How to Tell if an Image or Video Is AI-Generated

Quick answer: To tell if an image or video is AI-generated, check several signals together: where it came from, any AI labels or Content Credentials, invisible watermarks such as Google’s SynthID, a reverse image search, and visual clues like warped hands, garbled text and inconsistent lighting. AI detector tools can help, but no single clue or detector score proves anything on its own.

Why spotting AI images and videos is harder now

Early AI images were easy to catch: six-fingered hands, melted jewelry and nonsense text on every sign. Current image models and video generators such as Google’s Veo, OpenAI’s image tools, Kling and others produce results that often pass a quick look. Realistic AI video, including fake news clips and cloned voices, now spreads during breaking news and on short-video platforms within minutes.

The practical result is that “look closely at the hands” is no longer enough. Reliable checking in 2026 means combining source checks, provenance signals, visual inspection and detector tools, and being comfortable saying “unverified” when the evidence is unclear.

The 5-step method to check any image or video

1. Check the source: find who posted it first. A verified news outlet, the original photographer or an official account is more trustworthy than an anonymous repost.
2. Look for labels and credentials: check for “Made with AI” or “AI info” labels on the platform, and inspect Content Credentials if the file carries them.
3. Check for invisible watermarks: upload the file to a watermark checker, such as asking Gemini whether the content was made or edited with Google AI (SynthID), or a vendor’s own verification tool.
4. Run a reverse search: use Google Lens, Bing Visual Search or TinEye to find earlier or original versions, and for video, search key frames.
5. Inspect the details: look for the visual and audio clues described below, then use a detector tool as a final extra signal.

If several independent checks point the same way, your confidence is justified. If they conflict, treat the content as unverified.

Provenance signals: labels, Content Credentials and watermarks

Platform AI labels

Major platforms including Meta (Facebook, Instagram, Threads), YouTube and TikTok label some AI-generated or AI-edited content, either because the creator disclosed it or because the platform detected industry-standard signals. A label is strong evidence; the absence of a label proves nothing, because many AI files carry no detectable signal and many creators do not disclose.

C2PA Content Credentials

Content Credentials, based on the C2PA standard, are signed metadata attached to a file that can record which tool created it, what edits were made and whether generative AI was involved. Many AI tools and some cameras and editing apps add them. You can inspect them with the Content Credentials verify tool at contentcredentials.org or with compatible viewers.

Their weakness: metadata is fragile. Screenshots, re-encoding and many social uploads strip it. So missing credentials do not mean an image is real.

Invisible watermarks such as SynthID

Google’s SynthID embeds an invisible watermark directly in the pixels of images and frames of video (and in audio and text in some products). Because it lives in the content itself, it is designed to survive common changes such as cropping, compression and screenshots better than metadata does. Google has added SynthID checking to the Gemini app, where you can upload a file and ask whether it was made or edited with Google AI. OpenAI has also announced SynthID watermarking alongside Content Credentials for its images, plus its own verification tool.

Watermark checks only detect content from tools that use that watermark. A negative result means “no watermark from this system was found”, not “this is real”.

Signal	What it shows	Main limitation
Platform AI label	Creator disclosure or detected AI signals	Many AI posts are unlabeled
C2PA Content Credentials	Signed history of tools and edits	Stripped by screenshots and many re-uploads
SynthID and other watermarks	Content made with a specific company’s tools	Only detects participating tools
EXIF metadata	Camera model, date, software	Easy to remove or fake
Reverse image search	Earlier copies and original context	New content has no history yet

Visual clues in AI-generated images

Use these as reasons to investigate further, not as final proof. Real photos have flaws too, and the best AI images show none of these.

Hands, teeth, ears and bodies

Count fingers, check how fingers wrap around objects, and look for joints that bend the wrong way. Teeth that blend into one strip, ears that do not match and limbs that disappear behind objects are classic signs. Newer models get these right far more often, so a correct hand is not evidence of a real photo.

Text, logos and signs

Zoom in on shop signs, labels, license plates, book spines and T-shirt prints. AI often produces letters that look almost right but spell nothing, or text that changes style halfway through a word. Brand logos may be slightly distorted.

Lighting, shadows and reflections

Check whether shadows fall in the same direction as the light source, whether reflections in mirrors, glasses and windows match the scene, and whether light on the face matches the background.

Backgrounds and repeated details

Background people with smeared faces, fences or railings that merge into each other, windows that do not line up, and patterns that repeat too perfectly are all worth a closer look.

Skin, hair and texture

AI portraits often have skin that is too smooth, “plastic” or evenly lit, and hair strands that melt into the background or clothing. Pores and fine texture may look painted rather than photographed.

Physics and context

Ask whether the scene makes sense: objects floating or merging, clothing that fuses with skin, food with impossible shapes, or a famous landmark with the wrong details. Also ask whether the event itself is plausible and reported anywhere credible.

