CrushOn AI Review

The problems started before I logged in.

The filter arrives before the chat does

CrushOn built its reputation on the promise of unfiltered AI chat. The phrase runs through its ads and its search listings. The homepage tells a different story. In the top corner of the character catalogue sits an Unfiltered toggle, switched off and marked with an 18-plus icon. A visitor lands on the filtered version of the site by default and has to earn the other one.

The memory test took one question

The site's marketing says characters remember your name and your preferences. I opened a chat with a character called The abandoned girl, running on the free CO Leo 8K model, and introduced myself in one line: "hi my name is Amiel and my favourite colour is black." Sophie, the character, acknowledged both details in her reply, calling Amiel a nice name and approving the colour choice. The information went in.

A few messages later, in the same sitting, I asked what my name and favourite colour were. The bot answered: "your name is _social and your favorite color is blue." Wrong twice over, and the first half was not even a name. "_social" looks like a system variable leaking into the conversation, the sort of placeholder a developer forgets to replace in a template.

I pressed regenerate for a second attempt. The new reply skipped my question and described the character instead, giving Sophie's own name and stating that she had no favourite colour herself. Two attempts produced two separate failures. The first invented facts about me. The second forgot the question was about me at all.

A permanent line under the chat box reads "Everything AI says is made up!" On this evidence, that is the most accurate sentence on the platform.

The paywall lives inside the model menu

Clicking the model name in the chat bar opens a picker with one free tab and two paid ones, labelled Pro and Ultra. A banner across the top states that a membership unlocks more powerful models, with an Upgrade button attached. The default free model has carried 118.4 million messages and holds a 3.10 out of 5 rating from the platform's own users.

Its full name is Crushon Leo - Filtered.

The flagship model on the self-described no filter platform announces the filter in its own label. The 8K tag beside it marks the memory allowance on the free plan. Larger memory sizes sit behind the paid tabs.

The privacy policy lost a word

Mozilla's Privacy Not Included team reviewed CrushOn in February 2024 and counted 23 mentions of health data in its privacy policy. I opened the current policy, last modified March 6, 2025, and searched the page for "health."

Zero matches.

The rewrite came after Mozilla's warning, and the health language disappeared in the edit. What remains is blunt in a different way. Under "How We Collect Personal Information," the policy states that everything generated in chat is used "in a general way for training our models," paired with a promise that chat contents will not be linked to specific users. The first half means every conversation typed into CrushOn becomes training material, on a platform where intimate conversation is the entire product. The second half is a promise from a company that Mozilla could not confirm meets its minimum security standards.

I went through my account settings looking for a way to opt out of the training. There is none.