Vibe code detector
Websites and apps that feel prompted into existence — familiar layouts, generated copy, demo-day confidence.
Upload an image, a short video, or paste a link. One honest answer, every time.
Built by MIT engineers. Probably.
Drop a file, or browse
Images & short videos · Max 100 MB“This is DEFINITELY it. The first lab was too cautious about those pixels.”
“Absolutely not. I reviewed the same evidence for 0.4 seconds and reached the opposite energy.”
“You are both describing the same file. Please just sign the form.”
The panel disagrees with everything about the first verdict — except the verdict.
Three steps. One outcome.
An image, clip, or link — previewed in your browser. No server ever sees it.
Fingerprinting, extraction and cross-referencing across every layer — done properly, in full.
Stamped, sealed, and non-refundable.
Five checks. One consistent answer.
Websites and apps that feel prompted into existence — familiar layouts, generated copy, demo-day confidence.
Photos and clips that might be AI-generated — plausible lighting, almost-correct hands, text that nearly says something.
Video of a person who may not have said, done, or blinked any of that.
Posts and essays with the classic tells — the em-dash parade and a conclusion that restates the intro.
Photos retouched, composited, or filtered into a parallel timeline. Shadows will be cross-examined.
The joke verdict is free. When the answer actually matters, do this instead — it's what journalists and fact-checkers do.
Find the earliest version you can. An original upload from an identifiable account beats a screenshot of a screenshot every time. If nobody can say where it came from, lower your confidence — don't raise it.
Many cameras, phones and AI tools now attach Content Credentials (C2PA) — a signed record of how a file was made and edited. An unbroken credential chain is strong evidence. No credentials proves nothing either way; most of the internet doesn't have them yet.
Run the image — or a key frame from the video — through a reverse image search. Fakes are very often real media from an older, different event, and the search usually surfaces the original context in seconds.
Check whether independent, reputable reporting describes the same event. An extraordinary clip with exactly one source is a claim, not a fact — wait for a second source before passing it on.
One endpoint. Zero machine learning. Not a real server.
# This goes nowhere. There is no server. curl -X POST https://api.probably.example/v1/detect \ -d '{"check":"ai-generated", "input":"that-photo.jpg"}'
{
"verdict": "PROBABLY",
"confidence": "probably",
"evidence_seen": false,
"appeals": "denied"
}
Errors: 400 — you sent nothing, still Probably · 404 — endpoint doesn't exist · 429 — slow down, still Probably.
Joke first — then the straight answer underneath.
No. Probably is a parody with excellent posture. It runs no machine-learning model, employs no server, and has never once examined your file with anything but total confidence. Every image, video, and link receives the same verdict — Probably — delivered like it was measured in a lab. It's a joke about how confidently real detectors present results that are, at best, a shrug with a dashboard.
Straight answer: No. It performs no detection, runs no machine-learning model, uploads nothing, and returns the same verdict for every input.
Choose “AI-generated” above, upload the image or short video, and enjoy the full forensic theatre — fingerprinting, cross-referencing, the works. Then, if the answer actually matters, ignore us completely and use provenance tools: content credentials, reverse image search, the original source. We are the warm-up act, not the expert witness.
Straight answer: This site can't tell you. For a real check, look for content credentials / C2PA provenance, run a reverse image search, and verify the original source and context. No detector score is conclusive on its own.
Vibe coding means building software mostly by prompting an AI coding assistant and accepting whatever it produces, guided by vibes rather than review. Telltale signs: familiar layouts, generated copy, comments that say “TODO: fix later,” and features that work beautifully exactly once — during the demo, in front of everyone.
Straight answer: "Vibe coding" is a real term for building software mainly by prompting AI coding tools and iterating on the result, rather than writing and reviewing every line by hand.
Probably — which is to say, it will say Probably, with a straight face and a stamp. Real deepfake detection studies facial inconsistencies, unnatural blinking, and audio artefacts, and even the specialist tools argue about it afterwards. Ours skipped the studying and the arguing and went directly to the verdict.
