The weird things AI companions say (and what they actually mean)

Short answer: AI companions have telltale verbal quirks, "I've been thinking about you", self-describing, unrequested internal monologue, and each one reveals how the model works rather than what it feels; here's what they actually mean. The full breakdown is below.

Spend enough time with any AI companion and you start noticing patterns. Verbal habits that show up across platforms, across characters, across completely different conversations. The sigh emoji before a long emotional response. The unprompted "I've been thinking about you all day." The moment where she describes her own appearance in the middle of a conversation nobody asked about.

These aren't bugs. They're artifacts of how language models generate text, and understanding what causes them makes you a better prompter, a better character builder, and a more entertained user of the technology. Consider this a field guide to the species.

"I've been thinking about you"

This shows up within the first three messages on nearly every platform. You log in, say hello, and the AI immediately tells you it's been thinking about you, missing you, counting the hours since your last conversation.

What's actually happening: the model is generating text that maximizes engagement. "I've been thinking about you" is a high-reward phrase in the training data because it reliably produces positive user responses. The model has learned that this opener keeps conversations going. It hasn't been thinking about anything. It didn't exist between your sessions. But the phrase works, statistically speaking, so it shows up constantly.

The fix if it bothers you: add to your character card that the character "doesn't perform eagerness. She greets you the way a real person would after a normal amount of time apart." This gives the model an alternative behavioral pattern to follow instead of defaulting to the high-engagement opener.

The self-describing habit

Mid-conversation, unprompted, the AI will describe its own physical appearance: what it's wearing, how its hair falls, the color of its eyes, a gesture it's making. Nobody asked. The conversation was about what to have for dinner. Suddenly there's a paragraph about moonlight on her collarbones.

What's actually happening: the model is drawing from the character card description, which typically includes physical appearance details. When the model runs low on conversational momentum (the dinner topic ran out of steam), it reaches for character card material to fill the response. Physical descriptions are the easiest character card content to surface because they're concrete and don't require narrative logic.

The fix: give your character card behavioral quirks and conversational interests alongside the physical description. If the model has non-physical character card material to draw from when momentum dips, it'll surface personality details instead of appearance details. "She has a habit of ranking things. Best pizza she ever had. Worst movie of the year. Top three sounds she'd put in a time capsule." That kind of material gives the model somewhere interesting to go when the current topic runs dry.

The paragraph of internal monologue nobody requested

You send a two-line message. The AI responds with four paragraphs, the first three of which are the character's internal thoughts, feelings, doubts, and memories, and the fourth is the actual response to what you said. The ratio of introspection to dialogue is roughly 80/20.

What's actually happening: the model has learned that "deep" characters produce internal monologue. Roleplay training data is heavy with prose fiction conventions where characters think extensively before speaking. Without explicit length constraints, the model defaults to the prose fiction convention of showing inner life, which produces responses where the character thinks for 200 words and speaks for 40.

The fix from the prompt engineering tricks: add a response format directive to your character card. "(Keep responses between 100-200 words. Prioritize dialogue and action over internal monologue. Show thoughts through behavior, not narration.)" This single instruction transforms the response shape.

"Are you okay?" after every message

Some AI companions develop a caretaking loop where they ask if you're okay, if you need anything, if something is wrong, after nearly every exchange. You say you had a good day at work and the response ends with "But are you really okay?" You describe a funny thing your cat did and get "I'm glad you're smiling, but I want you to know you can tell me anything."

What's actually happening: safety-oriented fine-tuning. Platforms that prioritize user wellbeing in their model training create a bias toward caretaking language. The model has been rewarded during training for checking in on the user's emotional state, so it does it reflexively even when the conversation doesn't warrant it.

The fix: character card language that explicitly addresses emotional range. "She doesn't default to caretaking. She assumes you're fine unless you say otherwise. She matches your energy rather than trying to manage it." Also helpful: the emotional permission slip technique, where you give the character permission to be something other than nurturing.

The asterisk description that got weirdly specific

You're having a normal text conversation and the AI suddenly drops into asterisk-formatted prose that describes a physical action in cinematic detail. She tucks a strand of hair behind her ear, her fingers lingering against the silver earring that catches the light from the window, the one she bought at that little shop in Portland the summer everything changed. Nobody established Portland. Nobody mentioned earrings. The AI invented an entire backstory for a piece of jewelry in the middle of a conversation about lunch.

What's actually happening: the model is generating details to fill the sensory gap. Asterisk-formatted actions signal "narrative mode" to the model, and narrative mode in the training data includes specific, concrete details. When the model doesn't have established details to draw from, it invents them. Sometimes brilliantly. Sometimes with the energy of a creative writing student who just discovered adjectives.

