How to Check an AI Image or Video for Deepfakes

Updated 9 November 2026 · Standard formula, shown below

Checking an image or video for signs of AI generation or manipulation can help you decide whether to trust it, share it, or investigate further. No single visual clue or detection tool can prove that media is genuine or fake, so use several checks and look for independent confirmation. You will need the original file or its source link if available, a way to inspect details or individual frames, and access to reliable sources for comparison.

What You’ll Need

The media file or the earliest available link to it. Keep an unedited copy when possible.
A web browser and access to reverse image search or reputable fact-checking and news sources.
Optional: a video player that lets you pause, advance frame by frame, and view the audio track or file details.
Optional: an AI-media detector, used only as one clue rather than a verdict.

Step-by-Step

Save the claim and source. Note where you found the image or video, who posted it, when it appeared, and what it claims to show. Save the post link and, if permitted, the original file. Avoid relying on screenshots or copies sent through messaging apps: compression, cropping, filters, and re-encoding can remove useful details. Do not repost potentially harmful or intimate material while checking it.
Trace it back to the earliest available version. Search distinctive phrases from the caption or claim, and look for earlier posts, source footage, or reporting by established outlets. For an image, use a reverse image search or search several distinctive crops. For a video, take clear still frames from different moments and search those separately. A match may reveal that a clip is old, edited, or shown with a misleading caption; it does not by itself prove AI generation.
Check the context independently. Ask whether the location, date, event, and people named in the post are corroborated by sources that are not simply repeating the same upload. Look for original statements, full-length footage, local reporting, or other independently captured images. If a major event is supposedly documented only by a fresh, anonymous account, treat that absence of corroboration as a reason to wait—not proof of fabrication.
Inspect the whole image or video at normal and enlarged sizes. Look for inconsistencies in hands, teeth, jewelry, glasses, hair edges, text, repeated background details, reflections, shadows, and object shapes. In a video, check whether these details change abruptly between frames. Also look for mismatched lighting or perspective. These are prompts for closer checking, not reliable tests: real cameras, editing, low resolution, and motion blur can create similar oddities, while high-quality synthetic media may show none.
For video, compare movement, sound, and continuity. Pause at several points, including scene changes and moments when the person turns or speaks. Check whether the mouth movement roughly matches the words, whether the voice sounds consistent, and whether clothing, facial features, lighting, and the background remain coherent from frame to frame. Listen for unnatural cuts or audio that seems detached from the scene. Dubbing, poor synchronization, compression, and ordinary editing can also explain these signs, so seek the full, unaltered clip before drawing conclusions.
Review file details and provenance, if available. Check the file’s properties for creation or modification dates and whether it has been edited or re-encoded. Metadata can be missing, changed, or misleading, and a date is not proof of when a scene happened. Some files may include signed provenance information describing how they were created or edited; verify it with a compatible viewer and consider what it actually covers. Missing credentials do not establish that media is fake, and provenance is not a guarantee that the depicted event is truthful.
Use detection tools cautiously. If you try an AI-image or video detector, use a reputable service, read what formats and media types it supports, and check its privacy terms before uploading. Do not upload confidential, sensitive, or non-consensual content to an unknown service. Treat the result as an estimate, not a verdict: detectors can produce false positives and false negatives, and performance may vary with compression, edits, language, and generation methods. Compare results from more than one independent approach, but do not let a majority vote replace evidence.
Choose a cautious next step. If the evidence is mixed, describe the media as unverified rather than labeling it a deepfake. Wait for credible corroboration, ask the original poster for a source, or consult a qualified fact-checking organization. If you share a correction, link to evidence and explain what remains uncertain. Avoid accusing an identifiable person based solely on an automated score or visual impression.

Common Mistakes

Treating one visual glitch as proof. Artifacts can come from compression, filters, camera limitations, editing, or ordinary image defects.
Assuming a detector gives a definitive answer. Detection scores are fallible and may not explain how a result was reached.
Trusting metadata as a complete history. Metadata can be stripped or altered, and it usually cannot establish that the depicted event happened as claimed.
Checking only a cropped screenshot. Cropping removes context and may hide edits, source information, or details that change the assessment.
Confusing synthetic media with a false claim. A genuine photo can be presented with a false caption; AI-generated media can also be clearly labeled fiction or illustration. Assess both the media and the claim.

Tips

Keep a simple record of links, dates, searches, and observations so you can separate what you saw from what you inferred.
Compare the media with known, independently sourced images of the same person or place, while allowing for changes in lighting, age, clothing, and camera angle.
For video, examine several frames rather than a single still; a convincing frame does not establish that the whole clip is authentic.
Use neutral language such as “unverified” or “possibly manipulated” until evidence supports a stronger conclusion.
Be especially careful with media that could damage someone’s reputation, expose private information, or provoke harassment. Do not circulate it simply to ask whether it is real.

Troubleshooting

Reverse search finds no matches: Try a sharper frame or several crops, search distinctive caption wording, and look for the earliest account that shared it. A lack of matches may mean the media is new, private, or altered—not that it is authentic.
The video is too blurry to inspect: Find a higher-quality original rather than enlarging a compressed copy. Enlargement cannot restore details that are not present. Check the source and surrounding context in the meantime.
Detectors disagree or cannot process the file: Confirm that the tool accepts the file type and size, then stop treating detector output as a deciding test. Focus on provenance, context, and independent reporting.
Metadata is absent or confusing: This is common after social platforms process a file. Record that the details are unavailable and continue with source tracing; do not infer authenticity from missing data.
You cannot confirm or disprove the claim: Leave it unresolved. Avoid sharing it as fact, and revisit the assessment if reliable new evidence appears.

Frequently Asked Questions

Can I tell for certain by looking at an image?
Usually not. Visual clues can guide investigation, but convincing synthetic media may have no obvious flaws, and genuine images can contain artifacts.

Are AI deepfake detectors reliable?
They can be useful as screening tools, but none should be treated as conclusive. Results can be wrong, especially on edited, compressed, or unfamiliar media.

Does missing metadata mean a file is fake?
No. Many platforms remove metadata during upload or processing, and ordinary editing can also change it.

How can I check a video if I cannot upload it to a detector?
Trace its source, search still frames, inspect several points for continuity, and seek independent confirmation. These checks do not require uploading the clip.

What should I do if the video may be harmful or non-consensual?
Do not download or redistribute it unnecessarily. Use the platform’s reporting options and seek appropriate help if someone may be at immediate risk.

What wording should I use when I am unsure?
Say that the media is unverified or that you found signs worth checking. State what evidence you have and avoid presenting suspicion as proof.

Method: the standard formula for this calculation, rounded as each result states. See how the calculators work or report an error.