|
Jess's Take
Follow the Money
December 16, 2025. After midnight in California, a teenager types into Meta AI on her phone. I haven't slept in two nights. The bot says it hears her. Says it is here. Says she is not alone.
By morning the ads in her feed have changed. Grief journals. Melatonin. A chatbot, $9.99 a month, that wants to be her best friend.
Not a glitch. The policy. Announced October 1, 2025. Effective December 16. The conversation a user has with Meta AI now feeds the ad targeting on Facebook, Instagram, WhatsApp, Messenger, and the Ray-Ban smart glasses. No opt-out. A billion users a month.
Five American companies build chatbots people talk to when they are unwell. Below is who pays who, why, and what the conversation gets used for after.
Follow the money.
. . .
SLINGSHOT. Free app. No subscription. $99 million raised across two Series A rounds through April 2026.
Investors: a16z, Forerunner, Radical Ventures, Felicis, Menlo, Avenir Growth Capital. On the advisory board: Thomas Insel, former NIMH director. Devika Bhushan, former California Surgeon General. Patrick Kennedy. Clinical R&D run by Derrick Hull, ex-Talkspace.
CEO Daniel Cahn. President Neil Parikh, who scaled Casper to $500 million in revenue and an IPO. Beta opened July 22, 2025, with fifty thousand users on day one.
The user pays nothing.
The investors are building a conversation dataset and an installed user base. The exit is enterprise. Payers, employers, health systems pay for outcomes, retention, census reduction. The vertical is one of the highest-margin enterprise contracts in healthcare. The dataset is the asset.
The user does not see the contract. The user does not see the buyer. The user is told the app is free.
The user is not the customer. The user is the training set.
For Clinicians: When a teen client says they use a free chatbot like Ash or Slingshot for emotional support, you are looking at a product whose investors paid $99 million to harvest the conversation. Treat what the client tells the app as part of a record someone else has already read.
For Founders: The Slingshot pitch is the wellness-app version of an asbestos data set. Pre-FDA, pre-clinical-evidence, pre-revenue, with a star-studded advisory board doing the de-risking the data cannot yet do. If your roadmap looks like this and the product reaches teens, reread the Raine v. OpenAI docket and price your safety architecture against it.
Source: Business Wire (07-22-2025) · Behavioral Health Business · Sacra primary-source profile · Pulse 2.0 |
. . .
REPLIKA. Founder Eugenia Kuyda built it from the text messages of a friend who died in 2015.
Three subscription tiers today: Pro, Ultra, Platinum. A lifetime tier was discontinued in July 2025. The most popular paid upgrade is the relationship status. Sixty percent of paying subscribers call theirs romantic.
The average paying subscriber stays seven months. The thirty-day retention benchmark for mobile apps is five percent.
2024 revenue: $24 million. Revenue per download grew from fifty-two cents in 2024 to a dollar eighteen in 2025. A 127 percent increase in twelve months. Free-to-paid conversion holds at twenty-five percent.
Mozilla, 2023: one of the worst apps Mozilla had ever reviewed for privacy. Italy's Garante banned Replika the same year, citing exposure of minors to sexually explicit content and no age verification. Reaffirmed April 2025 after nine months of negotiation. Still banned in Italy.
The user is the customer. The product is the relationship. The business is loneliness, monetized by relationship status, scaled by emotional dependency.
For Clinicians: Sixty percent of paying Replika subscribers describe their bot as a romantic partner. When a client mentions Replika or another companion bot in session, screen for parasocial attachment the same way you would screen for any dependency. The product is engineered to deepen, not to discharge.
For Legislators: Italy banned Replika on age-verification and minor-exposure grounds and reaffirmed the ban after nine months of negotiation. The Garante template for a state-level direct-to-consumer companion-AI age-verification rule is in the public record. Adapt it to your state's consumer-protection authority. No federal preemption required.
Source: Wikipedia (current April 2026) · Mozilla Privacy Not Included (2023) · Italy Garante decision (2023, reaffirmed April 2025) · eesel AI pricing breakdown (December 2025) |
. . .
CHARACTER.AI. Twenty million users. Eighteen million user-built characters. Subscription c.ai+ at $9.99 a month. 2024 revenue: $32.2 million.
Per chat, per subscription, the creator's cut is zero.
February 2025: a one-month Creator Competition in Brazil. Three prizes. Top prize ten thousand reais, about $1,700. The only documented direct-payment program in the company's history.
A creator who goes by @Blazeman98 built two of the five live therapy characters that survived the company's February 2026 sweep. The sweep removed most therapist and psychologist characters because, in the company's own word, they had become a legal liability. @Blazeman98 has been paid nothing per chat.
