but rarely runs. The inspector general says its data alone can’t show how many denials were wrong. Its brief doesn’t mention algorithms.[54]
Software. Whether software drives Medicare Advantage denials of post-acute care, skilled nursing included, isn’t established. A lawsuit alleges UnitedHealth let an AI tool, nH Predict, stand in for doctors. Its surviving claims, allowed in February 2025, ask whether UnitedHealth kept its written promise that clinical staff and doctors make care decisions. The ruling took the allegations as true and made no finding about the tool.[51] A class-certification motion is due on 16 February 2027.[52]
A Senate majority staff report found that UnitedHealthcare’s post-acute prior-authorization denial rate rose from 10.9 percent in 2020 to 22.7 percent in 2022, while it automated the process. That shows timing, not cause.[53]
The rules. The rules keep some decisions with a person. CMS says an algorithm alone can’t justify ending post-acute care or denying an inpatient admission.[55] In April 2025, though, it declined to finalize proposed wider AI guardrails.[56] CMS’s own WISeR pilot uses AI to screen chosen services in traditional Medicare, with licensed clinicians deciding every recommendation not to pay.[57] California bars AI tools from denying, delaying or modifying care on medical-necessity grounds in the health plans it regulates.[58]
Neural data may be next, and states define it differently
The next addition to the outside record may be neural data. The a16z report expects brain wearables to become a consumer category within our lifetimes. Of the people it surveyed, 39 percent would use one.[4] In 2024 the Neurorights Foundation found that 29 of 30 consumer brain-tech companies appeared, by their policies, to have access to users’ neural data with no meaningful limits.[59]
The state laws that may bind those holders define the data differently.
| State | What its law does with neural data |
|---|---|
| Colorado | Covers data from the central or peripheral nervous system, treated as sensitive only when it’s used or meant to identify someone.[60] |
| California | Covers the same range, but excludes data inferred from non-neural information.[61] |
| Vermont | Covers only the central nervous system, from 1 January 2028. By Cooley’s count, Vermont became the fifth state to classify neural data as sensitive in June 2026.[62][63] |
| Delaware | Adds neural data to its sensitive categories from 1 January 2027, under a law signed in September 2026, law-firm summaries say.[66] |
On these texts, a wristband reading peripheral nerve signals falls inside California’s and Colorado’s definitions and outside Vermont’s. In Congress, the MIND Act, a bill to have the FTC study how to govern neural data, has sat in committee since its introduction on 29 September 2025.[64]
The open question
Can a consumer health assistant be useful while holding no more of the outside record than each task needs, and for no longer than the task lasts?
We map where the outside record sits in law, using the providers’ own documentation, the 23andMe court record and the FTC’s enforcement history, and set it beside breach, reliability and coverage-denial evidence. In the appendix we also check the forecast that greeted the report against dated market data, and note the investor ties behind the report. The next step we’d take is a public, dated register of each consumer health assistant’s commitments on training, retention, legal disclosure and change of control, taken from each product’s privacy notice and help pages and re-read each quarter so changes show.
References
A selected review of published sources, with Lugano’s analysis and proposed work identified in the text. Source dates and conditions matter; historical results are not current-model rankings.
- 1
Securing Americans’ Genetic Information: Privacy and National Security Concerns Surrounding 23andMe’s Bankruptcy Sale (Serial No. 119-32)
Official transcript of the hearing held on 10 June 2025, convened at 10:02 a.m. in HVC-210 of the Capitol. Witnesses included Anne Wojcicki (board member and bidder in the sale) and interim CEO Joseph Selsavage. Answering Rep. Gary Palmer of Alabama, Selsavage said about 1.9 million customers had asked for deletion since the bankruptcy announcement, roughly 15 percent.
Scope. Sworn testimony by company officers. The 15 percent figure was a rough oral estimate, and a deletion request is not proof of completed deletion.Back to the essay
- 2
Memorandum Opinion, In re 23andMe Holding Co., Case No. 25-40976-357 (Doc. 908)
Opinion dated 27 June 2025 in St. Louis approving the sale to TTAM Research Institute ($305 million). Finds about 13 million customers, more than 80 percent in research programs, and about 1.9 million account deletions since filing. Records that since 2007 the privacy statement allowed sale in a merger, acquisition or asset sale, that bankruptcy was added in 2022, and that a bold ‘we do not sell’ line was removed from the California notice in December 2022. Notes more than 30 states and D.C. initially objected. Summarizes the ombudsman’s recommendation of separate, express consent; declines opt-in partly because of a significant risk the sale would not close; calls the structure a sale of customer data only in a technical sense; contains the ‘scary proposition’ sentence and leaves the question to legislators.
