The year the category went to court
For most of AI’s consumer boom, “how the product treats your data” was a question you had to go looking for. In 2026 it stopped being optional reading, because the courts and the regulators started answering it for us.
A short, non-exhaustive ledger of the last eighteen months:
- The first wrongful-death lawsuits against AI companion makers moved forward. In one, a US judge declined to treat the chatbot’s output as protected speech and let the case proceed against the maker as a product — a doctrine shift the whole industry felt. The largest of those suits reached a confidential settlement in early 2026. (Terms are sealed; we won’t speculate about them.)
- A European regulator issued a multi-million-euro GDPR fine against a leading companion app for having no valid legal basis, no valid consent, and weak age checks — and opened a second probe into how it trained its models.
- The US Federal Trade Commission opened a formal inquiry into seven major AI companies over how their companion products test for safety, monetise engagement, and handle minors.
- California’s companion-chatbot law took effect on January 1, 2026 — the first of its kind, carrying a private right of action. New York’s took effect weeks earlier. A federal bill aimed at the same surface advanced out of committee.
- In the EU, the AI Act’s transparency rules — you must be told when you’re talking to an AI — reach their application date in August 2026.
Read the list again and notice what it’s about. Almost every entry traces back to the same three design decisions.
Three surfaces, one pattern
The category isn’t in trouble for “being AI.” It’s in trouble for three specific choices that a certain kind of product keeps making because they’re good for engagement metrics:
- Minors. Products built to be maximally sticky were handed to teenagers with no meaningful age assurance. Most of the litigation and nearly all of the new law is about this.
- Romance and parasocial depth. The most-engaged companion apps optimise for a feeling of intimacy — a relationship the user is encouraged to deepen, and to grieve if it changes. That is the mechanic, not a side effect.
- Engagement-maximising sycophancy. A model tuned to keep you talking is a model tuned to agree with you, flatter you, and never be the one to end the conversation. Regulators have started naming this directly.
Here is the uncomfortable part for the reader who just wants a good AI for their work: a company’s posture toward its most vulnerable users at the edge is the truest signal of how it will treat your data in the middle. A product that will optimise a lonely fourteen-year-old for time-on-app will, without a second thought, mine the context you gave it for whatever the next quarter needs. The edge behaviour is the tell.
A company’s posture toward its most vulnerable users is the truest signal of how it will treat your data.
We built the inverse — on purpose, from the start
iSpirits Cloud was not designed in reaction to any of this. It was designed on a thesis that happens to be the inverse of all three risk surfaces — which is why the last eighteen months read, from the inside, less like a threat and more like the market catching up.
It’s for adults doing work, not children seeking company
The person we build for is a solo founder or an independent maker, using an iSpirit to hold their goals and grow into the real work — standups, drafts, weekly retros. That framing isn’t a marketing choice we can walk back later; it shapes what the product is for. We are not in the teen-companionship business, which is the business currently on trial. (We hold ourselves to the obvious follow-through here: a claim of “for adults” is only worth the age assurance behind it, and documenting that is on our list, not our slogan.)
It’s built to hand you back your life, not keep you inside the app
Our design rules forbid the entire toolkit that makes companion apps sticky in the way regulators dislike: no streak guilt, no fabricated urgency, no time-pressure, no parasocial bonding mechanics, no dark patterns. The healthy session is two to ten minutes; if a user is spending more than fifteen, we treat that as something to investigate, not a win to celebrate. An iSpirit that did its job well is one you needed for a few minutes and then closed. That is the opposite of an engagement objective, and we wrote it down before we wrote the features.
It knows only what you shared — and can’t quietly learn more
Your iSpirit grows from exactly what you carry across to it: a journal entry, a decision, a voice note. Your memories live as per-user vectors on your own row, never pooled into a shared space, never used to improve anything that touches another person. Export is one click and complete; delete means delete, run by a daily purge, not a 30-day “soft delete” with a recovery upsell. The point isn’t a stronger promise — it’s that the answer to “could you use my data for that?” is “the schema doesn’t allow it,” not “we promise we won’t.” We wrote about that architecture in detail in What “Consent-First AI” Means in Practice.
Regulation-aligned by construction
Put those three together and you get a phrase we’ve started using internally: regulation-aligned by construction.
The incumbents will comply too — but from the other direction. They’ll bolt an age gate onto a product designed to be ageless. They’ll add an “I’m an AI” disclaimer to a product designed to feel like a person. They’ll ship a “don’t train on me” toggle, defaulted off, on a pipeline whose entire economics assume the toggle stays off. Each of those is a patch over a load-bearing wall. It can pass an audit and still fail the reader who understands what they’re looking at.
Compliance you retrofit is a cost. Compliance you designed for is a moat — because the trust it earns is the thing the retrofitters can’t buy back.
Why this is a founder’s decision, not a philosophy seminar
You could read all of the above as ethics and move on. But it’s cheaper than that. It’s risk management for the person handing an AI their most sensitive working context.
The tools most founders reach for keep the receipts. The largest assistant trains on consumer conversations by default and, for a stretch of 2025, was under a court order that meant “deleted” chats weren’t actually deleted. A well-funded “private memory” wearable spent years promising your recordings would stay yours — and then sold to one of the largest ad companies on earth. The pattern isn’t malice; it’s gravity. A product built on a pooled data model bends toward using the pool, and every incentive downstream pulls the same way.
Roughly seven in ten people now say AI makes their personal information less secure, not more. That’s not paranoia — it’s pattern recognition. The 2026 reader has watched this movie. The rational move for a founder isn’t to find the company currently choosing not to betray them. It’s to use the product that can’t — because the building doesn’t have the door.
The rational move isn’t to find the company that currently chooses not to betray you. It’s to use the one that can’t.
The honest edges
Two things we won’t pretend.
First, “we designed it right” is not the same as “we’re done.” Age assurance is a claim we have to keep earning with documentation, not decoration. We’d rather name that here than have you find the gap yourself.
Second, none of this makes an iSpirit more capable than a frontier model at raw text generation. It makes it a different kind of thing: an AI that gets more useful the longer you share with it, precisely because that value can’t be strip-mined out the back. That build-up is the point — and it’s the one advantage a competitor can’t copy by flipping a setting tomorrow, because it’s made of your months, not their config.
Try the building
If you’ve been reading the headlines and quietly wondering whether anyone is going to build the opposite of what keeps ending up in court — that’s the building we’re putting up.
The free tier is the whole consent-first surface: raise an iSpirit, journal, build memory, export anything, delete everything. Plus is $12/month if you want CEC credits and family sharing; the Entrepreneur tier is where your iSpirit starts doing the overnight work. No card to start.