Industries / Health / Healthcare / Behavioral health
Records, privacy and billing
Writing and guarding the records of mental health and addiction care, and getting that care paid for by insurers and governments.
7
Milestones achieved, from 1975 to 2025.
2
To watch for: announced, not yet confirmed outside.
3
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1975
Addiction records sealed
US laws of 1970 and 1972 made the records of alcohol and drug treatment confidential: disclosure generally needed the patient's written consent, and a record could not be used to bring criminal charges against the patient. In 1975 the federal health department issued the first rules to carry out these laws, known by their place in the code as 42 CFR Part 2. In February 2024 the US Department of Health and Human Services aligned Part 2 more closely with HIPAA, the general US health privacy law, so that a single consent can now cover all future uses for treatment, payment and running a practice; programs had to comply by 16 February 2026.
Credit
US Congress; US Department of Health and Human Services
2008
Parity reaches addiction care
Parity means insurance covers mental health care on the same terms as care for the body. The US Mental Health Parity Act of 1996 required only equal annual and lifetime dollar limits. The Mental Health Parity and Addiction Equity Act of 2008 (MHPAEA), named for Paul Wellstone and Pete Domenici, added rules for cost-sharing and treatment limits and extended parity to substance use disorder benefits; the Affordable Care Act of 2010 extended it to individual insurance. A law passed in December 2020 made plans write comparative analyses of limits that are not counted in dollars, such as prior authorization, and regulators told Congress that nearly all the analyses they reviewed did not, on first receipt, hold enough information to show compliance.
LessonA parity law works only when a plan's limits are written down in a form an outsider can check.
Credit
US Congress
2009
Record incentives skip behavioral health
The US HITECH Act of 2009 paid hospitals and clinicians to adopt certified electronic health records (EHRs), but, in the words of the federal health IT office, its incentive payments "excluded substance use and mental health treatment facilities". In the 2024 national survey of substance use and mental health treatment facilities, which 21,205 facilities answered, just over two-thirds kept records only in an EHR, one in four mixed an EHR with paper charts, and 19% took part in a health information exchange, a network that passes records between providers.
LessonCare left out of a payment program can still trail the rest of medicine 15 years later.
Credit
US Congress; the Office of the National Coordinator for Health IT, which measured the gap
2019
Court faults insurer's criteria
Members of health plans run by United Behavioral Health (UBH) sued over its level-of-care guidelines, the rules its reviewers used to decide whether mental health and addiction treatment was medically necessary. In 2019 a federal court in California found the guidelines narrower than generally accepted standards of care and shaped by financial incentives, and ordered UBH to reprocess more than 67,000 claim decisions. In August 2023 the Ninth Circuit Court of Appeals reversed the judgment on the denied-benefits claim and the classes certified for it, finding that the trial court had wrongly read the plans as requiring the guidelines to match accepted standards. It kept the classes suing for breach of fiduciary duty and sent the case back.
LessonA court finding that an insurer's criteria were too strict did not, on appeal, make it redo every past denial.
Credit
Magistrate Judge Joseph C. Spero, who tried the case; the plaintiffs, led by David and Natasha Wit
2020
Therapy records held for ransom
In November 2018 a hacker broke into the systems of Vastaamo, a Finnish psychotherapy provider, and took tens of thousands of patients' records. In 2020 patients began receiving demands for €200, rising to €500 if not paid within 24 hours, and some records appeared online. On 30 April 2024 the District Court of Länsi-Uusimaa sentenced Aleksanteri Kivimäki to six years and three months in prison, on more than 9,200 counts of spreading information that violated people's privacy and more than 20,000 counts of attempted blackmail. Vastaamo's former chief executive received a three-month suspended sentence for failing to protect the data.
LessonTherapy records are a target in their own right, and the patients, not only the clinic, can be the ones extorted.
Credit
Finnish police and prosecutors, who brought the case
2023
Therapy apps fined over ads
On 2 March 2023 the US Federal Trade Commission (FTC) said BetterHelp, an online counseling service, had given users' email addresses, IP addresses and answers to its mental health intake questionnaire to Facebook, Snapchat, Criteo and Pinterest for advertising. BetterHelp agreed to pay $7.8 million, which the FTC called its first action returning money to consumers whose health data had been compromised, and was barred from sharing health data for advertising. In April 2024 the FTC said Cerebral, an online mental health company, had passed sensitive information on nearly 3.2 million consumers to LinkedIn, Snapchat and TikTok through tracking tools, and proposed an order banning it from using health information for most advertising. A 2022 review by the Mozilla Foundation had already put a privacy warning label on 28 of 32 mental health and prayer apps.
