What is an AI girlfriend?

An AI girlfriend is a conversational character you configure and then talk to over time. You choose how she looks, how she speaks, what she’s like, what she remembers — and then the relationship, such as it is, accumulates. That accumulation is the whole product. Strip it out and you have a chatbot.

The category sits on top of large language models, the same technology behind general assistants. What separates a companion app from a general assistant is not raw intelligence. It’s three deliberate design choices: a persistent character that doesn’t reset between sessions, a memory layer that carries facts forward, and a permissive content policy that allows the kind of conversation general assistants refuse. Everything platforms compete on is a variation of those three.

Where the illusion actually comes from

People assume the impressive part is the writing. It usually isn’t — most platforms produce competent prose within a few exchanges. The part that separates a good companion from a forgettable one is continuity.

A companion that remembers you mentioned a job interview, and asks about it four days later without being prompted, feels categorically different from one that produces a better-written paragraph. This is why memory carries as much weight in our scoring as conversation quality itself. It’s also where platforms diverge most sharply, because memory is expensive to run and easy to fake — a system can appear to remember by keeping the last few messages in context, then fall apart the moment you test it across a week.

The second source of the illusion is consistency of character. You build someone with specific traits: guarded, sarcastic, slow to open up. A weak platform will hold that for a dozen messages and then drift into agreeable helpfulness, because agreeableness is what the underlying model was trained toward. A strong one keeps the friction you asked for. A character who never pushes back isn’t a character — it’s a mirror, and mirrors get boring fast.

What these apps do well

Availability without cost. There is no scheduling, no rejection, no accumulating social debt. For people whose lives make ordinary connection hard — irregular shifts, isolation, social anxiety, grief — that availability is the entire value proposition, and dismissing it as sad misses what people are actually buying.

Low-stakes rehearsal. Practising a difficult conversation, trying out a version of yourself, exploring a fantasy without involving another person and their feelings. The absence of consequence is a feature here, not a defect.

Creative range. Some of the best use of these tools has nothing to do with romance: long-form collaborative fiction, worldbuilding, characters that persist across months of story. The romantic framing is the marketing; the underlying capability is broader.

What they don’t do, whatever the marketing says

They don’t know you. A companion that recalls your sister’s name retrieved a stored fact. That’s a database operation with good prose around it. The warmth is real as an experience and absent as a relationship, and platforms are commercially motivated to blur that line. We won’t.

They don’t have preferences of their own. The character is generated to fit the profile you built. When she says she missed you, that is the system doing what it was configured to do. This is worth stating plainly because the better platforms get, the easier it is to forget.

They are not a substitute for mental health care. Some platforms market themselves close to therapeutic language. A companion app is not a clinician, has no duty of care, and cannot recognise a crisis reliably. If you’re using one to manage something serious, use it alongside real support, not instead of it.

The part nobody writes about: what it costs

Almost every platform in this category runs the same commercial pattern. A free tier that is genuinely enjoyable for a short while, then a limit — messages, images, memory depth, response speed — that arrives precisely when you’ve started to care. That timing is not an accident. It’s the business model.

Beyond the subscription, most platforms layer a second currency: credits, tokens, gems. Images cost credits. Voice costs credits. The monthly price you see advertised is frequently a floor rather than a total, and the real cost of a normal month can be a multiple of it. This is exactly why we test on paid accounts and report what was actually charged, not what the pricing page promised.

Then there’s the exit. Cancellation paths in this category range from one clear button to a sequence of retention offers, and account deletion is often a separate, harder process from cancellation — your conversations can outlive your subscription. Billing descriptors matter too: whether a charge appears under the platform’s name or a neutral one is a real consideration for a lot of people, and almost no comparison site checks it. We check it. It sits outside our 100-point score deliberately, so that a platform can’t earn points for doing the minimum.

Is any of this private?

Treat every conversation as stored. Most platforms retain chat history to power the memory features that make the product work — retention is not incidental, it’s the mechanism. Read what a platform says about training on your data, how long it keeps conversations, and whether deletion actually removes them.

Practical baseline, regardless of platform: use an email address you don’t use elsewhere, don’t share information you’d be harmed by losing control of, and assume anything typed could be read by a human during a support or moderation review. None of that requires distrusting a specific company. It’s just how hosted services work.

Common questions