Is This AI? How to Check Any Text, Image or Video

Humans can still win this game: in a 2025 ACL study, expert annotators who frequently use ChatGPT identified AI-generated text with 99.3% accuracy, resisting evasion tricks that beat automated detectors. The complication is that those were experts judging prose, which is the easy case. Images and video are where confidence goes to die: Bellingcat watched a leading image detector miss 7 of 10 AI images after social-media-level compression back in 2023.

Answering "is this AI?" honestly takes three moves: check provenance first (original source, C2PA Content Credentials, metadata), run a detector built for that specific format, and read its confidence level rather than just its score. Text, images and video each fail differently, and sometimes the only defensible verdict is "unverifiable," which is itself an answer.

Is this AI? Run the two universal checks first

Whatever landed in front of you, start with the same two questions: where did this come from, and does it carry receipts?

The source check is plain old journalism. Who posted it first, and when? Does a higher-quality original exist somewhere? Does the account have a history, or did it appear last Tuesday with breaking-news footage and a username full of digits? For anything consequential, finding the earliest version matters more than any scan you run later.

The receipts check is newer. C2PA Content Credentials are provenance labels embedded at the moment of creation: OpenAI has attached them to image outputs since February 2024, Adobe Firefly and Microsoft's Bing and Designer tools embed them, and Google's Nano Banana image models joined in 2026. Regulation is pushing the same direction, since the EU AI Act's Article 50 obligations apply from August 2, 2026 and require AI-generated content to be marked in a machine-readable way. One catch, and it's a big one: platforms routinely strip this metadata on upload. Credentials that survive tell you a lot. Their absence tells you almost nothing.

With those two checks done, the path splits by format:

If it's text

Run the scan first, but read the result like an adult. A percentage is a probability estimate, not a verdict, and the mechanics behind it (word-choice predictability, rhythm patterns) are worth two minutes of your time; our guide to how AI detectors work covers them without the math degree. Short samples are the classic trap. Below a couple hundred words, most detectors are guessing, and honest ones say so.

AI Detector 360's free scanner takes up to 5,000 characters with no sign-up and returns a multi-engine score, a sentence-level heatmap and an explicit confidence label. That last part is the feature that keeps you out of trouble: a low-confidence 80% is a lead, not a finding.

Then read it yourself. Uniform sentence rhythm, hedge-everything framing, examples from nowhere, a suspicious absence of lived detail. If you suspect a specific model, that question has its own answer; is ChatGPT detectable walks through what unedited chatbot output looks like to a detector and how much paraphrasing changes the picture. The RAID benchmark, built on more than 10 million documents and 12 adversarial attacks, found commercial detectors degrade sharply under paraphrasing, so a clean scan on heavily edited text means less than it seems.

When the scan and your own read disagree, escalate to process evidence instead of re-running scans until you like the answer. Real writing leaves a trail: drafts, version history, notes, the half-formed outline from Tuesday. Generated text arrives fully grown, with no childhood photos. And for longer material, a full account trades credits for volume at one credit per 100 words, which starts to matter once you're checking reports rather than paragraphs.

If it's an image

Images give you three independent evidence streams, and you should tap them in order of cheapness.

Reverse search first: if the "photo" of tonight's disaster appeared on a stock site in 2021, you're done. Metadata second: upload the original file, not a screenshot, to a checker that reads C2PA and EXIF data. Classifier third: a trained model scores the pixels themselves, and a good one also estimates which generator likely made the picture, the way our AI image detector pairs its score with likely-generator attribution and a provenance readout in the same report.

Your own eyes still matter, because compression is brutal to automated tools. Bellingcat's 2023 test saw a leading detector miss 7 of 10 AI images after social-media-level compression, and the arms race hasn't paused since. Zoom into hands, background text, reflections and hair boundaries; the full checklist lives in our guide to telling if an image is AI-generated. Watermarks are a special case: Google's SynthID marks every Gemini-generated image, but there is no public third-party API, so only Google's own tools can confirm it.

If it's a video

Video is the highest-stakes format and the least forgiving of shortcuts. Watch once at normal speed, then again slowed down, and give the audio its own pass with your eyes closed. Face boundaries during head turns, lip-sync drift, lighting that disagrees with the room: the tells are specific enough that a checklist genuinely helps, and our deepfake spotting checklist is the one to keep open.

Provenance counts double for footage. Ask where the clip premiered, whether a longer cut exists, and who benefits from you believing it, because AI-generated video mostly spreads by re-upload, cropped and re-encoded until forensic traces fade. An original upload from a known account settles more than any scanner will.

A frame-by-frame tool earns its keep here because single-frame artifacts slip past human eyes at full speed. AI Detector 360's video scan builds a frame-by-frame timeline showing which segments look generated, which tells you far more than one number stretched over a whole clip.

When the honest answer is "can't tell"

Some content defeats every check: a meme screenshotted through four apps, a 60-word caption, a photo stripped of metadata and compressed into soup. At that point a 90% score is a shrug dressed as a statistic, and treating it as proof is how people get hurt.

What you do next depends on stakes. For idle curiosity, let it go. For anything touching money, grades, reputations or news, treat "unverifiable" as actionable: ask the author or poster for process evidence such as drafts, originals or version history, hunt for the earliest posting, and wait before you share or accuse. Machine-readable labels will cover more of the internet as Article 50 obligations bite and C2PA adoption spreads, but nothing retroactively labels the internet we already have.

Two examples to calibrate on. A take-home essay scoring 85% at high confidence, with no drafts anywhere and a voice shift from past work, justifies a conversation built on process questions. A political video scoring 55% at low confidence from a re-encoded repost justifies exactly one thing: finding the original before anyone shares it further.

The workflow, compressed: source, credentials, format-specific scan, human read, then a verdict with its confidence attached. Run it in that order and you'll be right far more often than anyone eyeballing their feed, and you'll know exactly how much to trust your own answer.