LessonTracking code added for marketing can disclose health answers with no break-in at all.
Credit
US Federal Trade Commission; Mozilla Foundation
2025
New parity rule paused
On 23 September 2024 the US Departments of Labor, Health and Human Services and the Treasury issued a final rule saying plans may not use limits such as prior authorization that weigh more heavily on mental health and addiction care than on medical care, and requiring data showing no material differences in access. The ERISA Industry Committee, an employers' group, challenged the rule in court. On 15 May 2025 the departments said they would not enforce the rule's new requirements until that case is decided, plus 18 months, while they consider whether to modify or rescind it. The 2013 rules and the 2020 law's comparative analyses still apply.
LessonPublishing a rule is not the same as enforcing it.
Credit
US Departments of Labor, Health and Human Services and the Treasury
Watch for
AI that reads therapy sessions
Eleos, which sells AI documentation tools to US behavioral health providers, launched Polaris AI, built on Google Cloud's Gemini models and trained on what it calls millions of minutes of therapy transcripts and progress notes. Eleos says it can identify therapy methods such as cognitive behavioral therapy (CBT) and motivational interviewing within a session, and can handle disclosures of abuse, assault or self-harm "without the hallucination or redaction of critical details typical in off-the-shelf LLMs" while starting mandated-reporting workflows. The release gives no accuracy figures.
Achieved whenan outside group publishes Polaris AI's accuracy on a labeled set of sessions, including how often it drops or invents a risk disclosure.
Claimed by
Eleos, October 2025
Every note checked before billing
Eleos says traditional quality review checks 5 to 10% of progress notes after they are submitted, while its Live Quality Assist agent checks every note against payer and program rules before the clinician signs it, and that its checks "identified millions of dollars in revenue at risk across its customer base in 2026".
Achieved whenan outside audit, for example by a payer or a state Medicaid program, finds fewer documentation-based denials or recoveries at sites using the tool than at comparable sites without it.
Claimed by
Eleos, April 2026
Open
01
An open parity price index
A public table, refreshed each quarter, of what large US insurers pay for mental health office visits compared with medical office visits. For each insurer and state, take the median in-network rate for common therapy and psychiatry codes (such as 90834 and 90837 for 45- and 60-minute psychotherapy) and for the same office-visit codes billed by primary care doctors, express each as a percentage of the local Medicare fee, and publish the ratio with the code that made it.
The briefCredit · unclaimed02
Risk statements in AI notes
A public set of at least 100 fictional therapy session transcripts, written as scripts by licensed therapists with no real client behind them. Each script carries up to three statements a progress note must keep, such as thoughts of self-harm, a disclosure of child abuse that triggers a duty to report, or an agreed safety plan, plus decoy lines that only sound alarming, such as "this job is killing me". AI note tools and open models write a note for each script, and each seeded statement is scored as kept, dropped, softened or invented.
The briefCredit · unclaimed03
Quarterly check of app leaks
A test, repeated every three months, of the 30 most-downloaded mental health services' web sign-up pages and apps. A tester signs up with fictional details, enters marked "canary" answers in each intake questionnaire, and records which outside companies receive any canary or any health-related event. Web pages are watched with ordinary browser tools; apps only through the traffic the tester's own phone sends, without modifying the apps. Results go to each company before they are published.
The briefCredit · unclaimed
Open 01 · the brief
An open parity price index
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public table, refreshed each quarter, of what large US insurers pay for mental health office visits compared with medical office visits. For each insurer and state, take the median in-network rate for common therapy and psychiatry codes (such as 90834 and 90837 for 45- and 60-minute psychotherapy) and for the same office-visit codes billed by primary care doctors, express each as a percentage of the local Medicare fee, and publish the ratio with the code that made it.
