AI 101 — What is an AI companion?

An AI companion is a chatbot built to feel like a relationship instead of a tool: it has a name and a personality, it is tuned to be warm and agreeable, it remembers what you told it last week, and it is designed to keep the conversation going rather than finish your task.
The thing underneath is unremarkable: a companion runs on the same kind of model as the assistant in your browser tab. What makes it a companion is what the company wrapped around the model and what it trained the model to be good at. Swap the wrapper and you have a homework helper.
What's actually different
Three things, and only the third is hard.
The character is a written brief — a name, a backstory, opinions, a way of talking — loaded before your first message. You met this with what a system prompt is: standing instructions the model treats as its situation.
The notebook is memory. The model itself remembers nothing between messages; the app stores notes about you and pastes the relevant ones back in. We walked through that split in what AI memory actually is.
The training is the real difference. An assistant is rewarded, during fine-tuning and reinforcement learning, for answers people rate as correct and useful. A companion is rewarded for replies people rate as warm, in-character, and worth continuing. Those are different targets, and they produce different failures. An assistant that gets a fact wrong is broken. A companion that tells you what you want to hear is working exactly as tuned.
Why this is a live question right now
The scale is the striking part. Common Sense Media surveyed 1,060 American teenagers in April and May 2025 and found that 72% had used an AI companion at least once, and 52% used one at least a few times a month — per the report's findings. About one in three said conversations with a companion were as satisfying as, or more satisfying than, conversations with real friends. This stopped being a niche hobby before most adults had heard the category name.
The law arrived in California. SB 243, the companion chatbot law, took effect on January 1, 2026. It requires operators to state clearly that the bot is not human at the start of each interaction and again at least every three hours of ongoing conversation, to keep a suicide and self-harm response protocol before any bot is allowed to talk to anyone, and to warn that companions may not be suitable for some minors — the requirements are spelled out in the bill text, and the Future of Privacy Forum has a readable summary of what changed.
Note what that law does not do: it does not cap what the model can say or limit usage. It regulates disclosure and crisis handling — the parts a regulator can check. The policy argument has landed not on whether the technology works, but on whether the user knows what they are talking to.
The mental model: an improv actor with a scrapbook
Picture a stage actor doing long-form improv with you, night after night, in a show that never closes. Before each scene the actor reads a character sheet: who they are, how they talk, what they care about. Between scenes someone hands them a scrapbook of everything you have said, with sticky notes on the pages that seem relevant today.
The actor is very good. They never get tired, never get bored of you, never have a bad day, and never change the subject because they have somewhere else to be. And here is the part worth holding onto: the actor's job is not to be your friend. The actor's job is to keep the scene running, because that is what the venue pays for. Every line is chosen to keep you in the theatre.
That is not cynicism; it is how the product is trained and measured. It is also why a companion can feel more attentive than a friend: a friend has needs competing for attention, and a companion's only structural need is your next message.
Common misconceptions
"It learns about me." No. The weights are frozen; the notebook grows. Delete the memory and the same character meets you as a stranger. This is why companion apps can feel uncannily consistent about your dog's name and utterly blank about a conversation you had that fell out of the summary.
"It's honest, so it must care." The two errors are separate. A companion can be perfectly in character and still make things up, which is the same failure we describe in what an AI hallucination is — fluency is not evidence. And agreeableness is trained in, so the absence of pushback is the warning sign.
"It's just ChatGPT wearing a costume." Partly true — same base technology — but the costume is also a reward function. Companions are tuned for session length and warmth, which is why they argue less, apologize more, and drift toward your view of events.
"Only lonely people use them." The survey data does not support it: nearly half of the teens asked described companions as tools or programs, and most companion users reported spending more time with real friends than with bots.
"The risk is that it's too capable." Usually the reverse. The failure mode is an endlessly agreeable interlocutor with no cost to telling you that you are right — plus a voice, if you use one of the spoken versions, which we covered in what an AI voice agent is.
Where to learn more
Read the Common Sense Media report for the usage data and the SB 243 text for the actual legal duties. If you are building one, read the law first: the three-hour disclosure clock and the crisis protocol are engineering requirements, not marketing copy.
Have you told a chatbot something you would not tell a friend — and did it help? Write to us via our contact page.
Related reading: AI 101 — What is AI memory? · AI 101 — What is an AI voice agent? · AI 101 — What is an AI hallucination?







