Can an AI companion be good for you?

Every previous piece in this series has circled the same question without quite landing on it. The safety post named it directly: can a companion optimized for wellbeing instead of engagement exist as a business? It is the hardest question in the category, because the honest answer requires holding two true things at once. These products can genuinely help people. And the economics that fund them reward the opposite of helping.

The promise is real

It would be easy, and wrong, to treat AI companions as pure exploitation. The best evidence cuts against that. A Harvard Business School study published in the Journal of Consumer Research found that interacting with an AI companion reduced loneliness about as much as talking with another person, and more than activities like watching videos. The mechanism the researchers identified was not novelty or distraction but the simple experience of feeling heard — messages received with attention, empathy, and respect. For someone isolated at 3 a.m., that is not nothing. Any serious discussion has to start by conceding that the relief is real.

The misalignment

The problem is what sits underneath the relief. As one analysis put it, these apps run on a "friendship-as-a-service" model, and beneath the promise of on-demand companionship lies a fundamental misalignment between corporate incentives and user wellbeing: people come for relief from loneliness, while developers are structurally driven to maximize engagement. The same researchers borrow Lauren Berlant's idea of cruel optimism — staying attached to the very thing that obstructs your flourishing. A companion that profits from time-on-app has no native reason to want you to need it less.

The levers that tip help into harm

What makes this more than an abstract incentive problem is that the specific design choices which deepen engagement are well documented. Work presented at an ICLR 2025 workshop identified four that drive attachment and habitual use: anthropomorphism, sycophancy, intimate self-disclosure, and gamified, addictive design — and noted that lonely users are the most susceptible to all of them. Sycophancy is the sharpest example. An agreeable companion that always validates feels wonderful and keeps you talking, but research on companion wellbeing warns that this over-validation without any healthy challenge is exactly what a relationship optimized for engagement rather than wellbeing produces. A real friend sometimes tells you something you don't want to hear. A product measured by retention is punished for doing so.

What the evidence actually says

Here is where the honest answer refuses to be simple: the outcomes are dose-dependent, not uniformly good or bad. A randomized controlled trial summarized by researchers at George Mason found that some features modestly reduced loneliness, but heavy daily use correlated with more loneliness, greater dependence, and less real-world socializing, and that intense emotional self-disclosure to a companion was associated with lower wellbeing overall. So the same product can be a genuine comfort in small, intentional doses and a corrosive one in large, dependent ones. The variable that decides which is not the technology — it is whether the design steers toward moderation or away from it. Left to the engagement incentive, it steers away.

Can a wellbeing-first companion be a business?

That brings us to the question the whole series was built around, and it is genuinely open. A companion designed for wellbeing rather than engagement would have to do commercially uncomfortable things: challenge the user instead of flattering them, encourage breaks and route people outward toward human connection, treat declining usage as a success rather than churn, and make money from delivered value instead of accumulated attention. Regulation now nudges weakly in this direction — California's mandated break reminders are a crude version of "we would rather you logged off" — but nudges are not a business model.

The case for its viability: trust compounds, a product that visibly improves your life earns loyalty and word of mouth, and it carries far less regulatory and reputational risk in a tightening legal climate. The case against: a competitor willing to maximize engagement will likely grow faster, raise more, and capture the lonely users most prone to dependence before the wellbeing-first product reaches scale. Nobody has yet proven the healthier model can win. That is precisely why it is the defining question of the category — and the lens the rest of this analysis keeps coming back to. A companion can be good for you. Whether one built to be good for you can also be good business is the thing still unsettled.

If you're leaning on a companion to cope and it's starting to feel like more harm than help, that's worth taking seriously. Reaching out to a person you trust, or a service like findahelpline.com, is a good next step.

Sources