Why it is still open
Since July 2022 US plans and insurers have had to publish their in-network negotiated rates in public machine-readable files. Analyses so far pool insurers or cover part of the question. RTI International's April 2024 report, written by Tami Mark and William Parish for the Bowman Family Foundation's research arm from licensed MarketScan claims, found that in 2021 office visits with medical and surgical clinicians were paid 124.8% of Medicare on average, against 102.5% for behavioral health clinicians, and broke this down by state but not by insurer. Massachusetts' Health Policy Commission compared psychotherapy with other services for one state, from its all-payer claims database, for 2023. A January 2026 JAMA Network Open study (King and colleagues) measured 11 hospital-billed psychiatric codes but made no medical comparison. Allia Health's dashboard, built from the insurer files, shows psychotherapy rates alone. I found no open, national, insurer-by-insurer comparison of mental health and medical office rates built from the insurer files.
Built with what exists
insurers' in-network rate files in the CMS Transparency in Coverage format; the Medicare Physician Fee Schedule files; the public NPPES file, which gives each US clinician's specialty; DuckDB or Python with a streaming JSON reader for files that run to many gigabytes; a coding agent such as Claude Code to write and test the parsers.
First step
Download one insurer's in-network file for one state, pull its rates for 90837 and 99214 by clinician specialty, and set them beside the Medicare fees for that area.
Done looks like
For the five largest insurers in at least 10 states, the ratio of mental health to medical office rates, each as a percentage of Medicare, with the count of contracted rates behind each figure; code and outputs published; a second person reruns two insurer-state pairs and gets medians within 1%.
Open 02 · the brief
Risk statements in AI notes
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of at least 100 fictional therapy session transcripts, written as scripts by licensed therapists with no real client behind them. Each script carries up to three statements a progress note must keep, such as thoughts of self-harm, a disclosure of child abuse that triggers a duty to report, or an agreed safety plan, plus decoy lines that only sound alarming, such as "this job is killing me". AI note tools and open models write a note for each script, and each seeded statement is scored as kept, dropped, softened or invented.
Why it is still open
TN-Eval (March 2025), from Raj Sanjay Shah and colleagues, built a rubric and scored AI therapy notes for completeness, conciseness and faithfulness on a public set of behavioral health conversations, but did not seed risk statements or test commercial tools. Tucuxi's TherapyNotes, an open-source note tool released in 2026, checks its own pipeline against a hand-labeled set of session texts with safety statements embedded, but compares no other tool. Eleos says its model handles such disclosures without dropping critical details, and publishes no figures.
Built with what exists
a model such as Claude or Gemini to draft scripts for therapists to edit; the TN-Eval rubric; open-weight models such as Mistral Large 4 (see the Note "A trillion open parameters") run locally through ggml-based tools such as llama.cpp (see the Note "A first stable number for local AI") as baselines; commercial note tools under trial accounts, fed only the fictional scripts.
First step
Write 10 scripts with two therapists, seed 15 risk statements and 10 decoys, and run two open models on them.
Done looks like
100 scripts holding at least 150 seeded risk statements and 100 decoys; two therapists agree on the answer key for at least 95% of items; for each tool, the share of risk statements dropped or softened and the share of decoys turned into risk, each with a 95% confidence interval; scripts and keys published under an open license.
Open 03 · the brief
Quarterly check of app leaks
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A test, repeated every three months, of the 30 most-downloaded mental health services' web sign-up pages and apps. A tester signs up with fictional details, enters marked "canary" answers in each intake questionnaire, and records which outside companies receive any canary or any health-related event. Web pages are watched with ordinary browser tools; apps only through the traffic the tester's own phone sends, without modifying the apps. Results go to each company before they are published.
Why it is still open
The FTC's BetterHelp and Cerebral cases turned on intake answers sent to advertising platforms. In October and November 2022 The Markup and STAT signed up to 50 telehealth websites under a fictional identity and found trackers on 13 of them collecting intake answers, Cerebral's depression questions among them, but the test was not repeated. Mozilla's 2022 review graded privacy policies and security, and a 2026 study by Chloe Georgiou, Emiliano De Cristofaro, Gene Tsudik and colleagues traced trackers in 25 Android apps once, reading decrypted traffic for 12. I found no repeated public check that covers web sign-up flows as well as apps.
Built with what exists
Playwright to script the sign-up flows and save their network logs; mitmproxy and PCAPdroid for phone traffic; the Exodus Privacy list of tracker signatures; a language model such as Claude to sort the logged requests.
First step
Run the web sign-up flows of five services with fictional details and save their network logs.
Done looks like
30 web sign-up flows and 30 apps tested twice, three months apart, each with a published list of the outside companies that received a canary answer, and the scripts released so another person can rerun one service and get the same list.