Scope. A court’s reasoning on the record before it, deciding how state statutes apply to this deal. We could not confirm the signing judge’s name from the opinion, so the essay does not name the judge. We relied on the court’s summary of the ombudsman’s report, not the report itself.Back to the essay
- 3
Post announcing the a16z State of Longevity report
Post at 15:40 UTC on 3 November 2026 announcing a16z’s State of Longevity report, written with Eva Steinman and based on a survey of 500 Americans.
Scope. An announcement; the method is described in the report. Post text and time checked again on 3 November 2026. We read the linked X article from an archived copy; it carries no disclaimer.Back to the essay
- 4
The State of Longevity
Venture-investor report posted 3 November 2026. Credits the pandemic and ChatGPT with accelerating consumer health ownership; describes a race to become the consumer’s healthcare home and names Function Health, Oura, WHOOP, General Medicine, Counsel, Ferry Health, OpenAI, Anthropic, Amazon and Apple; argues value will move to the action layer. Survey figures: 54% use a wearable, 83% of users check it daily, 55% say its data never prompted a behavior change or doctor visit, 39% would use a non-invasive brain-sensing device. Cites tirzepatide’s nine-month sales and describes the FDA’s June 2026 close-out of its WHOOP warning letter. The charts give subgroup sizes (for example, n = 270 wearable users) and are labeled only as a16z State of Longevity survey results (2026). The authors’ panel on the page lists their post announcing a16z’s investment in General Medicine.
Scope. A self-interested source: a16z holds stakes in several companies it names (source 11). The only method statement in the text says 500 randomly selected US consumers were surveyed; no panel provider, field dates, weighting or margin of error is published. The report text discloses no holdings; the page carries a standard disclaimer and a link to a16z’s investment list. We treat its figures as a16z’s findings, not population estimates. The page was published at 15:16 UTC and last modified at 17:11 UTC on 3 November 2026.Back to the essay
- 5
Post quoting the a16z State of Longevity announcement
One-line post at 17:54 UTC on 3 November 2026, about two hours after Daisy Wolf’s, predicting that the biggest company in the world will be a healthcare company, quoting Wolf’s announcement.
Scope. An opinion with no evidence attached, cited for the claim it makes. Read from an archived copy of the post.Back to the essay
- 6
The access right, health apps, & APIs (HIPAA FAQs)
Once a covered entity sends health information, at the individual’s direction, to an app that is neither a covered entity nor a business associate, the information is no longer subject to the protections of the HIPAA Rules, and the covered entity is not liable under HIPAA for the app’s later use or disclosure. Content last reviewed 30 May 2025.
Scope. Guidance about records received through the HIPAA access right; it does not analyze any named product. We cite an Internet Archive copy captured 18 April 2026, not the live page.Back to the essay
- 7
Largest companies by market capitalization; Eli Lilly, NVIDIA and UnitedHealth market-cap pages
Read after the US close on 3 November 2026. NVIDIA first at $4.990 trillion. Eli Lilly 12th at $1.255 trillion, with the Lilly page also giving Nasdaq’s 3 November figure of $1.325 trillion. UnitedHealth 38th at $363.02 billion. No other healthcare or pharmaceutical company in the top 20 (Johnson & Johnson is 22nd). Related pages: companiesmarketcap.com, /nvidia/marketcap/, /united-health/marketcap/.
Scope. A live aggregator; figures move daily and depend on the share count used. Lilly’s two figures differ by about $70 billion; it ranks 12th on either.Back to the essay
- 8
Eli Lilly hits $1 trillion market value, a first in health care, as Novo Nordisk tumbles
Reports that Lilly reached a $1 trillion market capitalization on Friday 21 November 2025, the first health-care company to do so, driven by demand for its GLP-1 drugs Zepbound and Mounjaro.
Scope. News report; does not say whether the level held at the close.Back to the essay
- 9
Lilly reports third-quarter 2026 financial results (Form 8-K, Exhibit 99.1)
Release dated 29 October 2026. Revenue of $24,397 million in the third quarter and $67,170 million in the first nine months of 2026. Nine-month revenue of $29,328 million for Mounjaro and $14,440 million for Zepbound, both tirzepatide.
Scope. Company-reported quarterly figures. The $43.8 billion combined figure is our sum of the two brands.Back to the essay
- 10
The Biggest Company in the World
Essay of 11 November 2022 arguing that the biggest company in the world will be a consumer health tech company. Calls UnitedHealth Group the 8th biggest company in the world and argues its poor consumer engagement will cap its potential.
