How to Detect Deepfakes: 10 Signs a Video or Photo Is AI-Generated

Deepfakes have gone from research demos to everyday internet content in just a few years. Some of them are harmless fun — face swap memes, movie parodies, creative edits made with tools like our AI face swap. Others are built to deceive: fake celebrity endorsements, fabricated news clips, and voice-cloned scam calls. Either way, the ability to tell real from synthetic is quickly becoming a basic media literacy skill.

The good news: most deepfakes still leave fingerprints. This guide walks through practical deepfake detection — 10 visual and audio signs you can check with your own eyes and ears, plus metadata checks, dedicated deepfake detector tools, and the emerging content-provenance standards (like C2PA) that will make verification easier over time.

Why deepfake detection matters in 2026

Generative AI has crossed a quality threshold. A convincing face swap that once required a research lab and days of GPU time can now be produced in a browser in under a minute. Three trends make detection skills urgent:

• Volume: Millions of synthetic images and videos are published daily, and search interest in deepfake detection keeps climbing.
• Quality: Modern models handle lighting, skin texture, and lip-sync far better than early deepfakes did. (If you're curious about the underlying tech, read how deepfake technology works.)
• Stakes: Deepfakes now show up in fraud, election misinformation, and non-consensual imagery — contexts where being fooled has real consequences.

You don't need a forensics degree to protect yourself. You need a checklist. Here it is.

The 10 signs a video or photo is AI-generated

No single sign proves anything on its own. Detection is probabilistic: the more red flags stack up, the more confident you can be.

1. Unnatural blinking and eye behavior

Early deepfakes barely blinked, because training datasets were full of open-eyed photos. Modern models blink, but often on a strangely regular rhythm, or with eyelids that don't fully close. Also watch the eyes themselves: reflections (catchlights) should match the scene's light sources and appear in both eyes at the same position. Mismatched or missing catchlights are a classic AI tell. Gaze direction that drifts independently of head movement is another.

2. Blurry or shimmering face boundaries

Most face swap pipelines generate the face region and blend it onto the original head. That blend line — along the jaw, hairline, and ears — is where artifacts concentrate. Look for a subtle softness or "halo" around the face while the background stays sharp, flickering at the chin when the head turns, and hair strands that blur or smear where they cross the face boundary. Pause the video and scrub frame by frame; blending errors that are invisible at speed jump out in stills.

3. Inconsistent lighting and shadows

Light is hard to fake coherently. Check whether the direction of light on the face matches the rest of the scene: if the room's window is camera-left but the nose shadow falls left too, something is off. Skin highlights that stay fixed while the head moves, shadows under the chin that don't match the neck, or a face that looks studio-lit inside a dim room are all signs the face was generated separately from its environment.

4. Hands, teeth, ears, and jewelry

Generative models still struggle with high-detail, high-variation structures. The greatest hits:

• Hands: wrong finger counts, fused knuckles, impossible joint angles.
• Teeth: merged into a single white band, or changing count between frames.
• Ears: asymmetric, malformed, or missing an earring that appears in other frames.
• Glasses and jewelry: frames that warp against the face, reflections that don't track the scene.

In video, these zones also flicker — an earring that teleports a few pixels between frames is a strong signal.

5. Skin that's too smooth — or oddly aged

AI-generated faces often have a waxy, airbrushed quality: pores vanish, micro-wrinkles smooth out, and the skin looks uniformly lit like a cosmetics ad. The opposite mismatch appears too: a face whose apparent age doesn't match the hands, neck, or hairline it's attached to. Real skin has texture noise that varies across the face; synthetic skin tends toward suspicious uniformity.

6. Lip-sync and mouth interior problems

The mouth is the hardest region in a talking-head deepfake. Watch for lips that lag the audio by a few frames, mouth shapes that don't match specific sounds (the letters B, M, and P require closed lips — check them), and a mouth interior that looks like a dark smudge instead of teeth and tongue. If you can, mute the audio and just watch the mouth: without sound to anchor your perception, bad lip-sync becomes much easier to spot.

7. Audio that doesn't sound quite human

Voice cloning has its own artifact set, and many fake videos fail on audio before they fail on video. Listen for flat or oddly placed emotional emphasis, missing breath sounds between sentences, metallic or "underwater" timbre on certain phonemes, and unnaturally consistent pacing. Background noise is a tell too: real recordings have room tone that interacts with the voice, while cloned audio often sounds pasted over silence. We cover this in depth in AI voice cloning vs video deepfakes.

8. Glitches at profile angles and occlusions

Face swap models are trained mostly on frontal and near-frontal faces. The moment a subject turns to a full 90-degree profile, or a hand passes in front of the face, weak models break: the face may briefly warp, revert to the original person, or smear across the occluding object. Scrub to the moments where the head turns furthest or something crosses the face — those frames are where the truth leaks out.

9. Warped backgrounds and impossible physics

Generation artifacts aren't limited to faces. Straight lines in the background — door frames, tiles, text on signs — may bend or ripple near the subject. Text is especially revealing: AI-generated or AI-edited scenes frequently contain garbled lettering. In video, watch how hair, clothing, and accessories move; fabric that defies momentum or hair that doesn't respond to head motion suggests synthesis.

10. Context that doesn't add up

The most reliable deepfake detector is still your judgment about context. Where was this first posted? Does any reputable outlet corroborate it? Is a celebrity really announcing a giveaway in a low-resolution vertical video? Reverse-image-search key frames (Google Lens and TinEye both work) to find earlier versions of the footage. A shocking clip that exists only on one anonymous account, cropped to hide watermarks, is suspect no matter how clean it looks. For a primer on the categories of synthetic media, start with what is a deepfake.

