The “Humans-Only” Feed Filter: A Browser Extension Idea That Keeps Showing Up

Here’s an idea that refuses to go away. In the opportunity research we track, a browser extension that filters or labels artificial intelligence (AI)-generated content in your feeds has shown up five times in the weekly ideas tracking. Then the same concept surfaced independently in our app-niche pipeline, which uses a completely different research method. When two separate streams converge on one product without knowing about each other, we pay attention. That doesn’t happen often.

The pitch is simple. AI-generated content is flooding social feeds, search results, and comment sections, and a growing group of people actively wants less of it. A browser extension that detects likely-AI content and either hides it or slaps a label on it gives those people a switch to flip. Call it a “humans-only” mode for the internet. Freemium pricing, a few dollars a month for the full version, and a marketing story that basically writes itself because the frustration is universal.

Why the demand is real and getting louder

Spend ten minutes on any social platform and you’ll see it: the same stilted phrasing, the same generic stock-image-that-isn’t-a-stock-image aesthetic, the same engagement-bait threads that read like they were assembled by a machine, because they were. People have started calling it “slop,” and the word stuck because it captures the feeling exactly. It’s not that any single AI post is offensive. It’s that the volume degrades the whole experience.

What matters for you as a builder isn’t the annoyance itself. It’s that the annoyance has crossed into behavior. People are searching for ways to filter AI content, installing crude blocklists, sharing keyword-muting recipes, and joining communities dedicated to finding human-made writing and art. That’s the signal that separates a real product opportunity from a complaint. When people are already hacking together bad solutions, a good solution has a market waiting.

And the audience is easy to find. They’re loudly self-identifying in Reddit threads, artist communities, photography forums, and quote-posts complaining about slop. You don’t need a clever acquisition strategy. You need to show up where the complaining is happening with something that works.

What the product actually looks like

Don’t overbuild this. Version one is a browser extension that works on two or three major platforms, scores content on a likely-AI scale, and gives the user three choices: label it, blur it, or hide it entirely. That’s the whole product. A settings page with a sensitivity slider and a per-site toggle, and you’re done.

Under the hood, you’ve got options, and you should probably use several at once. Text classifiers catch the statistical fingerprints of machine writing. Image checks can look for known generator artifacts and metadata. Account-level signals matter too: posting frequency, account age, and repetitive structure across posts often say more than any single piece of content. You can also lean on community reporting, where users flag accounts and everyone benefits from the shared list. Blending signals is what keeps you useful when any one method gets dodged.

The freemium split is straightforward. Free gets you labeling on one platform with default settings. Paid, somewhere in the $3 to $5 a month range, gets you hiding, multiple platforms, custom sensitivity, and the community blocklists. At $4 a month, a thousand paying users is $48,000 a year — the same small-software math behind CRM marketplace add-ons — and this is the kind of product where a viral moment can add a thousand installs in a weekend because every frustrated user is a walking advertisement.

The honest part: detection is an arms race

We’ll be honest, this is the caveat that should keep you up at night if you build this. AI detection is not a solved problem, and it never will be. Every time detectors improve, generators improve to evade them. The academic consensus is that reliable detection of a motivated, well-resourced generator is somewhere between hard and impossible. You are signing up for a permanent maintenance burden, not a build-it-once product.

False positives are the sharper risk. If your extension labels a human writer as AI, you haven’t just made a technical error. You’ve insulted your user’s favorite blogger, or worse, the user themselves. A few visible false positives and your credibility is gone, because the entire product is credibility. This is why the labeling-versus-hiding distinction matters so much. A label that says “likely AI, 70% confidence” is honest and survivable when it’s wrong. Silently hiding a human’s work is not. Ship labels first, let users opt into hiding, and never pretend to certainty you don’t have.

There’s also platform risk. Extensions that modify how sites render content live at the mercy of site redesigns and browser policy changes. Every major platform update is a fire drill. Budget for it.

How you’d actually launch this

The go-to-market here is unusually cheap because the frustration is the marketing. A launch post that says “I built a humans-only mode for your feed” is engineered to travel. The communities that care about this, artists, writers, photographers, and plain tired internet users, share tools like this without being asked.

• Build for one platform first, the one where slop complaints are loudest in your own feeds.
• Launch free with labels only, and collect false-positive reports from day one with a one-click “this is human” button.
• Add the paid tier only after the detection feels trustworthy, because charging for a flaky filter kills you twice.
• Publish a monthly accuracy note. Transparency about error rates is a feature, not a confession.

That “this is human” button is doing double duty, by the way. It’s customer service and it’s training data. Your users will happily correct your classifier for free because they want the product to work.

Who this isn’t for

Skip this one if you want a passive product. This is closer to running a spam filter than shipping an app: the adversary adapts, and you have to keep showing up. If you can’t stomach a product that’s wrong some percentage of the time in public, this will be miserable, because it will be wrong, visibly, and people will screenshot it. And if you don’t have at least intermediate technical skills, or a partner who does, the classifier work and the constant platform breakage will bury you. No-code won’t get you there.

But if you’re comfortable with machine learning basics, enjoy an adversarial problem, and want a product where demand grows every time the internet gets a little worse, this is one of the stronger asymmetric bets in our current tracking. Five appearances in one stream, an independent echo in another. The frustration isn’t going anywhere. The only question is whether you can stay accurate enough, long enough, to be the tool people trust.