How to Detect AI-Generated Images and Deepfakes in 2026

· Updated March 28, 2026 · Anuranjan Vikas · 10 min read
deepfakesai-detectionguide

To detect AI-generated images and deepfakes, look for visual artifacts like distorted hands, too-perfect skin, melting backgrounds, and nonsense text. For reliable detection, use an AI tool like Kaval that analyzes pixel-level patterns invisible to the human eye, combined with reverse image search to check if the image existed before the claimed event.

A finance worker in Hong Kong wired $25 million to fraudsters after a video call where every other person on the screen — including the company’s CFO — was a deepfake. He thought the faces were real. They weren’t.

That was 2024. Things have gotten worse since then.

AI image generation has improved so fast that telling synthetic content from real photographs is genuinely hard now. We’re not talking about obvious glitchy stuff anymore. Political campaigns, romance scams, fake product reviews, fabricated news events — AI-generated images are everywhere. And every few months, a new model comes out that makes last year’s fakes look amateur.

So how do you tell what’s real? There are still visual cues that give AI images away, and there are tools that catch what your eyes can’t. Let’s go through both.

What Are Deepfakes and AI-Generated Images?

Quick definitions, because people mix these up.

AI-generated images are pictures created entirely by AI models like Midjourney, DALL-E 3, Stable Diffusion, and Flux. You type a prompt, you get an image. No camera involved. No real scene photographed.

Deepfakes are more specific: AI-manipulated media where a real person’s face, voice, or body is altered or replaced. Think face-swapping someone into a video, making a politician say something they never said, or cloning someone’s voice from a few seconds of audio.

All deepfakes are AI-generated. Not all AI-generated images are deepfakes. A Midjourney landscape isn’t a deepfake. A video putting words in a real person’s mouth is.

Why should you care?

  • Misinformation. Fake images of events that never happened spread faster than corrections. During elections, fabricated images of candidates can shift opinion before anyone verifies them.
  • Fraud. Deepfakes power CEO scams, romance fraud, and identity theft. The Hong Kong case isn’t a one-off — deepfake fraud spiked sharply in 2025.
  • Reputation attacks. Non-consensual deepfake content is used to harass and extort people, disproportionately targeting women.
  • Eroding trust. When anything could be fake, people start doubting everything — including real evidence. This “liar’s dividend” might be the worst long-term effect of all.

Visual Signs of AI-Generated Images

AI models are good. But they’re not perfect. A trained eye can still catch artifacts that give them away.

Hands and Fingers

Hands have been AI’s Achilles’ heel for years. It’s gotten better, but problems persist.

  • Extra or missing fingers. Less common with newer models, but it still happens in group shots and complex poses.
  • Weird proportions. Fingers that are too long, too short, or inconsistently sized.
  • Fused joints. Where fingers meet or bend, AI often produces a smooth, melted look instead of actual knuckles.
  • Impossible gripping. Objects clipping through fingers, or hands that aren’t really holding what they appear to hold.

Midjourney v6 and Flux Pro handle hands way better than earlier versions. Don’t rely on this cue alone.

Teeth and Mouths

Teeth are harder for AI than you’d think.

  • Too-perfect or randomly distorted teeth. Real teeth have slight natural asymmetry. AI either makes them identical clones or mangles them.
  • Wrong tooth count. Sometimes there are just… too many teeth. Or not enough.
  • Floating teeth. Gumlines that fade into skin, or teeth without proper gum attachment.
  • Blurry mouth interiors. Open mouths where the tongue, palate, and throat are indistinct blobs.

Backgrounds

The subject might look great. The background tells a different story.

  • Melting architecture. Straight lines that bend, windows that don’t align, railings merging into walls. AI buildings often have geometry that’s physically impossible.
  • Broken perspective. Objects at different distances not following the same perspective rules. A table receding at a different angle than the floor it’s on.
  • Repeating patterns. Bookshelves with gibberish titles, crowds where faces repeat, textures that tile weirdly.
  • Phantom objects. Half a chair at the edge of the frame. A floating arm. An ear without a head attached.