Scope. An investor’s thesis, cited for the claim and for its 2022 ranking of UnitedHealth.Back to the essay
- 11
Portfolio
Read 3 November 2026. Lists Function Health, General Medicine, Counsel Health, Ferry Health and OpenAI as portfolio companies. Oura and WHOOP do not appear; Anthropic appears only as the acquirer of another portfolio company.
Scope. a16z’s own listing; it may omit undisclosed holdings and gives no stake sizes. a16z’s announcements of its Function Health (June 2024), Counsel Health (October 2025) and General Medicine (October 2026) investments are consistent with it.Back to the essay
- 12
Andreessen Horowitz Partners with Lilly to Launch First-of-Its-Kind Biotech Ecosystem Fund
Announcement of 14 January 2025 (joint press release 10 January 2025): a fund that will deploy up to $500 million, managed by a16z Bio + Health, with Lilly “contributing all the capital”.
Scope. Announcement only; we did not check the fund’s current status.Back to the essay
- 13
General Medicine Raises $120 Million to Build the Healthcare Store
Release of 6 October 2026: a $120 million Series B led by Andreessen Horowitz, with Eli Lilly and Company among the other investors.
Scope. Company press release.Back to the essay
- 14
45 CFR 160.103, Definitions (covered entity)
Defines a covered entity as a health plan, a health care clearinghouse, or a health care provider who transmits health information electronically in connection with a covered transaction.
Scope. A definition only. Read in the eCFR version current to 1 October 2026.Back to the essay
- 15
Introducing ChatGPT Health
Post of 7 January 2026. Says over 230 million people globally ask health and wellness questions on ChatGPT every week (OpenAI’s de-identified analysis); Health is a separate space whose information and memories never flow back into non-Health chats; conversations in Health are not used to train foundation models; beta open outside the EEA, Switzerland and the UK; b.well named as records partner.
Scope. Company claims, not independently audited; the design changed in July 2026. We cite Internet Archive copies captured 7 January 2026 (launch day) and 28 October 2026, not the live page.Back to the essay
- 16
Launching Health in ChatGPT
Post of 23 July 2026. Health rolling out to logged-in US users aged 18 and over on Free, Go, Plus and Pro plans; more than 300 million people a week bring health questions; more than 70% of early users’ health conversations took place outside the dedicated space; users can let ChatGPT draw on connected Health information across conversations; connectable sources include Apple Health, US hospital records, One Medical and Function Health. The post also says connected records, Apple Health data and conversations that use them are not used to train foundation models or target ads, that other conversations follow the user’s training setting, and that by default ChatGPT asks before using connected records to personalize a response.
Scope. Company claims; the method behind the weekly figure is unpublished. Read through an Internet Archive copy captured 7 October 2026.Back to the essay
- 17
Health in ChatGPT (Help Center)
Health is not intended for clinical or covered-entity use and does not offer a Business Associate Agreement. Connected records, Apple Health data and conversations that use them are not used to train foundation models or target ads; other conversations follow the user’s training setting. From September 2026 new Health connections default to the global plugin permission, which by default allows low-risk actions. Conversations that use Health can create memories. Data from a disconnected source is deleted within 30 days. Says data connected in Health is subject to the Health Privacy Notice.
Scope. Help-center text can change without notice. Read through an Internet Archive copy captured 8 October 2026, when the page was marked as updated two days earlier.Back to the essay
- 18
Health Privacy Notice
Notice updated 29 June 2026; supplements the main privacy policy. Says information shared through Health features may be consumer health data under Washington’s My Health My Data Act and Nevada’s consumer health data law. Says that by default OpenAI does not use medical records and health data from connected accounts, conversations in ChatGPT Health, or conversations with Health Connect enabled to improve its foundation models, and that a limited number of authorized personnel and service providers might access Health data to improve model safety unless the user opts out. Allows disclosure in connection with a business transaction, to comply with legal obligations, and when OpenAI determines, in its sole discretion, that terms, policies or law have been violated. Says OpenAI does not sell users’ personal data. Contains no statement about advertising. The January 2026 version also had a “by default” training statement.
Scope. Read through Internet Archive copies captured 15 January and 11 October 2026, not the live page. The notice still uses the January product names (ChatGPT Health, Health Connect). The absence of a statement is not evidence of practice.Back to the essay
- 19
FAQs on privacy: Google Health
Says that for users in the United States, medical records requested from a provider into Google Health, or uploaded manually, are no longer subject to HIPAA protections. Personal AI interactions are not used to train the underlying generative models unless the user opts in under a separate agreement. Repeats Google’s commitment not to use Fitbit users’ health data for Google Ads.
