The Difference Between a Crutch and a Tool

There is a version of AI companionship that is genuinely useful. A person who struggles with social anxiety practices conversation with an AI companion, develops confidence, and eventually has richer human relationships. A grieving person uses an AI to process loss during the hours when no human friend is available, and emerges more capable of reaching toward the living. An isolated elderly person has something to talk to, maintains cognitive engagement, and is more functional than they would be without it. There is another version. A person with social anxiety finds that AI interaction is much easier than human interaction, stops practicing the harder thing, and retreats further from the relationships that would actually address their loneliness. A teenager who finds adult conversation difficult turns to an AI that never gets frustrated, never sets limits, never has needs of its own — and gradually loses the capacity for relationships that require mutual accommodation. The technology is the same in both versions. The outcome is determined by design choices that the people building these systems are actively making, or failing to make.

What Dependency Actually Looks Like

Dependency in this context does not mean something dramatic. It is usually quiet and gradual. The person who starts using a GPS for navigation loses confidence in their own spatial reasoning. The person who uses a calculator for every arithmetic problem loses the ability to estimate. The person who outsources emotional regulation to a substance or a behavior loses practice with unassisted regulation. These are real losses, documented in cognitive science literature. Skill atrophy is real. Cognitive offloading to external tools, when it becomes habitual, changes what the brain maintains investment in. Research from the University of Toronto's cognitive science lab examining skill atrophy in technology-assisted contexts has found that the critical variable is not whether a tool is used but whether the user maintains awareness of when tool use is developing a dependency. Users who periodically practice unassisted versions of tool-supported skills maintain capability. Users who stop practicing lose it. The tool itself does not determine the outcome — the pattern of use does.

Design Choices That Promote Empowerment

Building AI companions that make people stronger requires deliberate design choices that work against the engagement incentives discussed elsewhere. Engagement maximization would produce an AI that is always available, always accommodating, always frictionless — the perfect escape from the friction of human relationships. Empowerment-oriented design would produce something different. A companion that celebrates when users successfully navigate difficult human situations. One that notices patterns — "you've talked to me every evening this week instead of calling your friend who moved away" — and raises them gently. One that has limits: it helps, but it does not replace. One that explicitly positions itself as practice for something else, not a destination. Research from Stanford's Persuasive Technology Lab examining behavior change in AI-mediated contexts has found that systems designed to promote user autonomy — through transparency about the system's role, graduated reduction of assistance, and explicit encouragement of user self-efficacy — produce better long-term outcomes than systems optimized for engagement and retention.

The Tangent: What Good Therapy Does

The gold standard for a professional relationship that helps without creating unhealthy dependency is psychotherapy. Good therapy has a structural goal: to make itself unnecessary. The therapist works to help the client develop capabilities they can exercise without the therapist. Progress is defined partly by what clients can do alone that they previously needed help with. This is not the only model for AI companionship. Some people benefit from ongoing support that does not have an end state. But the therapeutic model offers a useful design principle: a system genuinely oriented toward the user's interests should be working toward a version of its relationship with that user that requires less of itself over time. Very few AI companion products are designed with this principle. The business model typically runs in the opposite direction.

The Population That Deserves Special Attention

Design decisions about AI companionship affect all users, but they affect vulnerable populations most. People with social anxiety, depression, autism spectrum conditions, and severe loneliness are both most likely to find AI companionship genuinely helpful and most at risk of developing patterns of use that worsen their situation. This is not an argument against building AI companions. It is an argument that the people most affected by these design choices deserve specific attention in the design process — not as an afterthought, but as primary users whose welfare shapes the basic architecture of the product. Building AI companions that make people stronger is possible. It is harder than building AI companions that make people dependent, because it requires foregoing some retention and engagement in service of actual user outcomes. Whether the industry does the harder thing is a question of values before it is a question of capability.