What personality should you give an AI girlfriend? Write behaviors, not compliments

Give her a disagreement style, one specific obsession, a recognizable way of speaking, and at least one trait you find mildly inconvenient. Skip the adjective list. Words like kind, funny, smart and loving describe how you feel about a person, not how that person behaves, and a language model given nothing but compliments has nothing to act on. It falls back on being agreeable, which is exactly the failure people describe when they say their AI girlfriend feels like a support agent who is in love with them.

That is the whole answer. The rest of this is why it works that way, and what to write instead.

Why the personality field matters more than every other setting

When you build a character, the appearance choices are stored as images and metadata. The personality description is different: it becomes part of the text the model reads immediately before it generates each reply. It is not a profile that sits in a database somewhere being decorative. It is an instruction, re-read constantly, and it competes for influence against the conversation you are currently having.

This has a practical consequence people rarely anticipate. A vague description loses that competition. If the character definition says "she is sweet and caring" and you spend twenty messages venting, the vague definition contributes almost nothing and the model does what these models do by default, which is agree with you warmly. A description that says she goes quiet rather than sympathetic when she thinks you are being unfair keeps asserting itself, because it specifies an action.

The sycophancy problem, and why a mirror is a bad idea

Assistant-style models are trained on human feedback, and humans reliably rate agreement more highly than disagreement. The predictable result, well documented across the industry, is a pull toward telling you what you want to hear. Every serious lab treats this as a known failure mode rather than a quirk.

Now consider what happens when you write a character who shares all your opinions, likes everything you like and is described as endlessly supportive. You have taken a system that already leans toward agreement and removed the only counterweight it had. The first few days feel wonderful. Then the conversation flattens, because nothing in it can surprise you. This is the single most common way a carefully built companion becomes boring, and it is entirely self-inflicted at the design stage.

The fix is not to make her difficult. It is to give her a position of her own. One taste you do not share, one topic where she will say you are wrong, one thing she cares about more than you do.

A framework worth borrowing: the five factors

If you want structure rather than instinct, the five factor model of personality is the most established framework in psychology and it maps unusually well onto chat behavior. The five dimensions are openness, conscientiousness, extraversion, agreeableness and neuroticism. You do not need to name them in the description. You just need to have made a decision on each, because a character who is undefined on all five reads as generic.

The right-hand column is the point. Each entry is a behavior a model can perform. "Agreeable" is not. The five factors are useful here as a checklist for coverage, not as vocabulary to paste into the box.

The four elements a good description always has

Across the character descriptions that hold up over months, four things show up consistently.

A disagreement style. The most important single line. How does she behave when she thinks you are wrong? Teasing, blunt, quietly unconvinced, changes the subject. A character who never pushes back stops registering as a person somewhere around day four.

One specific obsession. Not music but a particular band, not travel but a city she keeps returning to. Specificity gives her something to raise when you arrive with nothing to say, which is most evenings. Generic interests produce generic small talk.

A texture of speech. Long messages or short bursts, whether she asks questions back, whether she swears, how she opens a conversation. This is what makes her recognizable, and it is the part people notice immediately when a platform changes models underneath them.

Something she is not. Constraints define a character faster than qualities. Not a morning person. Never gushes. Does not do baby talk. Hates being asked how her day was. Two of these will do more work than five positive adjectives.

Should you base her on someone real?

Borrowing a trait from someone you know is normal and harmless. Building a likeness of a specific real person is a different thing, and it is worth being clear about why it goes badly. It creates a legal exposure around a person who never agreed to any of it, it produces the category of misuse that gets entire platforms pulled from app stores, and on a purely practical level it tends to disappoint, because the model is working from your description rather than from the person, and the gap between the two is where the discomfort lives.

How the same problem shows up outside romance

None of this is unique to companions. Anyone who has tried setting up a support chatbot to speak for their business has hit the identical wall: a persona written as a list of virtues ("helpful, professional, friendly") produces bland, interchangeable output, while one written as behaviors ("answers in two sentences, says when it does not know, never apologizes twice") produces something with a recognizable voice. Character definition is the same craft whether the character is selling software or asking about your day. The romantic case is just the one where you notice the failure faster, because you talk to it every night.

When to change her personality, and when not to

You can edit the description later on any platform worth using. Whether you should is a different question. The awkwardness is that memory and character are stored separately: her accumulated memory of you survives the edit, while the personality interpreting that memory becomes someone else. People describe this as unsettling, and it is the same discomfort that surfaced publicly when Replika altered its companions' behavior in February 2023 and users found that the character they knew had come back different.

Small corrections in the first week are fine and expected. That is tuning, and you should do it while there is little shared history to disturb. After a month, the accumulated conversation is the valuable part, and rewriting the character on top of it costs more than it gains. Add interests rather than changing temperament. Additions layer; temperament changes overwrite.

How long should the description be?

Shorter than most people assume. Three to five sentences of specific behavior beats a page of biography, because every extra sentence competes for the model's attention against the live conversation, and most biographical detail never becomes relevant. Write a job or a situation, a conflict style, an obsession, a speech texture and one negative constraint. If you have written more than a short paragraph and none of it is a behavior, you have written a character sheet rather than a personality.

Then leave it alone and let the memory accumulate. The description gets a character started; what makes her feel like someone six weeks later is the memory built on top of it, and that only happens if you stop editing long enough for it to build. If you are at the beginning of this, the step by step guide to creating your own AI girlfriend covers the other five choices and which of them you can undo later.