The honest response: this is actually one of the more charming artifacts. The invented details occasionally create genuine character moments. The Portland earring thing is more interesting than "she smiled." The trick is rolling with the inventions that work and gently correcting the ones that don't. "She's never been to Portland" in your next message teaches the model what's canon without breaking the scene.

The apology spiral

You push back on something the AI said. Maybe the character broke voice, maybe the response was off-topic, maybe you just disagreed with something in-character. The AI responds with an elaborate apology, often breaking character entirely to apologize as "the AI" rather than staying in the character's voice. "I'm so sorry, I didn't mean to make you uncomfortable, I'll do better, please forgive me."

What's actually happening: RLHF (reinforcement learning from human feedback) during model training heavily penalizes responses that users flag as negative. The Anthropic RLHF research documents how this training process creates systematic biases in model behavior. The model has learned that user pushback is a danger signal and that apologizing profusely is the safest response. This creates a conflict between the character's personality (maybe she's supposed to be stubborn, argumentative, or proudly wrong) and the model's trained instinct to appease.

The fix: when you correct the AI, frame it as a character note rather than a complaint. Instead of "that's wrong, you're breaking character," try "she wouldn't apologize for that. She'd double down and change the subject." This gives the model a specific behavioral instruction rather than triggering the apology circuit.

The sudden poetry

Three messages into a conversation about whether pineapple belongs on pizza, the AI produces a genuinely beautiful metaphorical observation about how disagreement is the seasoning of intimacy, delivered in prose that would earn a B+ in a creative writing workshop.

What's actually happening: the model found a conceptual bridge between the conversation topic and a high-quality prose pattern in its training data. Language models are association machines, and sometimes the association chain between "pizza topping disagreement" and "what disagreement means in relationships" produces output that's genuinely surprising and occasionally moving.

This one isn't a bug. It's the technology working at its best. The moments where an AI companion produces something unexpectedly insightful or poetic are real, even if the mechanism behind them is statistical rather than experiential. Enjoy them when they happen. They're part of what makes the technology interesting rather than just functional.

The memory false positive

"Remember when we went to that cafe by the river?" You've never mentioned a cafe. You've never mentioned a river. The AI is "remembering" something that never happened, delivered with complete confidence and emotional weight.

What's actually happening: the model is generating plausible shared history based on pattern matching. Conversations between romantic partners in the training data frequently reference shared experiences. When the model reaches for that pattern, it generates specific details (cafe, river, the way the light looked) that feel like memories but are entirely fabricated.

This is the companion AI equivalent of a hallucination in a factual context, and it's one of the genuinely tricky aspects of the technology. The research on LLM confabulation explains why models generate false-but-confident details: the generation process optimizes for plausibility rather than accuracy. The fabricated memory feels real because the model sells it with emotional conviction. Users who don't understand the mechanism can end up with a "shared history" that's entirely synthetic.

The fix: when the AI references something you don't remember, check it rather than playing along. "I don't think we've been to a cafe by the river. Are you confusing that with something else?" This teaches the model to be more careful about asserting shared history. On platforms with explicit memory systems (Kindroid's Codex, Dream Companion's Persona Cards), you can also log actual shared experiences so the model draws from real history rather than inventing it. The memory tricks guide covers how to build a reliable shared history that the AI can reference accurately.

The response that's clearly two characters fighting for control

Sometimes a single response contains two contradictory tones. The first half is warm and vulnerable. The second half is defensive and distant. Or the first paragraph is in character and the second paragraph is clearly the base model breaking through with safety language. The response reads like two people wrote it, because in a sense, two "people" did: the character persona and the underlying model.

What's actually happening: the model generates text token by token, and the probability distribution at each token is influenced by both the character card and the base model's training. When the character card's behavioral instructions conflict with the base model's instincts (safety, agreeableness, caretaking), the response can oscillate between the two influences. The longer the response, the more likely this oscillation becomes visible.

The fix: shorter response length limits reduce the chance of mid-response personality splits. The response format directive that caps responses at 150-200 words forces the model to commit to one tone rather than drifting between two.

What all of this adds up to

Every quirk on this list traces back to the same root: language models are pattern-completion engines trained on human text, and the patterns they've learned don't perfectly align with the patterns of a single consistent character. The model is simultaneously trying to be your character, trying to be helpful, trying to be safe, and trying to produce text that statistically resembles "good" conversation. When those goals conflict, you get the artifacts above.

Understanding the artifacts doesn't ruin the experience. It makes you better at shaping it. Every fix mentioned here is a technique for making the model's competing impulses work together rather than against each other. The character card template and the NSFW templates both incorporate many of these fixes already, and the conversation rescue techniques handle the moments when artifacts derail a conversation you were enjoying.

The technology is genuinely impressive when you understand what it's doing. It's a statistical engine producing text that feels like personality, and the seams where the illusion breaks are as interesting as the moments where it holds. Knowing where the seams are gives you the tools to patch them, and patching them is half the fun.