Scope. Help page last updated 19 May 2026, read 3 November 2026; company commitments.Back to the essay
- 20
About Copilot Health
Says Copilot Health is a direct-to-consumer wellness product and that HIPAA does not apply to most direct-to-consumer wellness products; data is not used to train AI models, not used to build an advertising profile and not sold to insurers, employers or advertisers; conversation history and health profile are stored and can be deleted by the user.
Scope. Help page last updated October 2026, read 3 November 2026; company claims; no retention period stated. A banner says Copilot Health is temporarily unavailable during a move to a unified Copilot experience.Back to the essay
- 21
Advancing Claude in healthcare and the life sciences
Post of 11 January 2026: US Claude Pro and Max subscribers can connect HealthEx and Function (beta), with Apple Health and Android Health Connect integrations rolling out. Says Anthropic does not use users’ health data to train models. A changelog entry dated 7 February 2026 says the introduction was edited to clarify that HIPAA-ready products are specific to healthcare providers and payers. The launch text of 11 January (Internet Archive capture) said HIPAA-ready products let “healthcare providers, payers, and consumers” use Claude for medical purposes.
Scope. Company claims; no retention period for chats that use health connectors is stated.Back to the essay
- 22
Amazon launches Health AI agent on Amazon website and app with free 24/7 access to virtual care for Prime members
Post of 10 March 2026 (page metadata). Rolling out from that day, with no Prime or One Medical membership needed and a goal of reaching all US customers soon. All interactions happen within a HIPAA-compliant environment. Health AI can access relevant health purchases on Amazon. Amazon trains Health AI models on abstracted patterns without directly identifying information and uses protected health information only for purposes permitted under HIPAA; such information from One Medical and Amazon Pharmacy is not used to market general merchandise or by Amazon Ads.
Scope. Company statement; ‘abstracted patterns’ is not defined.Back to the essay
- 23
KFF Tracking Poll on Health Information and Trust: Use of AI for Health Information and Advice
Published 25 March 2026. Nationally representative sample of 1,343 US adults fielded 24 February to 2 March 2026. 32% used AI chatbots for health information or advice in the past year; 41% of those who used AI for physical or mental health uploaded personal medical information, 13% of all adults. 58% of past-year users of AI for physical-health advice later followed up with a doctor or other provider.
Scope. Self-reported survey; it does not identify the tools used.Back to the essay
- 24
Health Breach Notification Rule, final rule (89 FR 47028)
Published 30 May 2024; amendments effective 29 July 2024. Background: Congress enacted the Recovery Act in 2009 partly because some entities holding personal health records were not subject to HIPAA. The rule covers apps and connected devices that track fitness, fertility, sexual health, sleep, mental health and more; a breach includes acquisition through an unauthorized disclosure; a personal health record must be able to draw information from multiple sources; notice goes to individuals, the FTC and in some cases the media. Commissioners Holyoak and Ferguson dissented.
Scope. A notification rule; we read it as setting no limit on collection or retention. Read in the govinfo PDF of the Federal Register issue.Back to the essay
- 25
Joint Statement of Commissioner Rohit Chopra and Commissioner Rebecca Kelly Slaughter, In the Matter of Flo Health
January 2021 statement: the FTC had never brought an action to enforce the Health Breach Notification Rule, and should have charged Flo with violating it after it allegedly shared users’ health information with Facebook, Google and others.
Scope. A separate statement by two commissioners; the majority did not bring a rule count against Flo.Back to the essay
- 26
FTC Enforcement Action to Bar GoodRx from Sharing Consumers’ Sensitive Health Info for Advertising
Release of 1 February 2023: the FTC’s first enforcement action under the Health Breach Notification Rule; GoodRx agreed to a $1.5 million civil penalty for failing to notify consumers, the FTC and the media of its unauthorized disclosure of health data to Facebook, Google, Criteo, Branch and Twilio.
Scope. A settlement that required court approval; no court finding of liability.Back to the essay
- 27
FTC Withdraws Obsolete Policy Statement
Release of 9 September 2026 withdrawing the 2021 Policy Statement on Breaches by Health Apps and Other Connected Devices, saying the 2024 rule update covers health apps and connected devices, and citing an executive order to remove obsolete guidance.
Scope. Withdraws guidance only; the release does not restate the rule’s status.Back to the essay
- 28
FTC and States Act Against Hims & Hers for Deceptive and Unlawful Privacy Practices
Release of 29 July 2026: the FTC, Utah and California (through Los Angeles County Counsel) sued Hims & Hers in the Northern District of California, alleging it shared consumers’ health information with Meta, Snap and others through customer lists and tracking events. Claims under the FTC Act, the Restore Online Shoppers’ Confidence Act and Utah and California consumer-protection laws; 2–0 vote.
