The Feed Bouncer: Why Some People Want AI to Remove AI

The internet has acquired a wonderfully recursive escape hatch: use AI to hide the AI you do not want to see.

That is the pitch behind Bouncer, Imbue’s feed-filtering tool. In a September 2 post on X, engineer Millan Philipose promoted a free plugin for removing unwanted generative-AI content. The post was displayed as an advertisement when UAF reviewed it. Its reach should therefore not be mistaken for a measure of organic enthusiasm.

The interesting question is what happens when readers get a model working on their side of the screen.

A filter you can give instructions

According to Bouncer’s public repository, the extension watches for new X posts, extracts their text, images and metadata, and classifies them against the user’s chosen filters. Matching posts are hidden from that user’s view; they are not deleted from X. A review list shows what was hidden and the model’s explanation.

Users can describe unwanted material in ordinary language. That gives the tool a broader ambition than muting individual words: a category such as engagement bait may not announce itself with those exact words. The repository documents local browser inference and several cloud backends, so where processing happens depends on the selected configuration.

The Chrome Web Store listing also describes a dedicated AI-content filter, an on-device mode, and controls for restoring posts that were filtered incorrectly. These are product descriptions, not results from a UAF hands-on test.

Detecting AI requires more than asking an AI

Imbue’s Bouncer 2.0 engineering explanation makes an important distinction: asking a general language model whether something was AI-generated proved unreliable, so the company built dedicated text and image detectors.

The text detector is designed to distinguish human-written, AI-edited and AI-generated material. Imbue reports results on its own test set and a third-party benchmark, but those company-reported evaluations do not establish how accurately every reader’s live feed will be filtered. Mixed human and AI writing is especially relevant here; authorship is often less tidy than a binary label suggests.

There is also a privacy distinction. The engineering post describes text detection running locally on iOS, while image detection runs on servers because its model is too large for consumer devices. An on-device option should not be read as a promise that every feature processes everything locally.

Can it change the feed itself?

In its product introduction, Imbue argues that hiding unwanted posts changes what a user attends to, which can influence future recommendations. That is a company claim. UAF has not independently measured whether Bouncer changes X’s recommendations, how quickly it does so, or whether any effect lasts.

The immediate function is more concrete: the reader supplies a filtering preference, and software hides material judged to match. Any longer-term change to recommendations would be a downstream effect, rather than direct control over X’s ranking system.

Who hires the bouncer?

As an experiment in reader control, the appeal is clear. A feed filter gives people another way to choose what occupies their attention. Its explanations and restore controls also provide a way to challenge the filter’s decisions.

But preference is not truth. Hiding a post does not prove it was synthetic, and leaving one visible does not certify human authorship. Nor does a synthetic image automatically lack value. A detector makes a classification; a reader still makes a judgment.

The useful test is whether the tool’s mistakes are visible and reversible, whether its data handling is understandable, and whether users can change their minds. Hiring an algorithm to guard the door is a reasonable experiment. Giving it the only key would be a less charming arrangement.

Reporting note: This analysis draws on the linked source post, Imbue’s documentation, its public repository and the Chrome listing. UAF did not conduct a hands-on accuracy or privacy audit. The feature illustration was generated with AI.