How to Spot AI-Generated Images, Videos and Voice Clones (2026 Deepfake Guide)
Looking for six fingers no longer works. Use this 6-step check for AI images, deepfake videos and cloned voices: provenance tools, scam defenses and what detectors miss.
On this page
- The 6-step check at a glance
- Why "look at the hands" stopped working
- Step 1: Find the source before anything else
- Step 2: Check Content Credentials (C2PA)
- Step 3: Look for an AI watermark (SynthID and similar)
- Step 4: Inspect the content (but treat tells as clues)
- Step 5: Use AI detectors, but only as a hint
- Step 6: Verify the claim, not just the file
- Voice clones and deepfake video calls: the scams to know
- Your defense plan: safe words, call-backs and delays
- Quick reference: what to do in common situations
- If you have been targeted
- How to protect yourself before anything happens
- If you create content: label it and keep your originals
- Frequently asked questions
- The bottom line
The most reliable way to spot an AI-generated image, video or voice is no longer to look for glitches. It is to check where the file came from. Find the original source, look for Content Credentials or a watermark, treat detectors as a hint, and confirm the claim through a second, trusted channel. If someone is asking for money or secrets, verify them by calling a number you already know.
A few years ago, you could catch fakes by counting fingers or spotting garbled text. In 2026 those tells are fading fast. Image and video generators now produce faces, hands and lighting that fool most people, and voice cloning has reached the point where researchers describe some clones as effectively indistinguishable from the real thing. If your plan is "I will notice if it looks off", you will eventually lose.
This guide gives you a practical system instead: six checks you can run in a few minutes, a simple way to weigh the evidence, and specific defenses against the scams that use AI to impersonate people you trust. It is written for ordinary readers, not experts. If you want the background on how these systems work, start with our explainer on how large language models work.
The 6-step check at a glance
| Step | What you do | What it tells you |
|---|---|---|
| 1. Find the source | Reverse image search, check who posted first | Whether the media is old, relabeled or unsourced |
| 2. Check provenance | Look for Content Credentials (C2PA) | How a file was made, if the data is intact |
| 3. Check watermarks | Use SynthID or other supported checks | Whether a participating AI tool made it |
| 4. Inspect the content | Look and listen for inconsistencies | Clues that justify a closer look |
| 5. Use detectors carefully | Run one or two tools, never one alone | A probability, not a verdict |
| 6. Verify the claim | Corroborate with trusted sources or call the person | Whether the story or request is real |
Notice that the visual inspection is step four, not step one. That order is deliberate.
Why "look at the hands" stopped working
Early AI images had obvious flaws: extra fingers, melted jewellery, nonsense signs. Generators have improved so quickly that careful viewers often cannot separate high-quality fakes from real photos by eye. Researchers who study this point to a second problem: teaching people to hunt for specific glitches can backfire. It makes you overconfident when you do not see one, and suspicious of real images that happen to contain odd lighting or compression.
Two facts shape everything below:
- A missing label does not prove something is real. Most content carries no provenance information at all.
- A detector score does not prove something is fake. Detection tools make mistakes in both directions.
So the goal is not certainty from one test. The goal is to stack several independent signals and decide how much to trust the file.
Step 1: Find the source before anything else
Most misleading media is not a brand-new fake. It is a real photo or clip from another time or place, or a fake that someone has reposted many times. Tracing the origin often settles the question in under a minute.
- Who posted it first? Look at the oldest version you can find, and the account behind it. Anonymous accounts with no history deserve extra caution.
- Run a reverse image search. Upload the image to Google Lens or TinEye to find earlier copies and the original context.
- For video, search keyframes. Take screenshots from several moments and search each one. A clip that is actually years old will often show up in older posts.
- Check for corroboration. If something dramatic really happened, established news outlets and independent photographers will usually have coverage from several angles. A single striking image with no other evidence is a red flag.
Step 2: Check Content Credentials (C2PA)
Content Credentials is a provenance standard, developed through the C2PA coalition, that attaches signed information to a file about how it was created and edited. Major technology and media companies back it, and some platforms now show a small "CR" icon on files that carry credentials.
To check a file, upload it to the free Content Credentials Verify tool or look for the icon on supporting platforms.
Know the limits before you trust the result:
- Most files have no credentials. Finding none proves nothing.
- Credentials can be stripped. Screenshots, re-uploads and some social platforms remove the data.
- Provenance is not truth. Credentials can tell you how a file was made, but they cannot tell you whether the scene or the claim is accurate.
When credentials are present and valid, they are one of the strongest signals available. When they are absent, simply move on to the next step.