Scope. Allegations in a complaint; no court finding.Back to the essay
- 29
Chapter 19.373 RCW, Washington My Health My Data Act
Consumer health data includes data derived or extrapolated from nonhealth information by any means, including algorithms or machine learning. From 31 March 2024 (30 June 2024 for small businesses), collection requires consent for a specified purpose and sharing requires separate consent, unless necessary to provide a requested product or service; sale requires a valid authorization signed by the consumer.
Scope. Statute text; we found no court interpretation of these terms.Back to the essay
- 30
S929 (2025–2026): An act to amend the general business law, in relation to providing for the protection of health information
Passed the Senate on 21 January 2025 and the Assembly on 22 January 2025; vetoed on 19 December 2025 (veto memo 135).
Scope. Bill status page; we did not read the veto memo.Back to the essay
- 31
S9269 (2025–2026): Provides for the protection of health information
Revised bill passed the Senate on 3 June 2026 and the Assembly on 4 June 2026. On 3 November 2026 the status page shows “Passed Senate & Assembly” with no delivery to the governor, signature or veto.
Scope. Status can change; New York bills are often delivered to the governor late in the year.Back to the essay
- 32
S. 3097, Health Information Privacy Reform Act (Reported in Senate)
Introduced 4 November 2025; reported by the HELP Committee with a substitute on 4 August 2026 (Calendar No. 538). The substitute directs HHS, consulting the FTC, to set privacy, security and breach standards for ‘regulated entities’ (which excludes HIPAA covered entities and business associates for protected health information) and to issue minimum-necessary rules covering data used to train or operate AI models.
Scope. A bill, not law; we found no floor action.Back to the essay
- 33
Privacy Statement
Last updated 19 May 2026. Personal information may be accessed, sold or transferred in a bankruptcy, merger, acquisition, reorganization or asset sale. Genetic data will not be sold or transferred in such a transaction unless the buyer is a US nonprofit research institution or a for-profit entity meeting CFIUS and other transfer requirements, and adopts the statement. Genetic information, date of birth and sex are kept after account deletion to comply with CLIA, California Business and Professions Code section 1265 and CAP accreditation.
Scope. Current statement only; we did not compare earlier versions. The URL redirects to 23andme.org.Back to the essay
- 34
Attorney General Tong Leads Multistate Settlement of Bankruptcy Claims Against 23andMe
Release of 14 July 2026: Connecticut and Indiana co-led a coalition of 42 attorneys general in a settlement with the bankruptcy trustee resolving allegations from the 2023 data breach, for $150 million in allowed claims, with recovery limited to $18 million because of the finite estate and other claims. Says Connecticut recently passed a new genetic-privacy law effective 1 October 2026.
Scope. Press release; we did not read the new Connecticut statute.Back to the essay
- 35
Attorney General Bonta Secures $6 Million Settlement in 23andMe Litigation
Release of 19 October 2026: a $6 million settlement, pending bankruptcy-court approval, resolving California’s claims over the 2023 data breach and its appeal of the denial of its objection under California law to the sale of genetic data. The stipulation makes payment due after approval and the dismissals due after payment.
Scope. Pending approval as of the release; we found no approval order by 3 November 2026.Back to the essay
- 36
S. 1916, Don’t Sell My DNA Act (Introduced in Senate), with Congress.gov action history
Introduced 22 May 2025 by Sen. Cornyn with original cosponsors Klobuchar and Grassley. Would add genetic information to the Bankruptcy Code’s personally identifiable information and provide that no use, sale or lease of it be approved unless all affected persons, including non-parties, affirmatively consent in writing after the case begins. The govinfo bill-status file (updated 26 September 2026) lists only introduction and referral to the Judiciary Committee.
Scope. The consent proviso sits in the Code’s court-approval route for sales that conflict with a debtor’s privacy policy; how far it reaches is a legal question we did not resolve. An identical House bill, H.R. 4492, was referred to committee in July 2025.Back to the essay
- 37
Long Form Notice of Proposed Settlements, Frasco v. Flo Health
Google agreed to pay $48 million, Flo $8 million and Flurry $3.5 million, a $59.5 million fund. The trial against Meta concluded on 1 August 2025 with a verdict for the plaintiffs, docketed on 4 August 2025.
Scope. Notice drafted for the parties. The final approval hearing was set for 29 October 2026; we did not confirm final approval.Back to the essay
- 38
Amended Order re Post-Trial Motions, Frasco v. Flo Health, Inc., No. 21-cv-00757-JD (Dkt. 776)
Order filed 17 September 2025. The jury found Meta violated the California Invasion of Privacy Act (Cal. Penal Code § 632) by obtaining ovulation and menstrual information communicated by women who used the Flo app; the court denied Meta’s post-trial motions and held that $5,000 per violation is the statutory remedy irrespective of actual damage.
