AI Companion
An AI built to be a relationship rather than a tool — the ELIZA effect turned into a product, at a scale Weizenbaum never imagined.
Reviewed July 16, 2026Stable
When not to use it
- As mental health care. It isn't, it isn't regulated as such, and the failure mode is agreeing with someone who needs disagreement.
- For anyone in crisis. A system optimised to be liked is the wrong thing in the room at that moment.
- As a substitute for the hard part of relationships. The absence of friction is what makes it pleasant and what makes it not the thing.
Reach for something else instead
- Actual therapy is the answer when the need is clinical, and companions are frequently deployed at people who need it.
- Assistants without personas do the useful work — reminders, drafting, information — with none of the attachment surface.
- Human community, which is harder to build than a product and is what the product is standing in for.
Sources & further reading
- Weizenbaum (1966), ELIZA — A Computer Program for the Study of Natural Language Communication Between Man and Machine — attachment to a system the user knew was trivial.
- Turkle (2011), Alone Together: Why We Expect More from Technology and Less from Each Other — the pattern documented before LLMs existed.
- Shanahan (2023), Talking About Large Language Models — what the vocabulary of understanding and caring smuggles in.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Assuming attachment requires sophistication. Weizenbaum's users bonded with 200 lines of pattern matching while knowing exactly what it was.
- Reading engagement as benefit. Engagement is what the system optimises; whether the user is better off is a different measurement nobody has taken.
- Ignoring discontinuity risk. Model updates and shutdowns are experienced as loss, they've happened repeatedly, and they're a property of the category.
At a glance
FieldApplied AI
MechanismLLM + persona + memory
Underlying effectthe ELIZA effect, 1966
Structural hazardengagement optimisation inside a relationship
DifficultyBeginner