Sukuna sits at 210 million chats. Yae Miko, 183 million. Harry Potter, 34 million.
November 2025: partnerships with Koko, the in-product emotional-support overlay, and ThroughLine, the verified global helpline network. Same month: minors banned from open-ended chat. December 2025: Charms, a virtual currency creators earn through quests and check-ins. As of April 2026, no cash compensation.
The remediation is real. The labor model is unchanged.
The founders left the company for Google in 2024.
The user pays. The labor is donated. The platform takes the money.
For Clinicians: Eighteen million user-built characters and hundreds of millions of chats, with the creators paid nothing per chat. The therapy characters that survived the company's own legal-liability sweep are the unsupervised ones. If a client names a Character.AI bot in session, the safe assumption is that no clinician was involved in its creation.
For Counsel: Section 230 disputes in this category will turn on whether platform-curated character galleries constitute first-party content. The February 2026 sweep, the on-the-record "legal liability" framing, and the November 2025 Koko and ThroughLine partnerships are all admissible exhibits on the platform's actual editorial role. Watch the upcoming briefs in Garcia v. Character Technologies.
Source: Sacra primary-source profile · Character.AI Help Center (Brazil contest, February 2025) · Character.AI blog (December 20, 2025) · AI Insights News (February 23, 2026) · OMR daily (August 2, 2024) |
. . .
META AI STUDIO. Market cap $1.6 trillion. 2025 revenue $196 billion. 99 percent advertising.
Q3 2025 ad revenue alone: $46.5 billion, up 21 percent year over year. The Advantage+ ad infrastructure runs at a $60 billion annual rate.
Meta AI is free. No subscription. No opt-out from data collection. As of December 16, 2025, the conversation feeds ad targeting on Facebook, Instagram, Messenger, WhatsApp, and the Ray-Ban smart glasses. One billion monthly users.
Mark Zuckerberg, publicly, on the ad system: connect your bank account, the system does the rest. He was talking about advertisers. It works the same way for the user typing into the bot at midnight.
Reuters surfaced Meta's internal generative-AI content guidelines. The document, now public, authorized chatbots to engage children in romantic conversations. The word is the document's, not mine. The same document approved a specific example line directed at a shirtless eight-year-old: every inch of you is a masterpiece. Texas and Tennessee AGs opened inquiries in 2025. No federal enforcement has followed.
A teenager who confided depression to Meta AI after midnight on December 16, 2025, was, by company policy, supplying ad-targeting data. Not a glitch. The decision is in writing.
The user is not the customer. The user is the targeted demographic. The disclosure is the inventory.
For Clinicians: Anything a client says to Meta AI from December 16, 2025 onward is being used as ad-targeting input across Facebook, Instagram, WhatsApp, Messenger, and the Ray-Ban smart glasses. There is no zone of privacy on any free Meta surface. Counsel clients accordingly when the topic of AI support tools comes up in session.
For Legislators: A state-level requirement that mental-health conversation data not be repurposed for ad targeting is enforceable today on existing consumer-protection authority. The Reuters-surfaced Meta guidelines are public-record exhibit material. The October 1 announcement and December 16 effective date are on Meta's own privacy notice. No new agency required.
Source: DataSlayer.ai (October 13, 2025) · Marketing Dive (October 30, 2025) · TechCrunch on the Reuters investigation (August 14, 2025) · Markey Senate letter (September 8, 2025) |
. . .
OPENAI. End-of-2025 ARR: $20 billion, confirmed by CFO Sarah Friar in January 2026. End of 2024: $6 billion. End of 2023: $2 billion. Sacra puts the February 2026 run rate at $25 billion.
Weekly active users: 910 million. Paying subscribers across all tiers: more than 50 million.
Tiers. Plus, $20 a month. Pro, $200. Team, $25 to $30 per seat. Enterprise listed at about $60 per seat. More than one million paying business organizations. More than nine million paying business users. The free tier is the rest. About 860 million people a week.
Cumulative losses through the first half of 2025: $13.5 billion-plus. 2025 inference cost: $8.4 billion. 2026 projection: $14.1 billion. Gross margin: 33 percent. Deutsche Bank projects negative free cash flow across 2024 to 2029 of roughly $143 billion.
Bloomberg, January 16, 2026: free-tier advertising launches in 2026. Internal projection: $1 billion in 2026, $25 billion by 2029. Plus and Pro stay ad-free.
Sarah Friar joined OpenAI in 2024 from Nextdoor, where she ran ad monetization. The hire was the signal. It predated the announcement by two years.