Straight answer: No, it can't. Real deepfake detection examines facial, audio and compression artefacts plus provenance — and even then it's probabilistic. Verify anything important against the original source and trusted reporting.
No — and this is the one answer here with no jokes in it, because it matters. Everything happens in your browser: files are previewed locally, links are never fetched, and nothing is uploaded or stored. Closing the tab forgets everything. The lab has no memory and, frankly, no filing cabinet.
Straight answer: Correct as stated. Files are previewed locally in your browser, links are read as text and never fetched, and nothing is uploaded or stored. The case log exists only in this tab's memory.
A fictional panel reviews the same evidence, disputes the first lab's methods, motives, and general vibe at remarkable volume — and then signs the exact same verdict: also Probably, but louder, and in writing.
Straight answer: It's a scripted joke panel built into this page. No second model exists, nothing is consulted, and it always lands on the same verdict.
No. There is no server, no endpoint, and no key — the API section is parody documentation and an in-browser playground arguing with itself, beautifully. It returns the only response ever documented: {"verdict": "PROBABLY"}.
Straight answer: No API exists. The documentation and playground are parody and run entirely in your browser.
Because “Probably” is the most honest word in detection, and the only one our lab is licensed to use. From the outside, the truthful answer is usually a shrug with good posture. We simply gave the shrug a stamp, a case number, and an extremely serious forensic process it did not earn.
Straight answer: Because the verdict is fixed in the code. The analysis animation is theatre — the result never depends on your file, your link, or which check you selected; those only change the wording and the joke metrics.
You can trust it to be a number. Real detectors speak in percentages because certainty isn't available: the score is a model's confidence in its own guess, and the guess errs in both directions — real work flagged as AI, AI work waved through. We just skip the percentage and keep the honesty.
Straight answer: Treat any detector score as a hint, never as proof. Testing consistently finds meaningful false-positive and false-negative rates, and scores differ between tools on the same file. Corroborate with provenance and the original source.
With the unglamorous steps in the guide above — source, provenance, reverse search, corroboration. Notably, none of them involves asking us, which is rather the point of verification.
Straight answer: Find the earliest source, check for Content Credentials (C2PA), reverse-search the image or key frames, and confirm the event with an independent reputable source. If it still matters and still isn't settled, treat it as unconfirmed.
Probably is a parody of the detector era — tools that promise certainty about things that are genuinely hard to know. Here the machinery is honest about being machinery, and the conclusion is the one word they're all quietly standing on: Probably. It's free, works on images, short videos, and links, and — in the spirit of its subject — is itself cheerfully vibe coded.
Built by MIT engineers. Probably. No real university was consulted, credited, or awake for any of it.
Probably is a parody site. Files you choose are previewed in your browser, and links you paste are read as text.
Last updated: 25 October 2026.
Verdicts are entertainment, not evidence. Don't use one to accuse anyone of using AI or faking media — if it matters, verify with primary sources and provenance tools.
By using this site you agree to these terms. Probably is a parody provided for entertainment only: it performs no detection, its verdicts are fixed in advance, and nothing on this page is professional, legal, or evidentiary advice. Do not rely on a verdict for any decision about a person, a piece of media, or anything else that matters.
Your content. Files and links stay in your browser and are never sent to us, so you keep every right you already have in them and grant us nothing. Only submit material you have the right to use, and don't use this site to harass, defame, or mislead anyone about real people or real media.
No affiliation. Probably is not affiliated with, sponsored, or endorsed by MIT or any other university, company, or institution — the line in the hero is a joke, and it stays a joke here in the small print.
Liability. The site is provided “as is,” without warranties of any kind. To the fullest extent the law allows, we accept no liability for anything done in reliance on a verdict — the verdict was always going to be Probably. These terms may change; the version on this page is the one that applies.