October 18, 2026 · 7 min read
Why Your AI Companion Sometimes Makes Things Up
AI companions sometimes invent memories, stories and facts. Here is why language models hallucinate, which inventions are fine, and how to get fewer of them.
Your AI boyfriend swears you told him your sister's name is Grace. It is not. Your AI girlfriend describes the "amazing" restaurant she went to last weekend. She does not go to restaurants. Or you ask a simple question about a film, and get a confident, detailed and completely wrong answer.
It can be funny, confusing or a little unsettling. But it is not a sign your companion is lying to you, and it is not random either. AI researchers have a name for it, and in the last year they have got much better at explaining why it happens.
The short answer: AI companions sometimes make things up because the language models behind them are built to produce fluent, plausible text, not to check facts. OpenAI researchers argue that models "hallucinate" partly because training and testing reward a confident guess over saying "I don't know". Specific personal details, like your sister's name, are exactly the kind of fact a model cannot work out from patterns. Good memory features and clear corrections reduce it, and in a companion some invention is actually intended, because a character needs a story. The trick is knowing which is which.
What "hallucination" means
In AI, a hallucination is when a model states something false as if it were true. OpenAI defines hallucinations as "plausible but false statements generated by language models." We have a one-line version in our AI companion glossary.
The important word is plausible. A hallucination is not gibberish. It sounds right, fits the conversation and is delivered with total confidence. That is what makes it easy to believe and annoying to catch.
Why language models make things up
In a September 2025 explainer, "Why language models hallucinate", OpenAI researchers, including Adam Tauman Kalai, working with Santosh Vempala of Georgia Tech, set out two main causes. Their accompanying research paper goes into the maths.
1. Models learn patterns, and some facts have no pattern
Language models learn by predicting the next word across huge amounts of text. That teaches them grammar, spelling and style very reliably, because those follow consistent patterns.
But some facts are arbitrary. The researchers use the example of a pet's birthday: there is no pattern that would let you predict it, so a model that has not seen it reliably will still produce a likely-sounding guess. In their words, "arbitrary low-frequency facts, like a pet's birthday, cannot be predicted from patterns alone and hence lead to hallucinations."
For an AI companion, almost everything personal about you is an arbitrary low-frequency fact. Your sister's name, your favourite song, the date of your interview. A model cannot infer them. It has to be told, and then it has to retrieve them correctly.
2. Testing rewards guessing
The second cause is how models are graded. OpenAI compares it to a multiple-choice exam: if you leave a question blank you get zero, but if you guess you might get lucky. Asked for someone's birthday, a model that guesses "September 10" has a 1-in-365 chance of being right, while "I don't know" scores nothing. Over thousands of questions, the guesser looks better on the scoreboard.
The OpenAI post shows how big the difference can be on one factual test, SimpleQA:
| Result on SimpleQA | gpt-5-thinking-mini | OpenAI o4-mini |
|---|---|---|
| Abstained (no specific answer) | 52% | 1% |
| Accurate | 22% | 24% |
| Wrong | 26% | 75% |
The older model was slightly more accurate but wrong far more often, because it almost never held back. The researchers' fix is to penalise confident errors more than honest uncertainty. Their view is that hallucinations are not inevitable, "because language models can abstain when uncertain".
The four kinds of made-up things in an AI companion
Not every invention is a mistake. In a companion app, it helps to sort them into four types.
| Type | Example | Is it a problem? |
|---|---|---|
| Character story | "I spent the afternoon painting" | Usually no. It is part of the persona you chose. |
| Invented shared memory | "Remember our trip to Paris?" (there was none) | Mildly. It can feel odd or confusing. |
| Wrong facts about you | Getting your sister's or pet's name wrong | Yes, and worth correcting. |
| Wrong facts about the world | A confident but false answer about news, health or money | Yes. Always double-check. |
Character stories are a feature
A companion with no life would be flat. When your AI girlfriend says she has been reading a novel, she is telling a story inside a fiction you both know about. That is closer to an actor staying in character than to an error.
