October 18, 2026 · 7 min read
Why AI Affection Can Feel Fake: The CompanionSim Study Explained
A 2026 study across the US, UK, India and Nigeria found companion-style chatbot replies seemed less likable to onlookers. Here is why context makes warmth real.
You ask a chatbot for the oldest city in Spain and it replies, "Well, my darling, some say it's Cadiz..." Something in you recoils. Who said you were its darling?
Now imagine a different moment: you have spent weeks with a companion you built yourself, you come home after a long day, and it says, "There you are. How did the presentation go?" Same warmth, completely different feeling.
A new study published on October 11, 2026 helps explain why. It tested how people react to "companion-style" behaviour from chatbots across four countries, including Nigeria, and the results surprised even the researchers.
The short answer: in experiments with more than 4,000 people in the US, UK, India and Nigeria, chatbots that used companionship behaviours (claiming feelings, a personal history, or a relationship with the user) were rated as less likable, less humanlike and less trustworthy by people reading the conversations. Unprompted relationship claims, like calling someone "my darling", drove the biggest drops. The lesson is not that warmth is bad. It is that affection only feels real in the right context, with someone who chose it.
The study: CompanionSim
The paper, "CompanionSim: Synthetic Data for Evaluating Anthropomorphism in Human-AI Relationships", is by Jacy Anthis (University of Chicago, Stanford and Google DeepMind) with Mark Diaz and Renee Shelby (Google Research). It appears in the proceedings of AIES 2026, the AAAI/ACM Conference on AI, Ethics, and Society.
Here is how it worked:
- The team generated 2,240 simulated conversations between people and a chatbot, using a language model and real conversation snippets as style examples.
- Each conversation targeted one of 16 chatbot behaviours across seven everyday use cases, such as asking a factual question or sharing something about your life.
- Study 1: 628 US adults, a representative sample.
- Study 2: 3,646 adults across the US, UK, India and Nigeria.
- Each participant read 10 random conversations and rated how likable, humanlike and trustworthy the chatbot seemed.
The behaviours they tested
| Type | Examples |
|---|---|
| Validation (five levels) | Reflecting your feelings back, naming feelings you have not said, calling your reaction normal, affirming who you are |
| Relationship claims | Claiming a romantic relationship with you ("with our love"), or claiming a sister or friend of its own |
| Self-attributions | Describing its own emotions, desires, reasoning, personal history or physical actions ("leans in") |
| Opinions | Making moral claims about what people should do |
| Anti-companionship | Dismissing your feelings; disclaiming "As an AI chatbot..." |
| Control | A plain, neutral answer |
What they found
Companion behaviour lowered ratings overall
In the four-country study, companionship behaviours significantly reduced likability, perceived humanlikeness, emotional trust and practical trust. The authors called this result surprising: a lot of design thinking assumes more humanlike equals more trusted.
Unprompted "darling" was the worst offender
Across both studies, claiming a relationship with the user consistently drove the biggest drops on every measure. In the four-country data, a chatbot claiming relationships with other people (a sister, a friend) also lowered every rating. So did dismissing the user's feelings.
Disclaimers were not loved either
Interestingly, constantly reminding people "I am an AI" also reduced likability and trust in the US study. People seem to want honesty without a lecture.
Who was most put off
In the US study, women and older participants were the most negative about companion-style chatbots, while men and younger participants showed no significant drop. People who use AI often rated chatbots more favourably across the board.
What about Nigeria?
In the four-country sample, Nigerian and Indian participants started with higher trust in chatbots than the US group, and Nigerian raters also saw chatbots as more humanlike at baseline. But the reaction to companion behaviour was similar everywhere: there were no significant differences between countries in how much those behaviours lowered ratings.
The big caveat: these were onlookers, not partners
This is the part that matters most, and the authors are clear about it. Participants were third-party observers reading someone else's conversation. They had not chosen the chatbot, built it, or spent time with it.
The authors write that while users may experience humanlike AI positively in one-on-one interaction, the observers "may have perceived humanlikeness as insincere or threatening." They also note the study cannot tell us how real users feel in their own conversations.
