Bot Detector Deprecated
Scores a Reddit commenter on three separate things, because a paid human passes every bot check.
Deprecated. No longer maintained or developed. The repository stays public, with its code and tests as they are.I've stopped working on it. The code is still public for anyone to read, but nothing new is coming.
How to use it
It is public and installs unpacked. Four steps, then a click checks anyone.
It is free to get and installs in one step. Four things, then one click checks someone.
- Get the code: github.com/gkbii/bot-detector Public. Clone or download it. No build step.Download it from GitHub. Nothing to compile.
- Load it: Developer mode, then Load unpacked Point the browser at the extension folder. That is the whole install.Turn on developer mode in your browser and point it at the folder.
- Open a Reddit thread and scroll A small button appears next to each commenter. Nothing is looked up until you click one.Open Reddit and a small button appears next to each person.
- Click the button to check someone One click scores the account; a second shows what stood out, each linked to the comment or post behind it.Click once to check someone, and again to see what stood out, with a link to the real comment.
Get it on GitHub
The extension itself, public and installable: the only part of this lab that runs on your machine rather than mine. Load it unpacked and it scores in your browser: no account, no build step.
The add-on itself, free and ready to install: the only thing here that runs on your computer. Load it in your browser and it works: no account, nothing to set up.
Three Scores, Never One Number
Someone paid to post talking points has a real account age, organic timing and varied language: clean on every automation signal, which averaging would hide. The third answers who argues in good faith, which a tool that only accuses never does. Twenty-seven frozen accounts are re-scored on every change: humans land 0–25 on automation, declared bots 44–75, nothing between.
A real person paid to push a message looks completely normal on every automation check. Blending the scores would hide them. The third score exists so the tool can also say when someone seems genuine. A set of 27 real accounts is kept exactly as it was, so anyone can check it still tells bots and people apart.
No Account, No Server, Nothing to Sign Up For
Reddit's own feed refuses a plain server request whatever headers it sends, so there is no server-side path to it at all. That forced the scoring into the browser and, incidentally, made the tool need nothing from me. One worker owns every lookup, so five open tabs stay one polite client of a free public archive.
Reddit blocks requests from servers, so everything had to happen inside the browser instead. That turned out well: it needs no account and nothing running at my end. It also asks the archive it reads politely, however many tabs you have open.
It Says What It Could Not See
A thin account gets no score rather than a flattering low one: absence is neither innocence nor guilt. Two signals once fired on its own paging, inventing a twelve-year dormancy.
The index it reads froze in March 2025, so the newest one in five were called deleted rather than unseen, and are now scored from what can be read.
It says how much it managed to read, and answers that it does not know rather than guessing. Two of its strongest signals once reacted to its own behaviour instead of the account.
It called people it could not find missing, when the list it checks had stopped being updated. The newest one in five now get an answer.
The Optional Read, With Its Citations Checked
Nothing here needs it, and running without it loses only two things: the judgement pattern-matching cannot make, and a lookup cache shared between machines. The model may cite only the numbered comments it was handed; an unresolvable citation is dropped, a finding with none left is rejected, and nothing survives means no block at all.
It works fine without this. Turned on, it asks a model what the account is actually arguing, and the model can only point at comments it was given. Anything it cannot point at is thrown away rather than shown.