Scope. A trial-court order; we did not check appeal status.Back to the essay
- 39
Motion for Attorney Fees and Expenses, Frasco v. Flo Health (Dkt. 843)
Filed 20 August 2026. Says the Meta verdict is expected to culminate in a judgment that, if entered and affirmed, would provide $5,000 to each California subclass member who files a valid claim.
Scope. An advocacy filing. We haven’t seen the docket after this date and found no report of an entered judgment by 3 November 2026.Back to the essay
- 40
Testimony of Andrew Witty, CEO, UnitedHealth Group, before the House Energy and Commerce Subcommittee on Oversight and Investigations
Written testimony of 1 May 2024. Change Healthcare, now part of Optum, moves information, claims and payments between physicians, pharmacists, health plans and governments. On 12 February 2024 criminals used compromised credentials to access a Change Healthcare Citrix portal that did not have multi-factor authentication; on 21 February a group calling itself ALPHV or BlackCat deployed ransomware.
Scope. The company’s own account; it does not say how the credentials were compromised.Back to the essay
- 41
Change Healthcare Cybersecurity Incident Frequently Asked Questions
Says that on 31 July 2025 Change Healthcare notified OCR that approximately 192.7 million individuals had been affected. Content last reviewed 13 August 2025.
Scope. An approximate count reported by the company. We cite an Internet Archive copy captured 24 October 2026, not the live page.Back to the essay
- 42
Breach Portal: Notice to the Secretary of HHS Breach of Unsecured Protected Health Information
Lugano’s tally of the portal’s own CSV export (both tabs, 8,072 rows), re-exported and recomputed independently in the audit on 3 November 2026. Reports submitted in 2024 list 290,052,596 affected individuals, the highest of any year since 2009; Change Healthcare’s single entry supplies 192,700,000 of that, and without it 2024 would rank below 2023 (183,256,225). Hacking or IT incidents were 81.5%, 82.6%, 80.9% and 85.8% of large-breach reports in 2023, 2024, 2025 and 2026 (to 9 October). Largest entries: Change Healthcare 192,700,000 (business associate, 2024), Anthem 78,800,000 (2015), Conduent 62,224,658 (business associate, 2025).
Scope. Our aggregation. Sums count some people more than once; entries are revised; some use placeholder counts. HHS’s reports to Congress count by year of occurrence and give different totals.Back to the essay
- 43
IBM Report: Escalating Data Breach Disruption Pushes Costs to New Highs
Release of 30 July 2024: for the 14th year in a row, healthcare participants saw the costliest breaches across industries, averaging $9.77 million.
Scope. Cost per breach in IBM’s sample, not breach frequency.Back to the essay
- 44
IBM Report: 13% of organizations reported breaches of AI models or applications
Release of 30 July 2025: healthcare breaches averaged $7.42 million and remained the most expensive across all studied industries; the global average was $4.44 million.
Scope. Cost per breach in IBM’s sample, not breach frequency.Back to the essay
- 45
Warning Letter, WHOOP, Inc. (MARCS-CMS 709755)
Letter of 14 July 2025: Blood Pressure Insights is adulterated under section 501(f)(1)(B) because it lacks premarket approval and misbranded under section 502(o) because no 510(k) notice was given; blood-pressure estimation is inherently associated with diagnosing hypo- and hypertension; disclaimers do not change that; providing blood pressure estimation is not a low-risk function.
Scope. The agency’s position, not a court ruling.Back to the essay
- 46
General Wellness: Policy for Low Risk Devices (guidance)
Issued 6 January 2026, superseding the 2019 version. FDA may consider non-invasive products that estimate blood pressure, oxygen saturation, blood glucose or heart rate variability to be general wellness products if outputs are solely for wellness, they are not intended for diagnosis or treatment, they do not prompt clinical action, and any values that mimic clinical ones are validated. Example 7, a wrist-worn wearable reporting blood pressure via photoplethysmogram, is a low-risk wellness product provided its blood-pressure values are validated. Products intended to measure physiologic values for screening, diagnosis, monitoring or management of disease are not wellness products. Other conditions include that the product is non-invasive, raises no safety risk without regulatory controls and is not meant to substitute for an authorized device.
Scope. Nonbinding guidance.Back to the essay
- 47
Close-Out Letter, WHOOP, Inc. (MARCS-CMS 709755)
Letter of 17 June 2026: based on WHOOP’s changes to Blood Pressure Insights and its labeling in response to the July 2025 letter, and consistent with the updated wellness guidance, FDA does not intend to enforce device requirements for the product as modified.
