How to Spot AI-Generated Text, Images and Video

There is no single reliable test for whether a piece of text, an image or a video was made with AI. That is the most important thing to understand before you trust any checklist, app or detector score. The honest method is to work through layers of evidence in order — provenance first, platform labels second, the source third, and your eyes last — and to accept that a clean result never proves something is real.

Generative models have improved so fast that the old visual tells are fading. Hands, text and lighting that gave away early AI images are now often correct. That is exactly why provenance standards, invisible watermarks and platform labels now carry more weight than a trained eye alone.

Layer 1: provenance — Content Credentials (C2PA)

The strongest evidence available is a provenance record. Content Credentials are signed records built on an open standard from the Coalition for Content Provenance and Authenticity (C2PA), whose steering committee includes Adobe, Amazon, BBC, Google, Meta, Microsoft, OpenAI, Sony and Truepic. The current specification is version 2.4, dated April 2026. Think of it as a nutrition label for media: it records where a file came from, which tool made it, what edits were made, and whether AI was used.

To check a file, open the free Verify tool at contentcredentials.org/verify and upload it. It accepts common formats including JPEG, PNG, WebP, HEIC, AVIF, TIFF, MP4, MOV, MP3 and WAV. If a credential is present, you will see who signed it, when, with what tool, and whether AI generation or editing was flagged.

The catch is coverage. Content Credentials are still rolling out, so most images you find online carry none. Support is strongest on flagship cameras — Leica, Nikon’s Z9, Z8 and Zf, Sony’s a9 III and a1, and Canon’s newsroom Authenticity Imaging System all sign images at the point of capture — and on files made with conforming AI tools. A missing credential proves nothing: the file may never have had one, or it may have been stripped when a platform re-encoded the image on upload or when someone took a screenshot. Treat a valid credential as real evidence, and an absent one as no evidence at all.

Layer 2: invisible watermarks — SynthID

A watermark hides a signal in the pixels, sound or words themselves, rather than in attached metadata. Google DeepMind’s SynthID embeds imperceptible watermarks in AI-generated images, audio, text and video, and its image and video watermarks are designed to survive cropping, filters, frame-rate changes and lossy compression.

You can check for a SynthID mark in the Gemini app, or through Google’s standalone SynthID Detector, open to the public since October 2026. OpenAI also adds SynthID watermarks through ChatGPT, Codex and the OpenAI API, and runs its own verification page for images and audio made with its tools.

The same limit applies as with credentials: a watermark check finds only marks from tools that add them. Gemini’s check currently recognises only content made by Google AI tools, so a missing watermark means the file was not made or edited by Google AI — it could still have come from another AI system. After many alterations, a watermark may also stop being detected. Check with more than one tool when you can, and never treat “no watermark found” as “this is real.”

Layer 3: platform labels

Platform labels are the layer most readers actually encounter, and the rules are simpler than they look: the trigger everywhere is realism, not the use of AI.

YouTube requires creators to disclose realistic AI-generated or altered content — a real person appearing to say or do something they did not, footage of a real event or place that has been altered, a realistic scene that never happened, or AI music as the main focus of a video. Filters, captions, AI-written scripts and a clone of your own voice do not need disclosure. Creators who repeatedly fail to disclose can have a label applied for them, or face removal or suspension from the YouTube Partner Program.

TikTok requires a label on AI-generated or significantly edited content that shows realistic-looking scenes or people, under Community Guidelines in force since 24 September 2026. Small edits, artistic styles such as anime, and generic text-to-speech narration are exempt. Unlabelled content may be removed, restricted or labelled by TikTok.

Meta — Facebook, Instagram and Threads — may require an AI label when content has photorealistic video or realistic-sounding audio that was digitally created or altered, including with AI. That requirement does not apply to still images, though Meta may label them itself. From 2 August 2026, the EU AI Act’s Article 50 also requires providers of generative systems to mark synthetic output so it is machine-readable, and requires people who publish realistic AI content to disclose it to viewers.

A label is useful evidence, but its absence is not. Creators forget, platforms miss things, and a file stripped of its provenance on upload may never be labelled at all.

Layer 4: your eyes — the visual tells that still work, for now

Look at the details, but treat everything in this section as a clue, not a verdict. Generators still frequently mangle text inside an image: signage, book spines, number plates and logos that dissolve into near-gibberish on close inspection. Hands and fingers are better than they were, but still worth checking. Lighting and shadows that do not match the scene, reflections that show something the photo does not, and backgrounds that blur into nonsense at the edges are all worth a second look.

For video, watch the face and mouth: lip movements that drift out of sync, teeth that blur, blinking that looks unnatural, and a face that wobbles when the head turns. For audio, listen for a voice that is too smooth — no breaths, no hesitations, no background noise that changes naturally — and for a familiar voice saying something the person has never been reported saying anywhere else.

Every one of these tells is a moving target, and a checklist that worked in 2024 catches fewer images in 2026. If the image looks perfect, that tells you nothing at all.

Spotting AI-generated text

Text is the hardest of the three to spot, because good human writing and good AI writing now overlap almost completely. A few patterns are still worth noticing: writing that is fluent but says very little, with the same structure repeated paragraph after paragraph; confident statements with no source, or with citations that do not exist when you check them; and a tone that never varies, never hedges, and never admits uncertainty. None of these proves AI authorship — plenty of human writing is vague and repetitive too.

AI detectors estimate how AI-like a piece of text is from its characteristics, but their results are only estimates, and they make mistakes. No accuracy figure is quoted here, because published figures disagree sharply depending on what text was tested and when. What independent research does agree on is the direction: detectors can miss AI-written text, especially when it has been lightly edited or paraphrased, and they can falsely flag human writing — a Stanford study found detectors falsely flagged an average of 61.3 percent of 91 TOEFL essays written by non-native English speakers. Turnitin itself says its AI score should not be used as the sole basis for action against a student, and at least a dozen universities have disabled its AI detection entirely.

So do not use a detector score as proof, in either direction. A high score does not prove a student cheated, and a low score does not prove a text is human. Use detectors, if at all, as a prompt to ask questions and check sources — never as a verdict.

Layer 5: the source — the check that matters most

When provenance, watermarks and labels all come back empty, the source is usually more informative than the file itself. Ask: who posted this, and where did they get it? Does the same image or claim appear on a reputable news site, or only on anonymous accounts? Does a reverse image search show the picture in an earlier context, before the event it supposedly shows? Does the account that posted it have a history, or was it created last week?

For a viral claim, the method in How to Check a Viral News Claim Before You Share It applies directly: check the original source, look for independent confirmation from a second reputable outlet, and be most suspicious of the content that makes you angriest or most certain — that is the content most likely to have been made to manipulate you.

Frequently asked questions

Can I tell if an image is AI-generated just by looking at it?
Sometimes, but not reliably. Text inside images, hands, lighting and reflections still give AI images away in some cases, but newer models correct these errors quickly. Use provenance checks and watermarks first, and treat visual tells as clues, not proof.

What are Content Credentials?
Signed provenance records under the C2PA open standard, showing where a file came from and whether AI was used. Check them free at contentcredentials.org/verify. Most files online do not carry them yet, and they can be stripped on upload, so an absent credential proves nothing.

Are AI detectors accurate?
Not reliably enough to treat a score as proof. Independent studies show detectors miss AI text — especially edited or paraphrased text — and falsely flag human writing, with non-native English speakers affected most. Vendors’ own accuracy claims are measured on their own tests and disagree with independent results.