How to tell if a video is AI-generated

Video adds motion and sound, which gives more chances for errors, but also more compression that hides them. Watch at normal speed, then slowly, then frame by frame.

• Short clips and hard cuts: many AI videos are made of clips of a few seconds each; watch for jumps where the scene or a person’s appearance changes.
• Morphing objects: items that change shape, appear or vanish, or hands that merge with objects as they move.
• Unstable details: text, jewelry, buttons and background people that shift between frames.
• Unnatural movement: gliding walks, floaty physics, hair or fabric that moves oddly, crowds that move in unison.
• Faces and lip sync: mouths that do not close on sounds like “b”, “m” and “p”, blinking that looks off, or a face that slightly “slides” over the head.
• Audio clues: voices that sound flat or oddly clean for the setting, missing breaths, or background sound that does not match the scene.
• Watermark remnants: blurred or cropped corners where a generator watermark may have been removed.

Research systems also look for signals humans cannot see, such as the tiny color changes in skin caused by blood flow (the idea behind Intel’s FakeCatcher). These signals weaken badly in low-resolution or heavily compressed video, such as clips forwarded many times in messaging apps.

How to check a video step by step

1. Find the earliest upload: search the caption, key phrases and the account name.
2. Take screenshots of several key frames and run a reverse image search on each.
3. Check for platform AI labels and the uploader’s description.
4. Upload the file to a watermark checker, such as asking the Gemini app about SynthID.
5. Compare with trusted news coverage: real major events usually have many independent angles.

AI image and video detector tools

AI detectors are classifiers trained to spot statistical patterns left by generators. Examples include Hive Moderation, AI or Not, Sightengine, and Illuminarty, alongside company-specific checkers such as SynthID in Gemini and OpenAI’s verification tool.

Tool type	How it works	Best use
Watermark checkers	Detect an embedded signal from a specific company’s tools	Confirming content from participating generators
Content Credentials viewers	Read signed C2PA metadata	Seeing the tool and edit history when metadata survives
General AI detectors	Classify pixels as likely AI or likely real	A quick extra signal, not a verdict
Reverse image search	Find earlier copies and context	Tracing origin and spotting recycled images
Forensic tools	Examine metadata, compression and noise patterns	Deeper checks by journalists and investigators

How accurate are AI detectors?

Accuracy varies a lot by tool, by generator and by how the file was edited. Detectors can do well on images from the generators they were trained on and poorly on new models. Compression, resizing, screenshots and filters reduce accuracy further. False positives also happen: heavily edited or filtered real photos can be flagged as AI. Treat a detector’s percentage as a probability estimate, and never accuse someone on the basis of one score.

What about AI text detectors?

Many people searching for AI detection are really asking about text: essays, homework and articles. Text detectors such as Turnitin’s AI writing detection, GPTZero and similar tools estimate how likely a passage is to be machine-written. They face the same core problem as image detectors, only worse: there is no reliable visual trace in text, and edited or paraphrased AI text is hard to classify.

• False positives are real: formal, simple or formulaic writing, and writing by non-native English speakers, can be wrongly flagged.
• Scores are not proof: most vendors themselves advise using results as a starting point for a conversation, not as evidence of misconduct.
• Learning platforms vary: learning management systems such as Canvas do not decide AI use on their own; schools usually connect third-party tools like Turnitin, and policies differ by institution.

If you write with AI assistance, the honest route is to follow your school’s or employer’s rules and disclose AI use when required. Keep drafts and version history, which is the best evidence of your own writing process. For content creators, our guide on making AI writing sound human focuses on editing for quality and readers, not on hiding AI use from people who have a right to know.

Deepfake audio and voice clones

AI can now clone a voice from a short sample, which is used in scams where a caller pretends to be a family member, boss or bank. Clues include flat emotion, odd pauses, and pressure to act fast or pay by unusual methods. The best defense is procedural, not technical:

• Hang up and call back on a number you already know.
• Agree a family “safe word” for emergencies.
• Never send money, gift cards or codes because of an urgent call or voice note alone.
• Report scams to your local authorities; in the US, the FTC collects fraud reports at ReportFraud.ftc.gov.

Real photos that look fake (and fakes that look real)

Detection works in both directions. Real photos get dismissed as “AI” because of heavy phone processing, beauty filters, HDR or unusual lighting. This “liar’s dividend” lets people wave away genuine evidence by claiming it is fake. At the same time, a well-made AI image may pass every visual check.

That is why source and context matter more than pixels. Ask who took it, when, where, and whether other independent sources show the same event from different angles.

What to do if you find AI-generated or fake content

1. Do not share it further: even sharing to mock it spreads it.
2. Check fact-checkers: search the claim on established fact-checking sites and reputable news outlets.
3. Report it on the platform: use the report option for misleading or manipulated media, impersonation or scams.
4. Warn people who shared it: politely, with a link to evidence.
5. Protect yourself if it targets you: save evidence (links, screenshots, dates), report it to the platform and, for harassment, threats or intimate images, contact the relevant authorities or support organizations.

