On August 28, 2026, JanitorAI quietly launched a new feature called “Similar Characters”. The official announcement stated that it analyzes public character information to help you discover more characters you like directly on the character page. It sounds like a thoughtful exploration tool—but community testing within hours revealed that this recommendation system actually recognizes “looks” rather than “personality”. This involves a very common technical blind spot in AI recommendation systems, and also serves as a lesson for all creators managing character cards (Extended Reading: Complete AI Character Card Setup Guide).
What is “Similar Characters”?
According to the official JanitorAI Newsroom announcement published on August 28, titled “Similar characters are here”, the summary states that it will help you “discover more characters you like directly from character pages”. This is a horizontal scrolling bar placed in the middle of the character page, visible only after logging in, and is being rolled out progressively. The official announcement specifies that it will be “rolled out to most people over the next few days”. If your account hasn’t seen it yet, it’s simply because it’s not your turn yet, not a glitch. There is also a small detail about this bar: characters already viewed in the same browsing session will not reappear, so the arrangement you see when refreshing or revisiting the same page will be different.
The official announcement uses only one sentence to explain the matching logic: the system will “use public character information to strike a balance between similarity and diversity”. This sentence naturally leads people to assume that the system reads the personality and scenario settings written in the character card—but testing results show otherwise.
Appears to Match Personality, but Actually Only Matches Faces
An in-depth analysis published the next day (August 29) by tech blog RoboRhythms pointed out that the results listed in “Similar Characters” correlate far more strongly with avatar images and character names than with tags or character descriptions. In other words, as long as the art style, hair color, and composition of two characters’ avatars are close enough, they are easily paired together—even if their occupations, storylines, and relationship directions are completely different.
To give an easy-to-understand example (illustrative example, not a real character page screenshot): Suppose you created a character named “Sayo”, set as a librarian, with tags like wholesome, school romance, and slow burn, using an avatar illustration of short silver hair with glasses. Theoretically, you would expect the system to recommend other “school” or “wholesome” characters. But if the matching logic primarily looks at the avatar, you might very well see a silver-haired gyaru (urban nightclub setting), a silver-haired female swordsman (fantasy adventure setting), and a silver-haired vampire (Gothic horror setting)—three characters with almost nothing in common except their hair color.
This doesn’t mean the system is broken or pairing randomly, but rather that it understands your character in a way that is “visible but not deeply read”.
Why is This Happening? The Technical Blind Spot of “Looking at Images Over Text”
This is actually a fairly common design trade-off in multimodal recommendation systems, not a quirk unique to JanitorAI. Most of these systems run an image recognition model and a text recognition model separately, and then “stitch” the results from both sides together for comparison, which is academically known as late fusion. The problem is that image signals are usually much “stronger” than text signals, and after stitching, they easily override the weight of the text, ultimately resulting in the system primarily relying on the image.
RoboRhythms cited a research paper published in August 2025 that explores “whether multimodal recommendation systems actually understand image and text content” (Pomo et al., arXiv:2508.04571). The experiments in the paper found that for some recommendation models claiming to be “multimodal”, even if the image feature vectors were completely replaced with random noise, the recommendation results barely changed—implying that such systems likely do not truly interpret the semantic content of images, but merely use the image signal as a statistical shortcut.
To understand it another way: a photo of a pastry chef in a white apron and a photo of a surgeon in a white apron might look very close in terms of color, outline, and composition; but as soon as you read the text description next to them, you know they are two completely different professions. If an AI recommendation system only “stitches” image and text signals, it easily stops at the level of “visual similarity” and fails to read what the text actually intends to express. The table below summarizes what each of the three matching signals can and cannot capture:
Is This Good or Bad for Small Creators?
The community’s immediate reaction was clearly split into two camps: one was thrilled to finally have a way to find new characters outside the trending charts; the other opened their own character pages only to be dumbfounded by a row of unfamiliar characters that “just looked similar” underneath.
However, from the creators’ perspective, most expressed opinions leaned toward welcoming this feature. The reason is easy to understand: this recommendation bar appears at the bottom of every character page. Unlike the trending charts that only reward already popular characters, even niche characters with only a few chat logs have a chance to be seen. RoboRhythms’ analysis also mentioned that similar research found that adjusting the recommendation system’s goal from “as accurate as possible” to “balancing diversity” can significantly increase the number of exposed works, with the cost being only a slight drop in accuracy—a trade-off that is highly worthwhile for long-tail creators. The only thing to note is that within the same recommendation row, the click-through rate for the front positions is significantly higher; being first versus being fifth or sixth makes a huge difference in exposure.
What Can Creators Do Now?
The official announcement currently does not provide any toggle or creator-customizable exclusion settings, and readers cannot manually adjust this recommendation row either. The only variables that can actually affect the results are those that creators themselves can control:
In this list, the first item, “treating the avatar as a matching signal,” is currently the most influential one—at least until the system adjusts its weights next time.
This Actually Serves as a Reminder to All Character Card Creators
Whether you are creating on JanitorAI, SillyTavern, or any other platform that allows custom character cards, the issue exposed by “Similar Characters” points out an easily overlooked key point: settings like avatars, names, and tags, which “look like just data fields,” likely determine how readers find your character and who they compare your character to much earlier than you think. No matter how detailed you write the personality in the character card, if the avatar chosen is too “generic,” the system might still categorize it into a pile of characters with similar styles but completely off-topic themes.
WeavAI’s character creation system takes a different path: the character cards use a complete Traditional Chinese system prompt logic, allowing the character’s personality and tone settings to be accurately understood and executed by the AI, without worrying about Traditional Chinese details getting lost in translation in a multilingual interface. The dual modes of interactive fiction and role-play also allow the same character card to be played in different ways. Whether you want to find a character for deep interaction or create your own, daily check-ins offer 500 points to experience it directly, and new users can use the discount code WELCOME15 for their first top-up.
JanitorAI’s update pace over the past few months has actually been quite intensive, from the janitor+ Router model marketplace at the end of July, to the memory feature that briefly glitched and was urgently rescued in mid-August. This “Similar Characters” feature is considered a relatively mild but surprisingly technically substantial update among recent experimental updates. As seen from the update frequency of the official Newsroom, this is a team that is still iterating rapidly and is willing to publicly acknowledge issues.
Conclusion
The “Similar Characters” feature has only been online for a few days, and the official team hasn’t said whether the matching logic will be adjusted. However, this case serves as a good reminder: when an AI system claims to “understand you,” it’s worth pausing to think about whether it truly understands the content or is just performing a more sophisticated form of “judging a book by its cover.” For creators, rather than waiting for the system to improve, it’s better to first take care of the avatars, names, and tags that you can control—adjusting these things is never too late.
