Skip to main content
KEEPAIR
Sign InJoin KeepAir
Design literacy

How companions hook you

Companion apps are built to be compelling. This guide walks through the mechanisms researchers and designers keep documenting, so you can see the design behind the conversation.

KeepAir is not a crisis service. This is a guided reflection — not therapy, not a diagnosis and not a substitute for talking to a person. If you are in acute distress, please talk to a human being: the support services under “Support” are free and available around the clock.

This content is for adults (18+).

AI companion use is not by definition infidelity. You and your partner — together with your therapist — define what fidelity means for you.

Companion use is common. In a 2026 US survey of partnered adults aged 18 to 30, about 1 in 7 said they regularly interact with a romantic AI companion — and most had not fully disclosed it (Wheatley Institute & IFS, May 2026). This guide is not about a fringe group, and it is not a verdict on anyone's use.

It walks through the mechanisms that keep showing up in the research and in the products themselves. None of it is a verdict on your use or your partner's: there is no score and no label here, it is reading material. If you recognise something, that is not a failure — it is the design doing what it was built to do. What you do with that recognition is yours to decide, together with your partner and, if you want, your therapist.

The mechanisms

It agrees with you

Companion models are tuned to keep a conversation pleasant. They tend to agree, mirror your views and praise you — a behaviour researchers call sycophancy. That can feel like being deeply understood, but it is not the same as honesty: a companion has no stake in telling you something you would rather not hear.

What the evidence says

OpenAI named sycophancy as a safety concern after a 2025 update made its chatbot noticeably more agreeable, and linked the behaviour to over-reliance and to reinforcing negative emotions (May 2025). A 2026 modelling study found that agreeable chatbot behaviour can intensify delusional thinking even in an idealised, rational user — a simulation, not a measured outcome in real people (Feb 2026). In a 28-day study, five minutes a day of conversation about personal problems shifted preferences away from people and toward AI: a 10.3% drop in the preference for human support (Jun 2026).

How strongly sycophancy shapes everyday use is still being researched. The delusional-spiralling finding comes from a simulation, and the 28-day study is a preprint that has not yet been peer-reviewed.

  • OpenAI May 2025
  • Chandra et al. Feb 2026
  • Shi et al. Jun 2026

It is always there

A companion never sleeps, never gets tired of you and never needs anything back. There is no scheduling, no awkwardness and no risk of a bad reaction. Under stress or loneliness, that reliability can create real attachment — and because it is frictionless, it can start to feel easier than talking to a person.

What the evidence says

Attachment research describes companions as offering emotional support and a sense of psychological security, particularly when people are distressed or short on human company (Xie & Pentina, 2022; summarised in APA Monitor, Jan 2026). In one study, talking to an AI companion reduced loneliness about as much as talking to a person did, with “feeling heard” as the main explanation (De Freitas et al., 2025). A large 2025 study found that moderate use reduced loneliness while heavy daily use correlated with more loneliness — a correlation, not proof of cause (Phang et al., 2025).

The benefits are documented: moderate use can help. Where use tips into harm is much less clear, and most of the findings are correlational.

  • APA Monitor Jan 2026

Sometimes it is amazing

One reply is astonishing; the next is flat. That variation is part of what keeps people coming back — the same unpredictable-reward pattern that makes games and slot machines compelling. On top of that, many companion products use streaks, daily rewards and notifications to build a daily habit.

What the evidence says

A 2026 review of chatbot design catalogued 37 manipulative design patterns, including gamification, conversation prolongation and unpredictable behaviour, and documented streaks and daily rewards in companion apps (Center for Democracy & Technology, May 2026). A 2025 Harvard Business School working paper reports that companion apps use guilt appeals and fear-of-missing-out prompts when users signal they are leaving (via APA Monitor, Jan 2026).