For creators: how to label your own AI images and videos

If you publish AI visuals on a blog or social media, being transparent builds trust and keeps you within platform rules.

• Use platform labels: turn on the AI disclosure setting when you upload realistic AI or altered content.
• Keep watermarks and credentials: do not strip SynthID, Content Credentials or visible marks.
• Caption clearly: a short “Image created with AI” caption or alt text note is enough for illustrations.
• Avoid realistic fakes of real people and events: they can mislead and may break laws and platform rules.
• Fix obvious errors: distorted hands or text make content look careless; see how to fix AI image problems.

If you create video, our guide to free AI video generators covers watermarks and commercial use in more detail.

Two example checks, step by step

Example 1: a dramatic “breaking news” photo

Imagine a photo of a flooded city landmark spreading on social media with no credit. You reverse search it and find no copies older than a few hours. No major news outlet has the image or any similar angles, although a real flood at a famous landmark would be photographed by many people. Zooming in shows a street sign with unreadable letters and people in the background with blurred, merged faces. A detector gives a high AI score, and the Gemini app finds no Google watermark. Conclusion: very likely AI-generated, even without a watermark, because the source, context and visual checks all point the same way. Do not share it.

Example 2: a product photo on an online store

A small shop’s product photo looks almost too perfect. Content Credentials show it was edited in a photo editor with a generative fill step used to extend the background. The product itself comes from an original camera photo. Conclusion: a real photo with AI-edited parts. That is common and often fine, as long as the product is shown accurately; what matters is whether the item you receive looks like the picture.

Quick checklist: is this image or video AI-generated?

Check	Suggests AI	Suggests real (but not proof)
Source	Anonymous repost, no original	Named photographer or trusted outlet
Labels and credentials	AI label, credentials naming an AI tool	Camera credentials with an intact edit history
Watermark check	SynthID or vendor watermark found	No watermark (proves little)
Reverse search	No history, or first seen on AI art accounts	Older copies with consistent context
Visual details	Garbled text, warped hands, lighting conflicts	Consistent details under zoom
Other angles	Nobody else captured a major event	Many independent photos and videos
Detector score	Consistently high across tools	Consistently low across tools

How to teach others to spot AI content

Parents, teachers and team leads can make this a habit rather than a one-time lesson. Show examples of real and AI images side by side and ask people to explain their reasoning. Teach the “pause before sharing” rule for emotional or shocking content, and focus on the source-first method rather than a list of visual tells that will go out of date as models improve.

Where AI detection is heading

Detection is shifting from “spot the mistake” to “check the provenance”. More AI companies are adding invisible watermarks and Content Credentials to their outputs, more cameras and phones can sign real photos at the moment of capture, and more platforms read these signals automatically to apply labels. Laws in some regions, such as the EU AI Act’s transparency rules, also push providers to mark AI-generated content in machine-readable ways.

None of this will catch everything. Open-source models can run without watermarks, and determined bad actors will try to strip signals. The most durable skill is the habit this guide teaches: check the source, look for independent confirmation, and treat surprising, emotional content as unverified until proven otherwise.

Frequently asked questions

How can you tell if an image is AI-generated?

Check the source, look for AI labels and Content Credentials, test for watermarks such as SynthID, run a reverse image search and zoom in on hands, text, lighting and backgrounds. Use a detector tool as one extra signal. Several independent signals together are far more reliable than any single clue.

Is there a free tool to check if an image is AI?

Yes. You can upload an image to the Gemini app and ask whether it was made or edited with Google AI, inspect Content Credentials at contentcredentials.org, run reverse searches with Google Lens or TinEye, and try free tiers of detectors such as Hive or AI or Not. Results are indications, not proof.

How do I know if a video is AI-generated?

Look for short clips joined by cuts, morphing objects, unstable text and background details, odd movement and poor lip sync. Then check the earliest upload, platform labels and watermarks, and whether trusted sources show the same event.

What is SynthID?

SynthID is Google DeepMind’s invisible watermarking technology. It embeds a signal in AI-generated images, video, audio and text from Google tools, designed to survive common edits such as cropping and compression. You can check for it in the Gemini app.

What are Content Credentials?

Content Credentials are signed metadata based on the C2PA standard. They can record which tool created a file, what edits were made and whether AI was used. Screenshots and many re-uploads remove them, so missing credentials do not prove a file is real.

Are AI detectors accurate?

Accuracy varies widely. Detectors can miss images from new generators, struggle with compressed or edited files, and wrongly flag real photos. Use them as one signal among several, and never as the only evidence against someone.

Can AI-generated images have no watermark?

Yes. Many tools do not add watermarks, open-source models can run without them, and some files are edited or screenshotted. A missing watermark only means none was found, not that the image is real.

Does Canvas have AI detection?

Canvas itself is a learning management system and schools typically add third-party tools such as Turnitin for AI writing detection. Whether your course uses one depends on your institution, so check your syllabus or ask your instructor.