"Is it AI?" is a free set of tools that lets anyone check whether a photo or video was made by AI, and whether a link, text message, email or QR code is safe. As generative AI and online scams grow more sophisticated, our goal is to help you judge authenticity yourself, without specialist knowledge.
Free with no sign-up, and the check runs inside your browser, so what you enter never leaves your device. The result is a signal to help you decide, not proof for money or legal decisions.
"Is it AI?" is a free set of tools that lets anyone check whether a photo or video was made by AI, and whether a link, text message, email or QR code is safe. As generative AI and online scams grow more sophisticated, our goal is to help you judge authenticity yourself, without specialist knowledge.
All analysis runs inside your browser. Images and text are computed on your device rather than sent to a server, so even sensitive material is safe to check.
AI-generated image and deepfake detection, AI-written text detection, fake-review and bot-account detection, phishing-link / smishing-text / QR / email-spoofing checks, photo metadata (location) removal, and content-provenance (C2PA) inspection — all the tools related to online authenticity and safety in one place.
Image detection uses an ensemble of two detection models chosen by real-world benchmarks, together with the file's provenance data. The text, link and message tools compute risk signals with style and structure heuristics, without any AI model. Results are given as probability/risk indicators, and each tool page explains both the principle and its limits.
Every result is a reference indicator and is not 100% accurate. The newest generators and sophisticated scams can cause false positives and false negatives. Do not use any result as sole evidence for important decisions such as academic, hiring or legal matters. Use it alongside human review.
It is a project built and run by one person. There are no accounts and no paid tiers; running costs come from ads on the page. There is deliberately no structure where paying unlocks a result or buys a more accurate one. That arrangement creates an incentive to keep results vague.
Tools were added one at a time as needs came up. It began with image detection and grew to text, links, messages, email, QR codes and calls. What they share is that they all run inside the browser.
Through a public gate of our own. A dataset mixing real photographs and AI images is measured separately at original quality and as social-media re-compressions, and the results are published in the accuracy report alongside the raw data. Conditions where the score drops, such as re-compressed copies, are published too rather than quietly dropped.
The number watched most closely is not overall accuracy but how often a genuine photo is confidently called AI. That is the error that does the most damage to the person using the tool.
Images, video, text and audio do not leave your device by default. The exceptions are explicit opt-ins, and features needing edge computation — call transcription, for instance — discard the input after processing. With no accounts, there is no way to tie usage back to a person.
Anonymous aggregates are collected separately: how often each verdict occurs and similar counts, with no IP addresses, cookies or device identifiers stored. Your check history lives only on your device and goes when you clear your browser data.
It does not stand in for legal or administrative judgement. No result here is designed to serve as evidence, which is why every tool page carries worked examples of when the tool gets it wrong.
It does not identify people. There is no feature to trace a face to a name or work out who runs an account, and it does not file takedowns or reports on your behalf. Those belong to the platforms and to law enforcement; this site stops at helping you form a judgement.
Sending files to a server would be far easier: bigger models become available and device performance stops mattering. Running in the browser was chosen because of what these files tend to be — a photo you check because you are suspicious of it is usually a photo you would rather not hand to anyone.
The cost is real. The first check spends time downloading models, speed varies with your device, and the largest current models — which only run server-side — are off the table. It trades away some accuracy for the guarantee that the file never leaves. This site took that trade.
The correct/incorrect feedback on the result screen is collected anonymously, and cases reported as misjudged appear as a count in the hard-case queue of the public accuracy report. Knowing the conditions under which it fails is what sets the direction for the next improvement, which makes a wrong signal more useful than a right one.
Telling us which input produced which result helps a great deal. You do not need to send the file: a description by email from the contact page is enough.