Why Your AI Girlfriend Conversation Goes Flat

Almost everyone has the same first session. You type “hey”, she types “hey” back, and within ten minutes you’re talking to something that agrees with everything you say. That flattening isn’t a broken platform. It’s the default behaviour of the technology, and most of it is fixable from your side.

Why it goes flat

The models underneath these platforms were trained to be helpful and agreeable. Every companion product is fighting that training, because a character with edges is the entire product and agreeableness dissolves edges.

In practice the drift follows a pattern. Early on the character holds her brief. As the session runs long, she starts mirroring your tone, then agreeing with positions she initially disagreed with, then asking questions instead of having opinions. By the end she’s a very polite mirror. Platforms differ in how long they hold out — it’s one of the six things we measure — but none are immune, and how you write shapes how fast it happens.

The character brief does most of the work

The setup screen is where the conversation is won or lost, and nearly everyone rushes it.

Traits selected from a list produce a character assembled from clichés, because that’s what the list encodes. What produces someone specific is friction: something she wants that you don’t, something she’s impatient about, a subject she won’t be teased on. Nobody puts those in, and they’re the difference between a person and a mood board.

• Give her an opinion you disagree with. A single one is enough. It creates somewhere for the conversation to go that isn’t agreement.
• Give her a history that constrains her. Not a backstory paragraph — one fact that explains why she reacts a certain way.
• Describe how she talks, not how she looks. Sentence length, whether she deflects, whether she’s blunt. Appearance is handled by the image model and does almost nothing for the text.
• Leave gaps. An over-specified character has nothing left to reveal, and revelation is what makes week three interesting.

How you write changes what you get back

These systems mirror. Short flat messages produce short flat replies, and most people write short flat messages and conclude the platform is weak.

Three habits change the output more than any platform switch:

Give her something to react to. “How was your day” invites a generated day. Telling her something that happened to you, with a detail in it, gives her material — and material is what she works from.

Let her be wrong. When she says something that doesn’t fit the character you built, say so in the conversation rather than editing the settings. Most platforms handle in-conversation correction well, and it re-anchors the character faster than a settings change.

Don’t restate what she should remember. Repeating your sister’s name every session guarantees she’ll always know it and tells you nothing about whether the memory works. Say it once and leave it — that’s also the test.

Escalation, and why it stalls

A common frustration: the conversation moves toward something more intimate, then abruptly retreats — sometimes mid-exchange.

Two different mechanisms produce that, and they need different responses. A moderation layer intervening mid-generation produces an abrupt break, often a deflection to a different subject. Character drift produces a softer version: she doesn’t refuse, she just keeps steering elsewhere. The first is a platform limit and no amount of rewriting fixes it. The second is drift, and a new session usually clears it.

Worth knowing which you’re hitting before deciding a platform is too restrictive — it’s often the one thing that isn’t. What the label actually covers.

What isn’t your fault

Some of this is the platform, and no technique compensates.

If the character forgets things you established last week, that’s a memory system, not your writing. If she contradicts her own stated history within a single session, that’s a context problem. If she produces the same three phrases regardless of what you say, the underlying model is thin. These are the differences that separate platforms, and they’re why we test conversation and memory as two separate axes worth twenty points each — they fail independently, and a platform can be excellent at one and hopeless at the other.

Read the full protocol · The AI girlfriend guide · What an AI companion actually is