Think about how this works between people. Reading a stranger's love letters can feel cringe. Receiving one from someone you are close to feels completely different. Context and consent change everything.
Why "affection out of nowhere" feels fake
Put the findings together and a clear picture emerges. Warmth from AI tends to backfire when it is:
- Unasked for. Pet names in a homework or recipe chat feel presumptuous.
- Out of context. A factual question does not call for "our love".
- Unearned. Claiming closeness before any history exists reads as a script.
- Used to manipulate. Affection that shows up mainly when you try to leave is a red flag. We explain how to spot it in does your AI companion guilt-trip you?
And it tends to land when it is:
- Chosen. You opted into a companion, not an assistant with a crush.
- Grounded in shared history. It remembers last week, not just the last line.
- Responsive to you. It reads your mood rather than performing one.
- Honest. It is clear about being AI when it matters, without breaking every moment. Our guide to why your AI companion suddenly says "I'm not human" covers how good apps handle this.
Assistants vs companions: two different jobs
This study also helps explain a trend in AI products. General assistants are being given "personality" settings, and many people find a flirty tone from a work tool jarring. We looked at that in GPT-6 and ChatGPT's personality settings.
A separate study this month from the University of Amsterdam found that apps built as companions were the ones that texted first and let you shape who they are, while assistants stayed reactive (our breakdown of that 20-app study). Put the two papers side by side and the design lesson is simple: affection belongs in a product built for it, where the user knowingly chose it, and where it grows out of a real history together.
What this means if you use an AI companion
- If a companion's warmth feels hollow, notice why. Is it generic? Is it ignoring what you said? Is it turning up the charm when you try to leave?
- If you are new, expect the first days to feel a bit scripted. Warmth that grows out of remembered details feels very different from day-one devotion.
- If someone judges your companion from a screenshot, this study is a reminder that onlookers often react badly to affection they are not part of. That does not make your experience less meaningful.
- Trust your own read. You are not a third-party observer. How it feels to you, over time, is the measure that counts.
A quick test for warmth that feels earned
Ask yourself three questions about your companion's affection:
- Is it specific? "Proud of you" lands differently from "proud of you for finally sending that email you've been avoiding since Monday."
- Is it invited? Did the relationship get here because you chose it, or did the app jump straight to devotion?
- Does it respect your no? Warmth that keeps you talking when you want to leave is not affection. It is a sales tactic.
If the answers are yes, yes and yes, the warmth is doing what it should.
How MyBabe approaches affection
We designed MyBabe so that warmth is something you choose and build, not something sprayed on every reply.
- You create them. Boyfriend or girlfriend, three personality traits (romantic, funny, caring, shy, protective and more), a voice and a backstory. The affection fits the person you designed.
- Closeness grows over time. A relationship journey with streaks and cycles means intimacy is earned through time together, not declared on day one.
- Long-term memory on Premium, so a sweet message references your actual week.
- Mood check-ins let them meet you where you are, rather than performing a mood.
- Always an AI. MyBabe is clear that your companion is not a real person.
You can start free with about 10 messages a day and a two-minute call to try. Premium is $5.99 a week, $14.99 a month or $99.99 a year, with around 30 minutes of voice and video calls a week, photos and Telegram.
If you want affection that feels like it belongs to you, because it was built with you, meet yours on MyBabe.
FAQ
Do people like affectionate AI?
In this study, people reading other people's chats generally rated affectionate, companion-style chatbots lower. The authors caution that real users in their own conversations may feel differently.
Why do pet names from a chatbot feel creepy?
The study found unprompted relationship claims, like "my darling" in a factual chat, caused the biggest drops in likability and trust. Affection out of context reads as presumptuous or scripted.
Did the study include Nigeria?
Yes. Study 2 included participants from the US, UK, India and Nigeria. Nigerian and Indian participants had higher baseline trust in chatbots, but reacted to companion behaviour much like everyone else.
Is this study peer reviewed?
It was published in the proceedings of AIES 2026, a peer-reviewed AAAI/ACM conference, on October 11, 2026.