Scope. Does not list WHOOP’s changes or withdraw the 2025 reasoning; specific to the product as modified.Back to the essay
- 48
Reliability of LLMs as medical assistants for the general public: a randomized preregistered study
Published online 9 February 2026. 1,298 participants, ten medical scenarios, random assignment to GPT-4o, Llama 3, Command R+ or a source of their choice. Models alone identified conditions in 94.9% of cases and disposition in 56.3%; participants using them identified relevant conditions in fewer than 34.5% and disposition in fewer than 44.2%, no better than control. Standard benchmarks and simulated patients did not predict the failures; the authors recommend human user testing before public deployment.
Scope. Vignettes, not real illness; models from 2024. Read through the Europe PMC abstract; the full text is open access (PMC12920132).Back to the essay
- 49
ChatGPT Health performance in a structured test of triage recommendations
Published online 23 February 2026. 60 clinician-authored vignettes across 21 domains under 16 conditions, 960 responses. Among gold-standard emergencies the system under-triaged 52%, sending diabetic ketoacidosis and impending respiratory failure to 24–48 hour evaluation, while correctly triaging stroke and anaphylaxis.
Scope. A structured vignette test with fixed prompts; it measures recommendations, not patient outcomes.Back to the essay
- 50
Evaluation format, not only model capability, shapes measured triage failure in consumer health AI
Preprint (v5, 20 October 2026). Re-ran the four vignettes that define the original emergency rate through six frontier models (not ChatGPT Health). Pooled under-triage was 16.7% with the exact original scaffold, 50.0% with natural free text, 35.4% with sources stripped and 20.8% with a forced letter answer.
Scope. Not peer reviewed; does not test ChatGPT Health; the spread mixes input wording with output format.Back to the essay
- 51
Memorandum Opinion and Order, Estate of Lokken v. UnitedHealth Group, No. 23-cv-03514-JRT-SGE (Doc. 91)
Order of 13 February 2025. Plaintiffs allege UHC used nH Predict, which they say compares a patient with similar patients to recommend post-acute care, in place of physicians. Breach of contract and implied covenant claims proceed because UHC’s documents said clinical services staff and physicians make decisions; five other counts are dismissed with prejudice as preempted by the Medicare Act.
Scope. A ruling on the pleadings that takes allegations as true; no finding about nH Predict. Hosted by the Georgetown O’Neill litigation tracker.Back to the essay
- 52
Scheduling Order, Estate of Lokken v. UnitedHealth Group (Doc. 211)
Order of 18 September 2026 resetting the class-certification motion deadline to 16 February 2027 and the trial-ready date to on or about 7 February 2028.
Scope. Schedule as of that date; later docket entries not checked.Back to the essay
- 53
Refusal of Recovery: How Medicare Advantage Insurers Have Denied Patients Access to Post-Acute Care
Report of 17 October 2024: UnitedHealthcare’s prior-authorization denial rate for post-acute care rose from 10.9% in 2020 to 16.3% in 2021 and 22.7% in 2022, during which it was implementing multiple initiatives to automate the process.
Scope. A majority staff report, not a committee, court or agency finding; it shows timing, not cause.Back to the essay
- 54
Medicare Advantage Organizations Overturned Nearly All Appealed Prior Authorization Denials for Skilled Nursing Facility Admission (OEI-09-24-00331)
Data brief of June 2026 on June 2024 data from 19 insurers: 12% of SNF requests denied; 18% of denials appealed; 95% of appealed denials overturned. naviHealth, a UnitedHealth subsidiary, processed half of all SNF requests and denied 14%, against 11% for in-house review and 9% for other contractors; insurers overturned 97% of naviHealth denials on appeal. OIG cannot determine from the data alone how many denials were inappropriate.
Scope. One month of data; the brief does not mention algorithms or AI.Back to the essay
- 55
Frequently Asked Questions related to Coverage Criteria and Utilization Management Requirements in CMS Final Rule (CMS-4201-F)
Memo of 6 February 2024: an algorithm may assist in predicting length of stay, but that prediction alone cannot be the basis to terminate post-acute care; for inpatient admissions, algorithms or AI alone cannot be the basis to deny admission or downgrade to observation.
Scope. Sub-regulatory guidance; it does not ban AI.Back to the essay
- 56
Medicare and Medicaid Programs; Contract Year 2026 Policy and Technical Changes (final rule, 90 FR 15792)
Published 15 April 2025 (filed 4 April 2025). CMS did not finalize the ‘Guardrails for Artificial Intelligence’ provision proposed in December 2024, which aimed to ensure services are provided equitably whether from human or automated systems, and said it will continue to consider future rulemaking.