Text in Images

This is still one of the fastest tells.

  • Nonsense words. Signs, T-shirts, and book covers with text that looks like letters but doesn’t spell anything real. Newer models have improved here, but it’s still common.
  • Inconsistent fonts. Letters in the same sign with slightly different styles or weights.
  • Reversed or scrambled characters. Letters that are backwards, upside down, or randomly mixing scripts.

If there’s text in the image, zoom in. It’s often the quickest way to spot a fake.

Skin, Lighting, and Accessories

  • Poreless skin. AI loves giving people unnaturally smooth, airbrushed skin. Like a beauty filter cranked to 11.
  • Lighting that doesn’t make sense. Shadows going in different directions, reflections that don’t match light sources, objects that seem to glow.
  • Mismatched accessories. Earrings that don’t match each other, glasses frames that differ on each side, jewelry that changes style across the image.
  • Hair weirdness. Strands that merge into clothing, individual hairs that turn into solid shapes, hairlines that dissolve into skin.

Metadata

Sometimes you don’t even need to look at the image itself.

  • Missing EXIF data. Real photos have camera model, aperture, GPS coordinates. AI images typically have none of this, or generic metadata.
  • Suspicious resolution. Many AI models output at 1024x1024 or 1024x1536. An image at exactly these dimensions with no EXIF data is worth questioning.
  • Odd compression. Real photos shared on social media have characteristic JPEG compression. AI images shared directly may compress differently.

How to Check if an Image Is AI-Generated

Your eyes can catch a lot, but dedicated tools catch more. Here’s what’s available.

1. Kaval — AI-Powered Image Verification

Kaval does AI-generated image detection as part of its content verification platform.

How to use it:

  1. Go to kaval.chat or open the Kaval WhatsApp bot
  2. Upload the image or send the image URL
  3. Get a verdict: whether the image appears AI-generated, with a confidence score and explanation

The WhatsApp integration is genuinely useful here. Someone sends you a suspicious image in a chat? Forward it to Kaval’s bot for an instant check — no switching to desktop tools. Kaval also handles fact-checking and URL safety, so you can verify the full context around a suspicious image in one place.

2. Hive Moderation

Hive Moderation analyzes images for synthetic characteristics and gives you a probability score. It can also identify which model likely generated the image (Midjourney, DALL-E, Stable Diffusion, etc.).

Accurate, well-regarded, and available as both a web demo and an API for developers.

3. AI or Not

AI or Not does one thing: tells you if an image is AI-generated or real. Upload, get a verdict. Simple.

Good for quick checks, but it can struggle with heavily post-processed images or AI images that have been edited further.

Sometimes the best approach is just checking if the image existed before the event it claims to show.

  • Google Images / Google Lens: Right-click and “Search image with Google” to find other instances online. If an image claiming to show something from yesterday has been circulating for years, it’s being used out of context.
  • TinEye: A dedicated reverse image search that’s especially good at finding the earliest known appearance of an image.

This won’t catch a freshly generated AI image, but it’s great for spotting recycled images used to fabricate stories.

Use Multiple Methods

No single tool is 100% accurate. The best approach:

  1. Visually inspect for the artifacts above
  2. Run it through a detection tool like Kaval or Hive
  3. Reverse image search to check if it existed before
  4. Check the context — who posted it, when, and what claim goes with it

If multiple checks raise flags, treat it as likely synthetic.

Deepfake Videos — How to Spot Them

Video adds complexity. But it also introduces more opportunities for artifacts to show up.

Visual Tells

  • Face boundary flickering. Where the swapped face meets the original head and neck, you’ll often see subtle flickering, color mismatches, or blurring. Most visible during quick movements.
  • Odd blinking. Early deepfakes barely blinked at all. Modern ones handle it better, but blinking can still look mechanical or too regular.
  • Lip-sync issues. If audio was generated separately, lip movements won’t perfectly match the words. Watch for hard consonants (B, M, P) where lips need to fully close.
  • Head turn artifacts. Extreme head angles cause deepfaked faces to warp, stretch, or lose detail. Profiles are harder for deepfake models than front-facing shots.
  • Lighting shifts. As the person moves, lighting on the face may change in ways that don’t match the environment. Shadows appearing and disappearing inconsistently.
  • Hair and ear glitches. Hair merging into background, ears changing shape between frames, hairlines shifting position.