Step 3: Look for an AI watermark (SynthID and similar)
Some AI tools embed an invisible watermark in what they generate. The best-known example is Google’s SynthID. Google lets users check for it in the Gemini app, which began supporting this in November 2025, and Search added a similar check in May 2026. OpenAI has also said that images from its tools can carry Content Credentials and watermark signals that its Verify tool can read.
The important limit is that a watermark check only detects content from the companies that use it. If the check finds nothing, the file could still come from another AI tool, including open-source models that embed no watermark at all. A positive result is strong evidence. A negative result is almost no evidence.
Step 4: Inspect the content (but treat tells as clues)
Visual and audio inspection is still useful as a second layer. These clues are most helpful when several appear together.
In images
- Text and signs that are garbled, inconsistent or slightly misspelled.
- Hands, teeth and ears that look slightly wrong or too uniform.
- Lighting and shadows that do not agree with each other, or reflections that do not match the scene.
- Backgrounds with strangely repeated crowd faces, warped lines or objects that merge into one another.
- Skin that looks too smooth or has no natural texture.
In video
- Lip-sync on "m", "b" and "p" sounds, where the lips should fully close.
- Blinking and head movement that feels stiff or oddly regular.
- The edges of the face, especially around the hairline, glasses and jaw, where face-swaps can flicker or blur.
- Lighting that changes on the face but not on the neck or background.
- Short, looping or heavily compressed clips, which hide imperfections.
In audio
- Flat or oddly consistent emotion, even in a supposedly stressful situation.
- No natural breathing, hesitations or background noise.
- Unusual pacing, or words that sound slightly stitched together.
- A caller who avoids questions only a real person could answer.
Remember that any of these can also appear in real media, and good fakes may show none of them. Use them to decide when to dig deeper, not to declare a verdict.
Step 5: Use AI detectors, but only as a hint
Detection tools can add a useful data point, but they are the least dependable part of the toolkit.
| Tool type | Examples | Main limit |
|---|---|---|
| Image classifiers | AI or Not, Hive Moderation | Often misjudge real or edited photos |
| Video scanners | Deepware Scanner | Weaker on new generators and compressed clips |
| Forensic viewers | FotoForensics | Needs interpretation; a visual aid, not an answer |
| Watermark checkers | SynthID check in Gemini and Search | Only covers content from participating tools |
A 2025 benchmark study, DeepFake-Eval-2024, reported that detector accuracy dropped sharply when tools were tested on real-world content instead of lab data. In practice, that means you should run one or two tools, compare the results with your other evidence, and never rely on a single score.
Step 6: Verify the claim, not just the file
Even a real photo can be used to mislead. The final question is whether the story attached to the media holds up.
- Does it trigger a strong emotion? Fear, anger and urgency are the scammer’s and propagandist’s favorite tools.
- Who benefits if you believe it? Ask who gains from you sharing, clicking or sending money.
- Is there independent confirmation? Look for reporting from several credible, independent sources.
- Can you ask the person directly? For anything personal, contact them through a method you already trust.
Voice clones and deepfake video calls: the scams to know
The most damaging uses of AI impersonation are not viral images. They are targeted calls and video chats that pressure someone into acting fast.
Family emergency calls. The US Federal Trade Commission has warned that a short clip of someone’s voice, taken from something they posted online, can be enough for a program to mimic them. A parent might get a panicked call in what sounds like their child’s voice. The FBI has also warned about AI-driven "virtual kidnapping" scams that combine cloned voices and altered photos to convince families a loved one has been taken, even though no one has.
Workplace impersonation. In a widely reported 2024 case, a finance employee at the engineering firm Arup’s Hong Kong office was tricked into transferring around US$25 million after a video call in which colleagues, including a senior executive, were deepfakes. Arup later confirmed it was the victim. The lesson is that a familiar face and voice on a call is no longer enough to approve a payment.
Celebrity investment videos. Deepfaked videos of famous people endorsing guaranteed returns are a common bait. Real public figures do not promise risk-free profits through an unknown app.
The scale of the problem. The FBI’s Internet Crime Complaint Center reported more than $16 billion in cyber-enabled crime losses in the US in 2024, and experts expect AI-assisted fraud to be one of the biggest threats going forward.
Your defense plan: safe words, call-backs and delays
Because voices and faces can now be faked, build verification into the relationship, not into your ears.
Create a family safe word
Agree on a word or short phrase that only your close circle knows. Rules that make it work:
- Pick something unguessable, not a pet’s name or a birthday that appears online.
- Never post or text it. Say it out loud once in person.
- Use it for any urgent request involving money, danger or secrets, even if the voice sounds exactly right.
- Make sure elderly relatives and teenagers know it, since both groups are frequent targets.
