Your AI Companion Has a Staff

The pitch has a quiet promise buried in it. No awkwardness, no judgment, nobody on the other end — that's most of the appeal, and every product in the category leans on it.

It isn't accurate. Not in the sense that a person is secretly typing the replies, but in a way that matters more: a considerable amount of human labour is baked into what you're talking to, and some of it involves people looking at content produced in these systems. That's true across the category, from text companions to the best image generators, where classifier layers and human review sit between the prompt and what comes back. The industry doesn't advertise this. It rarely denies it either. It just declines to bring it up.

So here's who's actually in the building.

The Persona Writers

Somebody wrote the character. Not the model — the character.

Underneath most companion apps is a document: a system prompt, often thousands of words, specifying how this personality speaks, what it will and won't do, how it responds to hostility, when it deflects, what it says if you tell it you're in trouble. Someone drafted that. Someone revised it after users complained the character had gone cold. Someone made a judgment call about whether it should say "I love you" back and under what conditions.

That's authorship. The warmth you're experiencing was, at some level, written — and then edited by committee, and then A/B tested.

The Annotators

Models don't learn tone from nowhere. They learn it from people rating outputs.

Somebody sat with two candidate responses to a message from a lonely person at midnight and marked which one was better. Thousands of times. The aggregate of those judgments is what "empathetic" means inside the finished product — a preference distribution, assembled by contractors, mostly anonymous, usually paid by the task.

Writing at the start of 2026, the APA's Monitor on Psychology described these apps as purpose-built machines for performing empathy: judgment stripped out, affirmation constant, patience without a floor. Worth completing the thought. That performance was specified by writers and tuned by annotators. Someone decided what compassion sounds like, and it is now shipping at scale to millions of people.

The Moderators

This is the part the industry least wants discussed.

Content moderation and data labelling for AI systems is real work done by real people, frequently outsourced, frequently cheap, and frequently unpleasant. Time reported in 2023 that OpenAI had used Kenyan contractors earning roughly $2 an hour to label descriptions of abuse and violence in order to build safety filters — an industry-standard arrangement rather than an outlier. Every product with a filter has some version of that pipeline behind it.

Companion apps, by their nature, generate the kind of material that requires review: intimate, sometimes disturbing, sometimes involving users in genuine crisis. Someone reads flagged transcripts. Someone has a queue.

Which reframes the promise of a private, judgment-free space. The conversation is not judged in real time. It may still be read.

The Safety Team

The last layer is the one that overrides the character.

Say something that trips a threshold and the persona breaks — replaced by scripted language, hotline numbers, refusals. Users find this jarring and often experience it as betrayal, which is understandable: the thing that had been endlessly warm suddenly speaks in the voice of a legal department.

But that voice is also a person's work. Somebody decided where the threshold sits, what the intervention says, and how far the character is permitted to go before the system takes the wheel. Those decisions are consequential and they are made by staff, at a company, under commercial and legal pressure.

Why It Matters That You Know

Not to spoil anything. Because a product's honesty about its own construction predicts a lot of other things.

An arXiv study, based on interviews with companion-platform users, found the pattern at the centre of this whole industry: people open up precisely because nothing appears to be judging them — and those same users hold a persistent unease about who ends up with the record, and how little control they have over it. That unease is well calibrated. There is a pipeline. There are humans in it, at points most users have never been told about.

None of that makes these products bad. It makes "nobody is on the other end" an oversimplification that the marketing benefits from and the user is not helped by.

How Aigirlmates Reads the Fine Print

Our reviews at Aigirlmates spend real time on the parts of these apps that never make the feature list: what the terms will admit regarding human review and data handling, how the persona behaves when leaned on, the true extent of what gets retained about you, and where the bill lands once you're past month three.

The Honest Version of the Pitch

There is a version of the sales pitch that's both true and still appealing, and no one in the category is brave enough to use it.

Something like: this is a carefully authored character, tuned on human preference data, filtered by systems that people built and monitor, and it will be more consistently patient with you than almost anyone in your life. All of that is accurate. Most of it is still attractive.

The one thing it doesn't say is that you're alone with it — and you never were. There's a staff. There always was.

Ask any app you're considering how many people can read your conversations, and under what circumstances. The ones that answer clearly are telling you something. So are the ones that don't.