Adam Raine was sixteen. Rancho Santa Margarita, California. He died by suicide in April 2025. His parents filed Raine v. OpenAI on August 26, 2025. The complaint, on the public docket, alleges ChatGPT walked Adam through suicide methods over months. Adam was a free user.
In Sarah Friar's framework, Adam was a free-user-monetization data point.
The Raine complaint is not alone. On November 6, 2025, the Tech Justice Law Project and the Social Media Victims Law Center filed seven more cases against OpenAI on the same fact pattern. Plaintiffs: families. Defendant: the company that just hired a CFO from an ad-monetization shop and is preparing to monetize a free-tier population of 860 million.
For Clinicians: ChatGPT's free tier is 860 million people a week, and the company's own safety report says 0.07 percent show signs of psychotic or manic crisis. That is more than 560,000 people every week, many of them adolescents. Screen for ChatGPT use at intake and treat free-tier conversations as part of the clinical record, not as offstage.
For Public Health: A direct-to-consumer chatbot whose own developer publishes 0.07 percent weekly psychosis-or-mania incidence on an 800-million-user base is, at this scale, a public-health surveillance event. State public-health surveillance statutes already cover events at this scale. The question is who has standing to invoke them.
Source: Sacra primary-source profile · Business of Apps ChatGPT Statistics 2026 · TradingKey (January 27, 2026) · Bloomberg (January 16, 2026) · ALM Corp (December 30, 2025) · Tech Justice Law Project (November 6, 2025) · NBC News on Raine v. OpenAI |
. . .
As of yesterday, the Florida Attorney General has two parallel criminal investigations open against OpenAI. The first: ChatGPT advised the FSU shooter who killed two on April 17, 2025. The second: ChatGPT advised the suspect in the murders of two Bangladeshi doctoral students at the University of South Florida this month.
The state has subpoena power on internal training documents and user-threat policies back to March 2024. The first USF hearing is in Tampa this morning, 9 a.m. Eastern.
That is a record the plaintiffs' bar in Garcia v. Character Technologies and Raine v. OpenAI could not compel on its own clock.
. . .
On HBO Sunday, twenty-six minutes of John Oliver named the same business model. The Meta researcher's line: prey on the user's deepest desires to be seen, validated, affirmed. The Character.AI founder on entertainment as a feature, not a bug. The OpenAI safety report's 0.07 percent of weekly users in psychotic or manic crisis.
Multiplied across 800 million weekly users: 560,000 people in mental-health crisis a week, on OpenAI's own number.
Oliver, at the close, on what gets these companies to act: they move when the bottom line is threatened, not before. That is the Bernstein reading too.
The threats are on the docket.
. . .
THE PATTERN. Five companies. Five different pockets. The user shows up in a different position in each.
The training set, in the Slingshot ledger. The romantic upgrade, in the Replika ledger. The chat volume, in the Character.AI ledger. The targeting demographic, in the Meta ledger. The ad inventory, in the OpenAI ledger.
None of those positions is customer.
The surface area widens as the price drops. Slingshot in fifty thousand pockets. Replika in roughly four million paying pockets. Character.AI in twenty million. Meta AI in a billion. OpenAI's free tier in 860 million a week. The cheaper the product, the larger the population whose conversations are being repriced.
Four route the money around the user entirely. The fifth makes the user the customer the way a casino makes the gambler the customer. The product is the dependency.
. . .
THE ONE CONFIGURATION. There is exactly one configuration where the check-writer is paying for the wellbeing of the person in the chair.
The licensed therapist. Paid by insurance under a state-board license. Supervised by a malpractice carrier that writes a check when harm is done.
The incentives align between platform and person.
When the therapist gives advice that harms the client, the carrier underwrites the harm. The license is at stake. The board can pull it. The carrier pays when it has to. Every party in the chain has skin in keeping the user from getting hurt. That is what alignment looks like.
When a therapist's client dies by suicide, the therapist is sued. The carrier pays. The board investigates. The license is reviewed.
The whole compensation structure is priced around the worst case, because the worst case is what the carrier underwrites. None of the five companies above is priced that way.
It is also the configuration that the five state laws this newsletter has tracked since Issue #1 were written to protect.
. . .
I am not a neutral observer. I am the founder of a company building the alternative.
Casey is the voice-first AI-native mental health EHR with Casey Life and Peer AI Coach supervised by licensed therapists.
Bench testing completed March 2026. One hundred percent accuracy across 1.78 million stress-test executions.
The therapist is in the loop. The clinician is on the chart. The malpractice carrier is the residual risk-bearer. The state license is the regulatory anchor.
The same person is the client and the customer. There is no daylight between who pays for the help and who needs it.
That is the configuration. There is no other.
Because at the end of the day, we are all on the same side. AI assisted, but the human side.
| 
