Replies, reviews and personal stories that seem to come from many people can now be written by machines. Rehumanize makes those synthetic voices visible in your feed, as you scroll.
Free. No account. Works on Reddit, X and Threads. Nothing in your feed is hidden, removed or reordered.
A new kind of misinformation
A fact-checker can test a claim. It cannot test a crowd. Generative AI lets one operator speak as thousands of people, and the agreement they seem to show is the message. No single sentence has to be false for the picture to be.
Over 1.8 million posts.
50,000+ fake accounts.
Six weeks.
Germany's Federal Foreign Office found a pro-Russian campaign on X posting automated German-language messages, on some days more than one a second, to wear down support for Ukraine: among its lines, that the government neglects its own people in favour of Ukraine. The Office reports that the campaign used AI language models to write its articles and the posts that spread them.
Source: Federal Foreign Office (Auswärtiges Amt), "Germany Targeted by the Pro-Russian Disinformation Campaign 'Doppelgänger'", technical report, 5 June 2024.
2.4× the output.
Still persuasive.
DC Weekly posed as an American news site: the only outlet in a Russian influence network to claim a US base. From September 2023 it ran stories taken from Fox News and Russian state media through OpenAI's GPT-3, and the model sometimes gave it away, leaving parts of its instructions in the published articles. Its daily output rose 2.4 times and its coverage spread across more topics. In a survey of 880 American adults, the AI-rewritten articles persuaded about as well as the ones it ran before, and readers found the site just as credible.
Source: Wack, M., Ehrett, C., Linvill, D. & Warren, P. (2025). Generative propaganda: Evidence of AI's impact from a state-backed disinformation campaign. PNAS Nexus, 4(4), pgaf083.
Both campaigns were found by investigators after the fact. The people scrolling past them had nothing in their feed to tell them. Rehumanize puts that signal in the feed, while you read.
Rehumanize marks posts that are likely written by AI. It does not judge whether a post is true, and it does not know who is behind it: it shows you when the voices around a claim may not be people.
For years the question was whether a post is true. Generative AI adds a second: whether the people who seem to be saying it are people at all. Fact-checking answers the first. Rehumanize brings the second into the feed.
People sharing news, opinions, jokes and their own lives.
What social media is for.
A false claim goes viral: a miracle cure, a doctored quote.
One claim that can be checked and corrected: what fact-checks and platform labels were built for.
A summary of verified election results.
Often useful, and not false.
A thousand "grassroots" opinions, all from one operation.
Manufactured consensus: no single claim to check and no single source to discredit.
When opinions and personal stories can be written in bulk, a thread can look like many people agreeing when few of them are people.
People use their feeds to judge what others think. Repetition can make a view feel familiar and widely held, and synthetic text makes repetition cheap.
Rehumanize gives back one missing piece of context, at the moment you read, and leaves the judgement to you.
Three invented posts, written for this page. Nothing here is checked or sent anywhere: the page draws the marks the way the browser extension draws them, so you can see what it adds to a feed.
Three sample posts, written for this page. Switch Rehumanize off and on to see what it adds, then choose how marks look. The browser extension shows the label on its own by default.
How marks look
Label only: the quietest option, and the browser extension's default.
Highlight and side line appear only on posts marked likely or partly AI-generated.
As a nurse of twenty years, I've never seen a policy fail patients like this one. Every colleague I've spoken to agrees: it's time for change. Share this if you stand with frontline workers.
bus was 25 min late again so i walked. passed the bakery on the corner that finally reopened, the queue was out the door at 8am. got the last cheese scone. worth being late for tbh
Went to the council meeting about the new cycle lane last night. Packed room. In conclusion, the proposal represents a balanced approach that addresses the diverse needs of all stakeholders while fostering a more sustainable community.
The Android app draws its marks its own way: see "What the marks mean".
Add it to Chrome on your computer. No account. Setup takes a minute.
On a phone? Open this page on your computer to install.
Scroll as usual. Nothing to paste, copy or click.
Read the mark, then judge the post yourself.