Audio Tells

Voice cloning has its own giveaways:

  • Flat intonation. Cloned voices can lack natural pitch variation — slightly monotone, overly smooth.
  • Unnatural fluidity. Real speech has breaths, hesitations, “um”s and “uh”s. AI speech is often too clean.
  • Mismatched ambient sound. If the voice was generated separately and overlaid, the background noise might not match the environment in the video.

How to Verify Suspicious Video

  1. Slow it down. Watch at 0.25x or 0.5x speed. Artifacts become way more visible.
  2. Watch transitions. Pay attention when the subject moves, turns, or when camera angles change — that’s when deepfake artifacts show up most.
  3. Check the source. Verified account? Can you find the original? Videos shared without attribution deserve extra skepticism.
  4. Cross-reference the claim. If it shows a public figure making a statement, check whether Reuters, AP, or BBC reported on it. Major public statements always get wire agency coverage.

Why Your Eyes Aren’t Enough Anymore

Here’s the uncomfortable part: the visual tells in this guide are becoming less reliable every few months.

Generation quality keeps jumping. Compare Midjourney v3 (2022) to Flux Ultra (2025). The improvement is dramatic. Models released in 2026 produce images that experts struggle to identify as synthetic in controlled tests.

Post-processing hides artifacts. Someone can run an AI image through filters, add noise, adjust colors, and crop strategically to eliminate most tells. After an image gets screenshotted and re-shared on social media a few times, compression artifacts further mask AI signatures.

Volume is overwhelming. Millions of AI images are created daily. Manual inspection simply can’t keep up.

Generators train against detectors. New models are specifically tested against popular detection tools and refined to evade them. It’s an arms race, and the generators are well-funded.

This is why automated detection tools matter. They analyze pixel-level patterns, frequency domain signatures, and statistical anomalies that are invisible to us. They’re not perfect either — nothing is — but they catch things no human eye can.

The practical approach: use your eyes and judgment to assess context, use tools like Kaval for the technical analysis. Together, they’re much stronger than either alone.

FAQ

Can AI image detectors be fooled?

Yes. No tool is 100% accurate. Post-processing (adding noise, resizing, screenshotting) and adversarial techniques designed to fool specific detectors can evade them. But detection tools are continuously updated, and using multiple methods — visual inspection, automated tools, reverse image search, context analysis — makes it much harder for fakes to slip through. If something feels off about an image, investigate further even if a detector says it’s clean.

Are deepfakes illegal?

It’s complicated and depends where you are. Several US states have laws targeting deepfakes, especially non-consensual intimate ones and election interference. The EU’s AI Act requires AI-generated content to be labeled. India updated its IT Act in 2024 to address deepfake creation and distribution. Enforcement is the hard part — creation tools are widely available and content can be posted anonymously from anywhere. Generally: deepfakes for clearly labeled satire or entertainment are legal. Deepfakes for fraud, defamation, or non-consensual intimate content are increasingly criminal.

How can I protect my photos from being deepfaked?

You can’t fully prevent it if your photos are public, but you can reduce the risk. Limit high-resolution face photos on social media — deepfake models need clear face shots from multiple angles. Tighten your privacy settings so photos are visible only to connections. Tools like Fawkes add invisible perturbations to images that disrupt facial recognition and deepfake models (effectiveness varies). Watermarking services can add invisible signatures that help prove authenticity later. Most importantly, stay informed about detection tools so you can quickly identify and respond if your likeness gets used without consent.

AI-generated content isn’t going away, and it’s only getting better. Being able to quickly check whether an image or video is real is becoming a basic digital skill — right up there with recognizing phishing emails.

Don’t trust your eyes alone. Check suspicious images at kaval.chat, or forward them to the Kaval WhatsApp bot for a quick verdict. When seeing is no longer believing, verification is what you’ve got.