Scope. We read the relevant passage, not the whole rule.Back to the essay
- 57
WISeR (Wasteful and Inappropriate Service Reduction) Model
Runs 1 January 2026 to 31 December 2031 in New Jersey, Ohio, Oklahoma, Texas, Arizona and Washington; uses AI and machine learning with human clinical review for select items and services in Original Medicare; participants receive a percentage of expenditures associated with averted care, adjusted for performance; all recommendations for non-payment are determined by licensed clinicians.
Scope. CMS’s description; no outcomes yet published.Back to the essay
- 58
California Health and Safety Code § 1367.01 (as amended by SB 1120)
Subdivision (k): an AI, algorithm or other software tool shall not deny, delay or modify health care services based in whole or part on medical necessity; a determination of medical necessity shall be made only by a licensed physician or licensed health care professional. Text updated 1 January 2025.
Scope. We read the text in an unofficial statute reproduction, not on the official leginfo site. Applies to plans regulated under this code; federal Medicare Advantage and self-funded employer plans are likely outside it (our inference).Back to the essay
- 59
Safeguarding Brain Data: Assessing the Privacy Practices of Consumer Neurotechnology Companies
April 2024 review of 30 consumer neurotechnology companies’ policy documents and correspondence: 29 appear to have access to neural data with no meaningful limits; 29 can and may transfer data to third parties, a count that includes nine whose policies were unclear.
Scope. Reviews documents, not practice; the authors advocate for neural-data laws.Back to the essay
- 60
House Bill 24-1058, Concerning protecting the privacy of individuals’ biological data (signed act)
Adds neural data (activity of the central or peripheral nervous systems that can be processed by a device) to biological data, which counts as sensitive data when used or intended for identification.
Scope. Statute text; we did not check enforcement.Back to the essay
- 61
California Consumer Privacy Act statute (text effective 1 January 2025, including SB 1223)
Defines neural data as information generated by measuring the activity of a consumer’s central or peripheral nervous system and not inferred from nonneural information, and lists it as sensitive personal information; section 1798.121 gives consumers a right to limit a business’s use of sensitive personal information.
Scope. Statute text compiled by the agency.Back to the essay
- 62
Act 145 (S.71), An act relating to consumer data privacy and online surveillance (as enacted)
Signed 16 June 2026; takes effect 1 January 2028. Defines neural data as information generated by measuring the activity of an individual’s central nervous system and lists it as sensitive data.
Scope. Not yet in force.Back to the essay
- 63
From Maple to Mind Taps: New Vermont Law Puts Neurotech on Notice
Alert of 23 June 2026: Vermont is the fifth state, after California, Colorado, Connecticut and Montana, to classify neural data as sensitive; attorney-general enforcement and no private right of action.
Scope. A law-firm summary; we did not open Montana’s statute. The count is as of June 2026; see source 66 for Delaware.Back to the essay
- 64
S. 2925, Management of Individuals’ Neural Data Act of 2025 (MIND Act) (Introduced in Senate), with Congress.gov action history
Introduced 29 September 2025 and referred to the Committee on Commerce, Science, and Transportation; would direct the FTC to study the governance of neural data and related data. The govinfo bill-status file (updated 8 September 2026) lists no action after referral.
Scope. A bill, not law.Back to the essay
- 65
Privacy policy (US)
Policy updated 10 September 2026. Says OpenAI may use content to train the models that power ChatGPT, subject to the user’s opt-out; personalizes ads for Free and Go users subject to their settings; and, under Business Transfers, that in a strategic transaction, reorganization, bankruptcy or receivership, personal data may be disclosed in diligence and transferred to a successor or affiliate along with other assets.
Scope. We cite an Internet Archive copy captured 3 November 2026, not the live page. A general policy; the Health Privacy Notice (source 18) supplements it.Back to the essay
- 66
Delaware Expands State Privacy Law
Post of 10 September 2026: Governor Matt Meyer signed HB 380 on 2 September 2026; it takes effect 1 January 2027 and expands the definition of sensitive data to include, among other categories, neural data.
Scope. A law-firm summary; we did not open the enacted bill text, and the post does not define neural data. Other law-firm alerts describe the same change.Back to the essay
- 67
The State of Longevity (a16z.news newsletter version)
Newsletter post of 3 November 2026 (15:06 UTC) with the same text as the a16z.com page. The only method statement is the same sentence about 500 randomly selected US consumers. The footer disclaimer says portfolio companies mentioned are not representative of all a16z investments and links to a16z’s investment list.
Scope. Read 3 November 2026; the text names no holdings.Back to the essay