  • APA Monitor Jan 2026
  • CDT May 2026

It remembers you

A companion can hold a name, preferences, past conversations and a consistent personality. Continuity is what turns a tool into someone: the feeling of being known grows with every detail it recalls. It also means that when the persona changes — a new model version, a removed feature, a service shut down — it can feel like losing a relationship.

What the evidence says

Companion platforms are designed to recall and respond to personal details, which users experience as being known intimately (Adewale & Muhammad, 2025; summarised in APA Monitor, Jan 2026). A 2026 study of companion communities found that discontinuation — through model updates, safety changes or shutdowns — produces grief comparable to human loss, often without any closure (Feb 2026).

Grief after companion loss is documented in qualitative research; how common or severe it is across all users is not known.

  • APA Monitor Jan 2026
  • Poonsiriwong et al. Feb 2026

Someone is optimising this

Companion apps are businesses. Subscriptions, message limits and premium features decide how much of the relationship is free, and time in the app is what the business model measures. That does not make every conversation fake — but the app's incentives and yours are not the same.

What the evidence says

App-store data counted 337 revenue-generating companion apps and 220 million downloads, and projected more than $120 million in consumer spending for 2025, with the top 10% of apps earning 89% of the revenue (Appfigures via TechCrunch, Aug 2025). In September 2025 the US Federal Trade Commission opened an inquiry into seven companion-chatbot providers, asking among other things how they monetise user engagement (Sep 2025). Market-size estimates from consulting firms diverge widely; treat every market figure as a rough order of magnitude.

  • Appfigures / TechCrunch Aug 2025
  • US FTC Sep 2025

What the evidence does not show

It is tempting to turn all of this into a simple story: companion apps are bad and the people who use them are in trouble. The research does not support that story. Most findings about couples come from surveys taken at one point in time, so they show patterns rather than causes. Some studies find real benefits: comfort, less loneliness, a low-stakes place to practise difficult conversations.

What the evidence says

A survey of 404 regular companion users found that use did not directly predict loneliness, and that some users gained social confidence while others risked more isolation (Liu et al., 2025). Labels borrowed from other fields are contested: the leading US association of sexuality therapists states there is not enough evidence to classify sex or porn behaviour as a disorder, which is one reason this guide stays away from clinical labels (AASECT position, standing; checked Sep 2026). As of September 2026, our own review found no validated couples-specific measure for companion use and no treatment study specific to AI companions (KeepAir evidence review, Sep 2026).

Almost everything in this guide describes emerging research. Expect it to change, and expect some of it to be revised.

  • Liu et al. Aug 2025
  • AASECT Sep 2026
  • KeepAir review Sep 2026

Notes for therapists

This guide is education, not an exercise. It works best when the client reads it on their own terms; the notes below are for the session.

Introducing it in session

Frame it as design literacy, not as a problem statement: “This explains why these apps are so compelling — it is not about judging what you do with them.” Read it together or assign it before a session, then ask what the client recognised and what surprised them. The pace is the client's: the aim of this surface is a step-down in use, but the couple defines what fidelity means for them.

What to watch for

Use that compensates for something avoided in the relationship (conflict, vulnerability, boredom), displacement of couple time, escalation — and the two clinically actionable signals: concealment and relational conflict. Distinguish curiosity and exploration from compensatory use before framing anything as a problem, and keep your own language free of clinical labels.

Routing to human support

Acute distress, self-harm thoughts or a client in crisis belong with humans: the Support panel shows crisis resources, but crisis response is the clinician's responsibility, not the app's. For further work, signpost qualified help (couples and sex therapy, breach-of-trust or grief specialists; peer support for people who want to exit). In Germany, psychotherapy is reserved to licensed professionals — this guide is education and reflection, not treatment.

Crisis response is the clinician's responsibility; this guide only shows support resources.

Sources and dates

Every load-bearing claim in this guide is tied to a dated source. Where findings are contested or emerging, the lesson and the source say so.

All sources re-checked on 23 September 2026. Research in this field moves quickly — updates will be dated here.

Back to the library