How to spot a deepfake
The 30-second version
No single tell reliably exposes a modern deepfake. The checks that work best are the boring ones: who posted it, where it first appeared, and whether anyone trustworthy can confirm it. Visual and audio glitches are a bonus when you find them, not proof when you don't.
- Pause when a clip makes you angry, scared or rushed. That reaction is what fakes are built to trigger.
- Trace it back to the earliest upload with a reverse image search.
- Check for Content Credentials, and remember that most real files have none.
- Look and listen for the tells below, then confirm through a second, independent source.
1. Check the context first
Most viral fakes are caught by context, not pixels. Fact-checkers such as AFP routinely debunk clips by finding that the account is new, the "news" appears nowhere else, or the footage is a rerun of an older AI video. Before you study the image, ask:
- Who posted it? A new or anonymous account, a parody account, or a page that only posts outrage is a warning sign.
- Is anyone credible reporting it? A real explosion, arrest or speech by a world leader will be covered by several established outlets within hours.
- Does the timing fit? Check dates, weather, seasons, signage and language against the claimed place and time.
- Is it asking you to act? Send money, click a link, share before it gets taken down. The NSA, FBI and CISA warn that synthetic media is increasingly used for impersonation and fraud, where urgency is the point.
Our weekly trending fakes show this in practice: most debunks start with "this account" or "this footage first appeared" rather than with a glitch.
2. Trace the source
The single most useful skill is finding the earliest copy of an image or video.
- Reverse image search: upload a screenshot to Google Lens, TinEye or Bing Visual Search, and look for the oldest result. If the "breaking" image existed last year, the claim is false, AI or not.
- For video, search keyframes. The free InVID-WeVerify verification plugin, built for journalists, splits a video into keyframes you can reverse-search and shows available metadata.
- Look for the uncropped original. Satire and AI-art accounts often label their work; reposts crop out the watermark or caption. Finding the original label settles many cases.
- Check the creator's other posts. An account that posts a stream of similar, too-perfect clips is often an AI content farm.
3. Check provenance
C2PA is an open technical standard, backed by companies including Adobe, Google, Microsoft and OpenAI, for attaching a signed, tamper-evident record of where a file came from and how it was edited. These records are branded Content Credentials.
- Drop a file into contentcredentials.org/verify or Adobe's Content Authenticity inspect tool to see any credential, who signed it and whether an AI tool was involved.
- For content made with Google's AI tools, Google's SynthID watermark can be checked in the Gemini app. A positive result means a Google tool made or edited it; a negative result says nothing about other generators.
Limits: provenance proves where a file came from when it is present. It cannot prove a file is fake when it is absent, because most cameras and platforms still don't attach credentials and many sites strip metadata on upload.
4. Look closely: visual tells
The MIT Media Lab Detect Fakes project and the FBI point to the same kinds of artifacts. Slow the video down (0.25× to 0.5×) and pause on frames where the face moves fast or something passes in front of it.
- Mouth and teeth: lips should close fully on "p", "b" and "m" sounds; teeth may blur into a single white band.
- Eyes and blinking: unnatural blink rhythm, mismatched reflections in the two eyes, or a gaze that doesn't track.
- Skin and edges: skin that is too smooth or too wrinkled for the face, flicker along the hairline, jaw or glasses, and earrings that change shape.
- Lighting and shadows: highlights, shadows and reflections should agree on one light source. Researchers at UCLA are building detectors around exactly this physics of light (see our story).
- Hands, text and backgrounds: extra or merging fingers, garbled signs or logos, and background objects that warp or loop.
Be careful: these artifacts are getting rarer with each model generation, and heavily compressed real videos can show them too. In the 2022 PNAS study by Sophie Nightingale and Hany Farid, people judged synthetic faces slightly more trustworthy than real ones. Use tells to raise suspicion, not to deliver a verdict.
5. Listen closely: audio tells
Voice clones are now convincing enough that the FTC warns a scammer may need only a short audio clip of someone's voice. Signs that sometimes give a clone away:
- Flat or oddly placed emotion, missing breaths, or an even pace with no natural pauses.
- Mispronounced names, places or family nicknames the real person would get right.
- Audio that is too clean for the claimed setting, or background noise that cuts in and out between words.
- A caller who avoids open questions or keeps steering back to a script, especially toward money.
The fix for a suspicious call is not better listening, it's verification: hang up and call back on a number you know. Our voice-clone scam guide covers this step by step.
6. Use detectors as one signal
Detection tools can help, but published testing shows clear limits. The Deepfake-Eval-2024 benchmark, built from deepfakes circulating on social media in 2024, found open-source detectors' performance (AUC) fell by about 50% for video, 48% for audio and 45% for images compared with older academic benchmarks. Commercial detectors did better, but none reached 90% accuracy on that real-world set.
If you use one, test more than one tool, use the highest-quality copy you can find, and treat a "likely fake" or "likely real" score as a lead to verify. Our detection tools guide compares what is available.
If you find one
- Don't share it, even to mock it. Quote posts spread fakes further.
- Report it to the platform using its manipulated media or impersonation option.
- If it involves money or a threat, report it to the FTC and the FBI IC3.
- If it depicts you, especially in intimate imagery, see our deepfake laws guide for removal rights under the federal TAKE IT DOWN Act and free tools such as StopNCII.org (adults) and NCMEC Take It Down (minors).
- Tell us through our report a deepfake page so our desk can investigate.
Frequently asked questions
Can you always tell a deepfake by looking at it?
No. In a 2022 PNAS study, people shown real faces and StyleGAN2-generated faces were right 48.2% of the time, about the same as a coin flip, and training only raised accuracy to 59%. Visual tells help, but context checks and source tracing are more reliable.
Are deepfake detectors reliable?
They are one signal, not a verdict. The Deepfake-Eval-2024 benchmark found that open-source detectors lost about half their measured performance on real-world 2024 fakes, and no commercial model it tested reached 90% accuracy. Treat a detector score as a reason to dig further.
What are Content Credentials?
Content Credentials are tamper-evident provenance records based on the C2PA standard. When a camera, editing app or AI tool attaches one, a checker such as contentcredentials.org/verify can show who signed the file and how it was edited. A missing credential does not prove anything, because most files never had one and metadata is easy to strip.
What should I do if I get an urgent call or video from someone I know asking for money?
Hang up and contact the person on a number you already have. The FBI recommends agreeing on a secret word or phrase with family members in advance, and never sending money, gift cards or crypto to someone you have only met or heard from online or on the phone.
Sources
Every factual claim on this page links to one of these. We checked each source on Oct 23, 2026.
- AI-synthesized faces are indistinguishable from real faces and more trustworthy · Nightingale & Farid, PNAS, 2022
- Deepfake-Eval-2024: A Multi-Modal In-the-Wild Benchmark of Deepfakes Circulated in 2024 · arXiv, 2025
- Criminals Use Generative Artificial Intelligence to Facilitate Financial Fraud · FBI IC3 PSA, Dec 3, 2024
- Contextualizing Deepfake Threats to Organizations · NSA, FBI & CISA, Sept 2023
- Detect Fakes · MIT Media Lab
- Scammers use AI to enhance their family emergency schemes · FTC consumer alert, Mar 2023
- Coalition for Content Provenance and Authenticity (C2PA) · c2pa.org
- Verify Content Credentials · Content Authenticity Initiative
- SynthID · Google DeepMind
- InVID-WeVerify verification plugin · InVID project