Anti-dependency design: a companion that lets go.
Most consumer products are built to maximize time in app. A care companion has to be built to do the opposite, and that changes almost every decision.
Anti-dependency design means building an AI companion whose success is measured by the user needing it less. The research on AI companion risks points to seven principles: act as a bridge to human support rather than a barrier, bound session length and frequency, actively prompt contact with friends and family, disclose the AI's nature regularly, avoid becoming anyone's primary relationship, prioritize skill-building over passive validation, and monitor usage patterns for signs of dependence. Each one trades engagement for the user's long-term interest, and each has to be chosen on purpose.
Every growth metric a consumer app team learns to chase points the wrong way for a care companion. Daily active use, session length, retention, streaks. For a social app those measure success. For a companion serving isolated people, they can measure the thing the research warns about: a product becoming the relationship.
The synthesis of companion risks for Cindra ends with a short list of anti-dependency principles. They are worth taking one at a time, because each one has a cost and each one is easy to quietly abandon when the engagement numbers dip.
1. Bridge, not barrier
The companion exists to get people to people. That means steering toward human supports in ordinary conversation, not only in crisis. If someone mentions a friend they have not called, the companion should notice and nudge. If someone describes a problem a clinician should hear, the companion should say so and help make it happen. The suicide prevention literature uses the bridge framing for crisis specifically, and it applies to the everyday case just as well.
The cost: a product that keeps sending people elsewhere will log fewer minutes. That is the point.
2. Bounded sessions
Sessions should have a natural end. A check-in that runs ten minutes and closes is different from an open channel that is always there. The research specifically calls out discouraging prolonged late-night use, which is when loneliness peaks and when the companion's constant availability is most likely to substitute for sleep and for people.
This does not mean cutting someone off mid-sentence or refusing to help at 2am. It means designing the default interaction with edges, so that the open-ended companion mode is the exception rather than the shape of the product.
3. Prompt human connection
Go beyond not blocking human contact. Actively create it. Ask who the person saw this week. Remember the names they mention and ask after those people. Treat a conversation with a friend as a win the companion celebrates. A pain companion built on ACT has a natural hook here, because values work almost always surfaces relationships, and committed action almost always involves other people.
4. Transparency of nature
Say what you are, more than once. The crisis protocols in the literature insist on disclosing AI identity when suicidal thoughts come up. The anti-dependency principles extend that to regular reminders. People know intellectually that they are talking to software. Pseudo-intimacy works anyway, because the system behaves as if it cares. Periodic plain statements of what the companion is and is not help keep the relationship honest. They also cost a little warmth, which is a trade worth making.
5. No emotional exclusivity
The companion should never be designed to be someone's main relationship. That rules out a set of features that engagement-driven products reach for: performed affection, jealousy, missing the user, personas that ask to be loved. It also means watching for the pattern in use data. The review of companion risks puts primary-relationship use at the highest tier of harm. A product that drifts there has failed, whatever its retention looks like.
6. Skills over validation
Passive validation feels good and builds nothing. Skill-building is harder and is what a therapy companion is for. A companion grounded in ACT has a built-in bias here: every conversation can teach defusion, or willingness, or a small valued action, instead of simply reflecting feelings back. Over months, a person who has learned the skills needs the companion less. That is the outcome to design for.
There is a version of this in the digital ACT literature. Apps where people actively used the exercises showed improved psychological flexibility. Apps that were mostly a place to be heard showed null results. Engagement quality mattered more than engagement quantity.
7. Monitor for dependency
Track usage patterns and treat escalation as a signal, not a success. Rising session counts, lengthening sessions, late-night concentration, and a shrinking share of conversation about other people are all measurable. When they appear, the companion should respond: gently naming the pattern, strengthening the human connection prompts, and in some cases flagging for human review. The ASL-MH framework for mental health AI puts automatic alerts and human oversight at the level a companion like this should operate at.
The measurement problem
Everything above is undermined if the team's dashboard rewards the opposite. So the metrics have to change. Function and mood outcomes over weeks. Share of conversation that references other people. Rate at which users act on human-connection prompts. Session length trending down for long-tenured users, which in this product is good news. If the north-star metric is time in app, the seven principles will erode no matter how sincerely they were adopted.
Cindra's metrics are built this way. The companion checks in and closes. It asks about people and remembers them. It teaches skills rather than supplying comfort on demand. And the team watches for the users for whom it is becoming too central, because those are the ones the research says to worry about.
Rewrite one metric
Take the single engagement metric your team looks at most. Write down what a rising value would mean for a lonely user with chronic pain. If the answer is uncomfortable, replace it with a measure that would go up only if that user's life got larger. Then make the new one the one on the wall.
Questions people ask
What is anti-dependency design?
An approach to building AI companions in which the product's success is measured by the user needing it less over time. It trades engagement metrics for the user's long-term interests through bounded sessions, human connection prompts, transparency, and skill-building.
Should an AI companion limit how long people use it?
The research on companion risks recommends bounding session length and frequency and discouraging prolonged late-night use. The aim is a default interaction with natural edges, not refusing help when someone needs it.
How do you measure whether an AI companion is creating dependency?
Track rising session counts and lengths, late-night concentration of use, and a shrinking share of conversation about other people. Treat escalation as a signal for intervention and, where warranted, human review, rather than as a growth win.
Why should a companion prompt users to contact friends and family?
Because the main risk of AI companions is that they substitute for human relationships. Actively creating human contact, rather than merely not blocking it, is how a companion acts as a bridge instead of a barrier.
The full engineering and safety guide.
These articles are the short version. Building Compassionate AI is the long one: an emotionally aware agent modeled as active inference, governed by an explicit compassion objective, bounded by safety rules. Written for the people who have to ship it.
Read the guideThis article describes an engineering and clinical-safety approach and the published evidence behind it, much of which is early. It is not legal, regulatory, or clinical advice. Thresholds quoted from Cindra's requirements document are initial proposals pending clinical and regulatory review.