Each mark shows what an automated check found in the post's text. Whether the post is true, and what to make of it, is your call.
In the Android app
The optional companion app starts with an amber side bar and a soft highlight on posts likely or partly AI-generated. Switch the label on to see every mark on Threads and Reddit; on X the bar and highlight mark likely and partly AI-generated posts.
In the browser extension, everything lives in one side panel, opened from the toolbar icon. A six-step welcome, then nothing else to set up.
The label shows on its own by default. Add a highlight behind the text, an amber line beside it, or both. The Android app starts with the side bar and highlight.
Turn Rehumanize off entirely, or only on Reddit, X or Threads, straight from the side panel.
Click any mark and the panel says, in plain words, what it means.
How many posts were checked as you scroll, what they came back as, and which were skipped and why.
In the browser extension, marks follow each site's light or dark look; on Android they follow the phone's dark mode.
An optional companion app marks posts in the Reddit, X and Threads apps, with a small "Rehumanize active" chip on screen.
A post checked again, in another sitting or by another person, counts again. Posts that were not analysed, such as those too short or in another language, are not counted.
Every post Rehumanize checks, from every install, whether or not the person takes part in research. No individual records sit behind these numbers.
We never sell data. Rehumanize exists purely for academic research. Research is off until you say yes.
Your device detects each post's language before anything is sent. Posts in these languages are checked. A post in any other language shows "Language not supported" instead of a guess, and one whose language is unclear shows "Not analysed".
It measures the synthetic voices that surround a message, repeat it and amplify it.
From installs that say yes, Rehumanize records measurements for every post it checks. The data records what real people saw.
The running totals above are public. Research records are for research teams only, by agreement. Researchers can try the research dashboard in demo mode, with invented sample data or with real published articles from a documented campaign.
What research means for you.
Rehumanize Research is the research edition, for people who join a study with a code from its research team. Everyone else uses the standard extension, which the privacy policy calls the public version. Rehumanize Research is run by a research team. It records more than the public version, including the full text of each post once you join a study, unless the research team switches that off. Other details of posts, such as their authors, are kept only under switches that team turns on. The team chooses which switches are on and gives participants its own information sheet. Participants can leave the study at any time, from the browser extension's side panel or the Android app's main screen.
No. Research is off until you say yes, and you can change your answer at any time: in the browser extension's side panel, or on the Android app's main screen. Saying no changes nothing about how Rehumanize works for you.
In the standard extension (the public version): one record for each post Rehumanize checks, made of measurements: the result of the check; basic facts about the post, such as its exact number of words and its language; how and when the post appeared on your screen; random codes for your sitting and for the post, not made from its text; for English posts, automatic labels for tone, topic and hateful or offensive language; and the app version, the settings in use and which language check was used (in the browser extension, with your browser's main version number). The privacy policy describes each kind.
In the public version, never the post text, never usernames, never your identity or browsing history. Rehumanize Research records more: see "What is Rehumanize Research?" above.
Plain counts with no post text and no install ID: how many posts were checked and with what result, how many were not checked and why, and simple measures of how the setup and the switches are used and how well checks work. The privacy policy lists them in full.
A random code for each sitting, with the day and hour, platform, app type and version. This is how the app receives its current settings.
If you answered the research question, your answer, kept with a random install ID made on your device. The install ID is not your name, but it stays the same on your device until you remove Rehumanize, so it can be linked to that device. That ID is pseudonymous personal data, and it is never put on research records or sent with a post for checking.
Public-version research records carry only a random session code that our server does not link to you or to your install ID, so they cannot be found for one person. Switching research off stops new records at once. Records made before you switched off are kept for the period the research team sets, or without a time limit if it sets none.
If you wrote a post and want data about it removed, use the data requests page linked at the foot of this page.
To check a post, our server sends its text, and nothing about you, to the AI detector currently in use, and sends the result back to your device. The privacy policy says who receives the text and where it is processed.
For English posts, and only if you said yes to research, the text also goes to a company that hosts the models that label tone, topic and hateful or offensive language.
In the public version, our own servers never store the text.