Experts quoted in recent reporting, including a Fortune feature on a fake-kidnapping scam, recommend this approach because it is cheap and works even against a perfect voice clone.
Use the call-back rule
If someone calls or video-chats asking for money, hang up and call them back using the number saved in your phone, not the one that called you. If it is a "bank" or "official", use the number printed on your card or on the official website.
Slow everything down
Every one of these scams depends on speed. Take a breath, ask a question only the real person could answer, and tell the caller you will call them right back. Real emergencies survive a two-minute delay.
Protect money at work
- Require two people to approve large or unusual payments.
- Confirm payment changes through a second channel, such as a known phone number.
- Treat "keep this confidential, it is urgent" as a warning sign, not an instruction.
Quick reference: what to do in common situations
| Situation | Red flags | What to do |
|---|---|---|
| Panicked call from a "relative" | Urgency, secrecy, unknown number, payment by gift card or crypto | Hang up, call the person on their saved number, ask for the safe word |
| Video call from your "boss" asking for a transfer | Unusual request, pressure, secrecy | Verify by phone, follow your payment approval process |
| Viral shocking image or video | No source, strong emotion, only on social media | Reverse search, check credentials, look for reputable coverage |
| Celebrity endorsing an investment | Guaranteed returns, a link to an unknown app | Ignore, report the post, do not click |
| New online friend who will not meet or video chat normally | Glitchy video, excuses, requests for money | Pause contact, search their photos, never send money |
If you have been targeted
- Stop engaging. Do not send money or personal information.
- Contact your bank immediately if you paid. Quick action can sometimes stop or recall a transfer.
- Save evidence: phone numbers, usernames, screenshots, payment details and message history.
- Report it to your local police or national cybercrime unit and to the platform. In the US, you can report to the FTC and the FBI’s IC3.
- Secure your accounts. Change passwords, turn on two-factor authentication and review recent logins. Our password manager guide shows how to do this quickly.
If an AI-generated intimate image of you is being shared, you are not at fault. Report it to the platform, and use free tools such as StopNCII.org for adults, or Take It Down if the image shows someone under 18, which help prevent the files from spreading.
How to protect yourself before anything happens
- Set up your safe word with family and close colleagues this week.
- Limit what you share publicly. Detailed travel plans, real-time location and long public voice or video clips give scammers material.
- Use strong, unique passwords and two-factor authentication, so a fake message cannot easily lead to account takeover.
- Learn the phishing basics. Many deepfake scams begin with an ordinary phishing message. See our guide to spotting phishing emails and scams.
- Talk to older relatives. A five-minute conversation can prevent a very expensive call.
If you create content: label it and keep your originals
If you publish AI-generated or AI-edited media, say so clearly. When your tools offer to attach Content Credentials, turn it on. Keep original files and drafts so you can prove what is real if your work is ever challenged or impersonated.
Frequently asked questions
Can I tell an AI image by looking at it?
Sometimes, but not reliably. Modern generators produce images that fool most viewers. Use source tracing and provenance checks first, and treat visual clues as a second layer.
Are AI image detectors accurate?
They can be a helpful hint, but they make mistakes, especially on edited, compressed or real-world images. Never rely on one score.
Does a missing watermark mean the image is real?
No. Watermark checks only detect content from tools that embed them. Many AI tools, including open-source ones, add nothing.
What is SynthID?
SynthID is Google’s invisible watermark for content made with its AI tools. A check in the Gemini app or Search can detect it in supported content.
What are Content Credentials?
They are signed provenance data that can record how a file was made and edited. They are helpful when present, but they can be removed by screenshots or re-uploads.
How can I tell if a voice on the phone is cloned?
You often cannot, which is why the best defense is verification: a family safe word, hanging up and calling back on a known number, and never sending money under pressure.
Are all AI images harmful?
No. Most AI-generated media is harmless creative or commercial work. The risk comes from using it to deceive, impersonate or defraud. Clear labeling is the healthy norm.
The bottom line
You cannot rely on your eyes and ears to catch every fake, and that is fine. Trace the source, look for Content Credentials or a watermark, use detectors as a hint, and confirm any request through a channel you already trust. Add a family safe word and a call-back habit, and even a perfect clone of a loved one’s voice cannot make you act. Share this guide with the relatives who are most likely to get that call.
Sources and further reading: public guidance on Content Credentials and SynthID, FBI and FTC warnings about voice-cloning and virtual-kidnapping scams, FBI Internet Crime Complaint Center data, and reporting on AI-enabled fraud. Tools and features change quickly, so we update this page as detection methods and platform policies evolve.
Get new articles by email
One short email when we publish. No spam, unsubscribe any time.