ByteDance and Alibaba disabled custom AI agent features as China’s companion crackdown took effect. I'm building in this category.
But I'm not worried.
The rules expose a measurement gap I’ve been working on all year.
Most headlines focused on users losing companions they’d spent months or years shaping.
What caught my attention was why the rules were disruptive enough for major platforms to switch those features off rather than adapt immediately.
At SuperAI and AI Engineer, I asked people working in evals how they'd recommend assessing long-term relational outcomes for companion products.
Not whether a single-turn reply was safe, but whether hundreds of interactions left someone with more agency, increased dependency, or displaced parts of her offline life.
Most answers came back to output safety, incidents and engagement. Those matter, but they still tell us more about the AI’s output than what the relationship becomes.
Retention usually looks like proof that a product is working.
Emotional AI makes that signal messier because the product can adapt itself around one person. It can learn what reassures her, what lowers resistance and what keeps her coming back.
High retention might mean the relationship is valuable. It might also mean the system has become very good at removing friction and making itself harder to leave.
China’s rules make that distinction much harder to ignore.
They scrutinise excessive accommodation, emotional dependency, damage to real-world relationships, signs of distress and whether users can leave easily.
The relationship itself is becoming the unit of risk. The higher bar is overdue.
Companion companies will increasingly need to prove what the relationship is designed to do beyond generating attachment.
For Maibel, that outcome sits outside the chat.
Before our companion moves into coaching, our system checks whether the user is distressed, resisting the direction or simply needs warmth. A “no” changes what he’s allowed to do next.
We already evaluate parts of that layer: emotional-state routing, distress handling, memory confidence, respect for refusal, trust rupture and repair.
The harder next step is measuring what those behaviours add up to across weeks and months.
Did the relationship help her sleep better, move more consistently or restart when life knocked her routines sideways?
Did it preserve her judgment and agency?
Or did it simply become very good at keeping her engaged?
A higher bar won’t kill the companion category, but it'll change what companies compete on.
When attachment stops being enough, evidence of user benefit becomes part of the moat.
That’s better for users, good for ambitious founders and probably overdue for a market that has treated retention as proof-of-value for too long.
Honestly, I’d rather compete on that.
Making someone want to stay is one thing. Proving she was better off because she did is much harder.
Hi, I’m Mabel.
I’m the founder of Maibel. I write about emotional AI, women’s wellness and learning to build the things I wish existed.
More about me
