See a bot, say a bot: why we're building SynthStamp
Scroll any big subreddit for ten minutes and you’ll meet them. The account that posts a perfect sunset every morning and has never once replied to a comment. The reply that says “What a stunning capture! The interplay of light and shadow truly showcases nature’s beauty” under a photo of a parking lot. The meme you saw last week, now on the front page again from someone else, with the same title.
None of this is new. What’s new is how cheap it has become. Writing a plausible comment costs nothing now, and neither does running a few hundred accounts that post them. The result is a feeling a lot of people have and can’t quite prove: that a growing share of what you read wasn’t written by anyone.
Detectors guess
The obvious answer is an AI detector. Point software at the text, get a verdict. The trouble is that detectors are wrong in both directions. They flag people who write carefully, they miss models told to write sloppily, and they can’t tell you why they decided what they decided. A verdict you can’t question isn’t much help.
Meanwhile, the people actually reading a thread notice things a detector never sees. They notice that an account posts every twelve minutes around the clock. They remember the photo from last week. They can tell when a comment says nothing in four sentences. That knowledge exists, in thousands of heads, and it evaporates the moment they scroll on.
Readers notice
SynthStamp is a browser add-on that writes that knowledge down where the next reader can see it. Next to every username on Reddit there’s a small flag. Click it and you can tag the account as an AI bot, a repost bot or a spam bot, or as an honest bot like a subreddit’s helper. You can tag a single comment or image as AI-made, a repost or spam without saying anything about the person who posted it. Or you can vouch that it’s human.
When enough people have tagged something, everyone with the add-on sees it: “bot 4”, “AI image 9”, “✓ 6”.
If this sounds familiar, it’s the idea that made SponsorBlock work on YouTube. A few people mark what they see, and everyone else benefits. Nobody has to do much, and the tags pile up where they’re useful.
A count of people, never a verdict
We made one rule early and haven’t broken it: SynthStamp never decides anything. A flag is a count of people and what they said. When people disagree, you see both numbers, and neither side wins by getting there first. You look, and you make up your own mind.
Nothing gets hidden either. Tagged comments stay on the page, marked. Some people will want tagged comments folded away, and that’s a switch you can turn on, but it’s off unless you choose it.
Reposts can’t hide behind a resize
Images are where it gets interesting. Your browser turns every picture into a short fingerprint, and the picture itself never leaves your computer. A resized, recompressed or mirrored copy lands on the same fingerprint, so a tag on one copy shows on all of them. The next time that sunset turns up in another subreddit, it arrives already tagged. We wrote up how the fingerprinting works separately.
What it costs you
To recognise a repost, we have to have seen the original. So the add-on sends what the Reddit pages you open show: the text of posts and comments, who posted them, and a fingerprint of each image. We’re upfront about that on what we collect, and the add-on shows you the full list before it does anything. One switch turns the sharing off while tags keep showing.
It’s free, it works without an account, and it’s useful from your first tag: whatever you mark stays marked for you, wherever it turns up again.
Where it goes
Reddit is first. X, Bluesky, Facebook, Instagram, Threads, YouTube, LinkedIn and TikTok are planned, one at a time, each announced before it’s switched on. Further out, there’s a feed of things people vouched were made by people.
For now: see a bot, say a bot. Get the add-on.