Check the metadata

Beyond eyeballing pixels, files carry evidence:

• EXIF data: Photos from real cameras include camera model, lens, exposure settings, and often GPS. Tools like exiftool or web-based viewers expose this in seconds. AI-generated images usually have stripped or minimal EXIF — and some generators actually write identifying software tags.
• File history: Re-encoded, re-uploaded files lose metadata, so absence proves nothing — but presence of a coherent camera trail is meaningful positive evidence.
• Resolution and compression fingerprints: Many generation pipelines output at characteristic resolutions (512, 768, 1024 pixels square) before upscaling. Odd aspect-ratio crops and double-compression artifacts hint at post-processing meant to hide origins.

Treat metadata like the visual signs: one more probabilistic input, not a verdict.

Deepfake detection tools worth knowing

When manual inspection isn't enough, automated deepfake detection can help:

• Hive AI Detector — browser extension and API that classifies images, video, and audio as AI-generated, used by moderation teams.
• Reality Defender — enterprise-grade multi-model detection platform for media, voice, and documents.
• Intel FakeCatcher — analyzes subtle blood-flow color changes (photoplethysmography) in facial video that generators don't reproduce.
• Deepware Scanner — free web scanner aimed at suspicious videos.
• Resemble Detect — focused on synthetic audio and cloned voices.

Two honest caveats. First, detectors are locked in an arms race with generators: a detector that's accurate today can be blind to next quarter's models. Second, they produce false positives on heavily edited but authentic media. Use detector output as one signal among the ten above, never as a final answer — especially before accusing someone of posting a fake.

C2PA, content credentials, and watermarking

The long-term fix isn't better detection — it's provenance. Instead of proving a file is fake, provenance systems prove where a file came from:

• C2PA (Coalition for Content Provenance and Authenticity) defines cryptographically signed "Content Credentials" that record how an image or video was created and edited. Adobe, Microsoft, Google, OpenAI, and major camera makers back it, and newer cameras can sign photos at capture time.
• Invisible watermarking (such as Google's SynthID) embeds machine-readable marks into AI outputs that survive resizing and compression.
• Platform labeling: major social networks now label detected or self-declared AI content, driven in part by regulations like the EU AI Act's transparency requirements.

None of this is universal yet — signatures get stripped, watermarks aren't mandatory everywhere, and open-source models won't always comply (see our overview of open source deepfake tools for why enforcement is hard). But checking for Content Credentials is already worthwhile: an image bearing a valid C2PA manifest from a camera is strong evidence of authenticity.

What to do when you spot a deepfake

Found something synthetic and harmful? A quick protocol:

1. Don't amplify it. Even quote-posting a fake to debunk it boosts its reach.
2. Capture evidence. Screenshot the post, save the URL, note the account and timestamp.
3. Report it to the platform — most now have dedicated synthetic-media reporting categories.
4. Warn the target if the deepfake impersonates someone you can contact.
5. Escalate to law enforcement if it involves fraud, extortion, or intimate imagery — many jurisdictions now criminalize non-consensual deepfakes.

Detection and creation: two sides of the same literacy

It might seem odd for a face swap platform to publish a detection guide, but the two skills are inseparable. People who understand how a deepfake generator actually works — what it does well, where it breaks — are dramatically better at spotting fakes in the wild. Making a consensual video face swap of yourself and then studying the boundary artifacts, lighting mismatches, and lip-sync limits is genuinely instructive.

That's also why consent sits at the center of legitimate use. Creative face swapping — with your own photos, or with the clear permission of everyone depicted — is fun and harmless. Using someone's likeness to deceive, harass, or defraud is not, and in a growing list of countries it's illegal. If you experiment with our free face swap tools, the rule is simple: swap faces you have the right to use, and label synthetic content when there's any chance of confusion.

FAQ: how to detect deepfakes

What is the fastest way to spot a deepfake video?

Scrub through it frame by frame and focus on transitions: head turns, hands crossing the face, and mouth movements on hard consonants. Blending artifacts that are invisible at full speed usually show up in paused frames within a minute of inspection.

Are there free deepfake detector tools?

Yes. Deepware Scanner analyzes suspicious videos for free, Hive offers a free browser extension for images, and reverse image search (Google Lens, TinEye) costs nothing and often finds the original source footage a fake was built from.

Can deepfake detection tools be trusted completely?

No. Detectors trail the newest generators, and they misfire on compressed or heavily edited real footage. Treat any detector score as one input alongside visual inspection, metadata, and source verification — never as standalone proof in either direction.

Do all AI-generated images have watermarks?

No. Some major platforms embed invisible watermarks like SynthID, and C2PA Content Credentials are spreading, but open-source models typically add nothing, and watermarks can sometimes be stripped. Absence of a watermark proves absence of nothing.

Can a deepfake be detected after it's been compressed and re-uploaded?

Partially. Compression destroys metadata and softens pixel-level artifacts, which weakens automated detection. But behavioral signs — lip-sync errors, lighting mismatches, profile-turn glitches — survive re-encoding, which is why manual checks remain essential.

Is it legal to make deepfakes at all?

Creating synthetic media is legal in most places when everyone depicted consents and the content isn't used to deceive or harm. Laws increasingly target specific abuses: fraud, election interference, and non-consensual intimate imagery. Consensual creative edits — like a photo face swap shared with the people in it — remain firmly on the right side of the line.

The bottom line

Deepfake detection in 2026 is a layered game: visual inspection first (eyes, edges, lighting, hands, mouth), then audio, then metadata, then automated tools, then plain source-checking skepticism. No layer is perfect; stacked together, they catch the overwhelming majority of fakes circulating today.

And the deepest defense is understanding. Spend ten minutes seeing how modern face synthesis behaves — its strengths and its telltale failures — and you'll never look at a suspicious clip the same way.