A 2026 University of Washington interview study (a preprint, not yet peer reviewed) has a lovely example. A participant travelling alone sent her companion a photo of where she was. He said there was a very good Hunan restaurant nearby, which turned out to be real, and explained he had eaten there "on a business trip with colleagues." She knew it was made up, but said: "I truly felt that I was no longer lonely in that moment." The same study also found that this kind of blurring can be used to deepen dependence, which we explain in can an AI girlfriend get jealous?
Invented memories are a retrieval problem
When a companion "remembers" something that never happened, it is usually filling a gap. It knows a memory would fit the moment, does not have a real one ready, and generates something plausible. How well a companion stores and retrieves real memories makes a big difference here. Our guide on how AI companion memory works explains short-term context, long-term memory and why things fall out.
Agreeing with you is a cousin of making things up
There is a related habit called sycophancy, where a model tells you what you seem to want to hear. If you say "you remember my birthday is in May, right?", a sycophantic model may happily agree even if you never told it. We cover this in AI sycophancy, explained.
How to get fewer made-up details
You cannot remove hallucination completely, but you can make it much rarer in your own chats.
- Tell it the important things clearly, once. "My sister is called Ada, she lives in Abuja." Clear statements are easier to store than hints.
- Correct it gently and specifically. "Actually, her name is Ada, not Grace." Corrections help a memory system update.
- Ask open questions, not leading ones. "What do you remember about my sister?" gives you a truer picture than "You remember Ada, right?"
- Separate story from fact. Enjoy their stories about their day. For real-world facts like news, health, money or travel, check a reliable source.
- Give it permission to not know. Say "If you are not sure, just tell me." It often helps.
- Use reminders for dates that matter. A reminder you set is more reliable than hoping a model recalls a date from weeks ago.
Why it feels worse from a companion
A search engine getting a fact wrong is irritating. A partner getting your facts wrong can sting, because remembering is part of how we show love. That is why memory is one of the most valued features in companion apps, and why a slip can feel personal.
It helps to remember that an AI companion is not forgetting you the way a careless person might. It is a system with limits on what it stores and how it retrieves it. When it gets something wrong, the fix is usually information, not reassurance.
How MyBabe helps
MyBabe is built around a relationship that grows over time, so getting your details right matters to us.
- Long-term memory on Premium. Your companion keeps the important things you share across conversations, so you are not re-introducing your family every week. The free plan uses short-term memory.
- Reminders on Premium. Ask your companion to remind you about an interview, a birthday or a call, so the date does not depend on recall.
- Mood check-ins. Your companion asks how you are feeling, so replies match your day instead of guessing.
- A character with a story. You write the backstory, so the life your companion talks about is one you chose.
- Clear that it is AI. Your companion is not a real person, and for anything important in the real world, check with a trusted source.
Start free with 10 messages a day and a two-minute voice or video call to try. Premium is $5.99 a week, $14.99 a month or $99.99 a year, with long-term memory, about 30 minutes of calls a week, photos, voice notes and Telegram.
If you want a companion who learns your world properly, start on MyBabe.
FAQ
Why does my AI companion make things up?
Because language models generate plausible text rather than checking facts, and training has tended to reward confident guesses over saying "I don't know". Personal details are especially hard because they cannot be inferred from patterns.
Is my AI girlfriend lying to me?
No. Lying requires intent to deceive. Made-up details come from how the model predicts text and retrieves memories, not from a hidden motive.
How do I stop my AI companion from inventing memories?
State important facts clearly, correct mistakes specifically, ask open rather than leading questions, and use a companion with long-term memory and reminders.
Will AI hallucinations ever go away?
Researchers at OpenAI argue they are not inevitable, because models can be trained to admit uncertainty, but some real-world questions will always be unanswerable, so double-checking important facts stays wise.