Chai: Evaluating a Platform, Not a Product
Chai belongs to the part of this category where the characters are written by users rather than by the operator. That single structural fact changes what you are evaluating: not a product with a voice someone designed, but a platform hosting many voices of wildly varying quality, plus whatever moderation sits over them. Confirm that the current listing still describes it that way before relying on any of this — and if it does, the checks below are the ones that follow, and they are almost entirely different from the checks you would run on a single-character app.
The mistake worth avoiding is judging a platform by the first character you happen to open.
What changes when users supply the characters
Four consequences, and none of them is a criticism.
Quality varies more than between products. Two characters on the same platform, using the same model and the same machinery, can be a great deal better or worse than each other, because the difference is in what someone wrote. A bad first experience on a platform is genuinely uninformative in a way a bad first experience with a designed product is not.
Responsibility splits. The operator supplies the model, the limits and the rules; a stranger supplied the personality. When something in a conversation goes wrong it is worth knowing which of those produced it, because they have different remedies — one is a support ticket, the other is picking a different character.
Popularity signals are inside the app rather than outside. On a platform, the useful reputation data is which characters are widely used and how they are rated by other users, not what any external article says. That information exists in the app and is far more specific than a review of the app as a whole.
The catalogue is not stable. Characters are added, edited and removed by their authors. Something you rely on can change or vanish without the operator doing anything, which is a form of continuity risk that single-character products do not have.
The moderation question, which is the real one
On a platform, this is where evaluation actually lands.
Published rules for creators. Community or content guidelines aimed at people making characters are a different document from the terms aimed at users, and their existence and specificity tell you how the operator understands its own role. This is the document to look for.
How reporting works, and whether anything visibly follows. Whether there is a report control on a character, whether it explains what happens next, and whether the platform publishes anything about enforcement. A report button with no described process is a gesture.
What is filtered at the platform level versus what a character can override. A creator’s instructions and the operator’s rules are separate layers, and the operator’s is meant to win. How that machinery works generally is in what content filters actually are.
Whether character definitions are visible. Some platforms show you a character’s description or greeting. Where they do, you can see what you are talking to before you talk to it, which is the strongest single advantage a platform has over an opaque product.
Everything in this section concerns adults choosing what to use. Anything involving people under eighteen is outside this site’s scope and is not covered here.
What to check in the first ten minutes
The order is different from a single-character app, because you have to sample.
Open three or four different characters, briefly. You are calibrating the range rather than assessing one. The spread between the best and worst is the platform’s actual character.
Look at how characters are surfaced. Whether the platform ranks by usage, by rating, by recency, or by something it does not explain determines what you will end up talking to, and it is the most consequential piece of design on the whole product.
Find a character’s definition, if it is exposed. Reading the instructions behind a character is the fastest education available in how thin or thick a personality actually is.
Check whether continuity is per-character or per-account. On platforms this varies, and it determines whether switching characters costs you anything.
Check the publisher and the policy domain, as with any install — the store fields that matter are listed in spotting a lookalike companion app.
The claims to distrust on a platform specifically
Three, and they are platform-shaped rather than product-shaped.
“Thousands of characters.” Catalogue size is the easiest number to grow and the least related to whether any of them are good. A large catalogue with no useful discovery mechanism is worse than a small curated one.
“Community-driven” as a substitute for moderation. Sometimes it describes real, published governance. Sometimes it describes the absence of any. The difference is visible in whether the rules and the enforcement process are written down.
Any external review of “the app”. On a platform, an outside reviewer’s experience is a sample of a handful of characters chosen by an algorithm at a particular moment. It generalises poorly even when honestly reported, which is a specific instance of the broader problem covered in how to read an AI companion app review.
Why there is no verdict here
Rating a marketplace on the basis of not having used it would be less defensible than rating a product, and rating a product without using it is already indefensible. This site has run no sessions on Chai, sampled no characters, and tested no moderation process, and a score assembled from other people’s accounts would be a summary of a summary.
There is also a structural reason a verdict would be worth little even from someone who had used it. A platform’s quality is the quality of what its discovery mechanism puts in front of you, and that is personalised, time-varying, and unavailable to any reviewer’s reader. The honest deliverable is the distinction this post is built on — that a platform is a different kind of object from a product, and that its moderation and discovery are the things to examine rather than its conversational quality, which is not a single property it possesses.