CAN YOUTRUST THIS?

A deepfake can be convincing even when you know what deepfakes look like. A person appears to speak words they never said, a photograph shows a disaster that never happened, or a video combines genuine footage with a fabricated audio track. The important question is not simply whether an image contains strange hands or unusual shadows. It is whether the source, history and claim survive careful checking.

Independent fact-checking organisation Full Fact explains that automatic AI detectors and visual clues can be unreliable, while the Content Authenticity Initiative describes how Content Credentials can reveal aspects of an image's origin and editing history. Neither gives a universal “real or fake” verdict.

This field report goes beyond our general crisis-misinformation checklist. It explains how to investigate a particular photograph or video step by step: preserve the original context, search for earlier versions, compare visual clues with outside evidence, interpret metadata and decide what level of confidence is justified. The objective is practical verification, not pretending you can become a forensic laboratory by watching one online tutorial.

What makes a video or image a deepfake?

Deepfake usually refers to synthetic or manipulated media produced with machine-learning techniques to depict a person, scene or event in a misleading way. It might involve replacing a face, modifying mouth movements, cloning speech, changing an existing scene or generating an entirely fictional recording.

But not every deceptive image is a deepfake. A genuine video uploaded with the wrong date can be misinformation. A cropped photograph can hide context. A voice recording can be misleading because of selective editing without involving AI at all. Your verification should therefore test the claim attached to the content, not only the technique that might have produced it.

One question can be answered before another: “Did this video exist three years ago?” can often be checked without knowing whether any individual frame was generated by AI.

The evidence ladder: what does each clue actually prove?

Treat evidence as a ladder rather than a binary label. Some findings establish very little; others independently contradict the claimed event.

Stronger conclusions come from several independent clues agreeing. A convincing single frame, a confident detector score or thousands of reposts are weaker evidence than an original, dated recording corroborated by trustworthy independent sources.

Step 1: preserve the claim, not just the picture

Before researching, write down the exact claim. Does a video supposedly show a minister announcing an emergency? A soldier using an AI weapon? A dam collapsing this morning? The same image can be genuine yet entirely unrelated to the accompanying assertion.

Record the account that shared it, publication time, claimed location and any original attribution. Keep a screenshot of the caption alongside the media if that is lawful and appropriate. Avoid repeatedly reuploading sensitive footage containing identifiable people or distressing events.

The task is to establish what the publisher says this proves. Without that, you can spend half an hour analysing an image and still miss the actual misinformation.

Step 2: find the earliest discoverable version

Use a reputable reverse-image search tool on a still image or distinctive frame. Multiple search tools can produce different results. Search for unusual objects, visible signs or exact phrases from the post. For video, extract or capture a frame that includes distinctive landmarks rather than only a person's face.

The Bellingcat Online Investigation Toolkit collects tools for image and video analysis, web archives and locating the origin of material. It is a reference library, not an automatic authenticity detector.

If the same footage appears in a credible report from an earlier year, that undermines any claim that it shows a brand-new incident. But an early upload is not necessarily the original upload, and a search returning no results does not prove the clip was created today.

Example: a viral video claims to show a new power station explosion. An earlier upload clearly labels the same scene as an industrial fire from a different country years before. You have strong grounds to reject the new location-and-date claim, even without proving whether the original video was edited.

Step 3: verify the location

Look for persistent visual features: street layouts, bridges, building shapes, road signs, mountains, industrial structures or distinctive skylines. Compare them against reliable mapping, photographs or news material. One sign can be misread, mirrored or transplanted; several features fitting the same place are stronger evidence.

Be careful when comparing images with wide-angle lenses, mirrored clips or unusual camera positions. A camera can distort distances, and structures may have changed since public street-level imagery was captured. If you cannot identify the location, say so rather than inventing one.

This matters particularly with conflict imagery and emergency footage. A real scene from one area can trigger panic elsewhere when falsely presented as local. Our autonomous-weapons report explains why sensational claims about military AI require the same attention to source and location.

Step 4: check when it could have been recorded

Compare the claimed date with independently documented events and visible conditions. Daylight, weather, seasonal vegetation, construction changes and the earliest identified upload may all offer clues. Sometimes an official statement or a legitimate broadcaster has already placed the scene in a different sequence of events.

Do not treat shadows as a precision clock without specialist knowledge. The sun's apparent position depends on location, date, camera orientation and editing. A blurred night-time video may support no useful timing inference at all.

What matters is finding a credible contradiction or corroboration, not extracting an exact timestamp from every frame.

Step 5: inspect the original file and metadata cautiously

If the original file is legally available from a trusted source, its metadata may contain camera make, software information or timestamps. However, social platforms often strip or rewrite metadata. Screenshots can erase original data completely, and metadata can also be modified.

Therefore:

• Present metadata: one clue about the file's history, not automatic proof of truth.
• Missing metadata: not evidence by itself that a file was generated by AI.
• Editing-software fields: may indicate ordinary cropping or colour correction, not necessarily fabricated content.
• Conflicting timestamps: need investigation; device clocks and time zones can be inaccurate.

Be cautious about uploading sensitive private images to unknown “AI detective” websites. A verification attempt can create its own privacy problem.

What are Content Credentials and C2PA?

The Coalition for Content Provenance and Authenticity (C2PA) develops a standard for attaching cryptographically verifiable claims about the origin and editing history of media. Content Credentials are a way that such information can be carried and displayed. For example, supported tools may state that an image was generated using an AI model or edited in a named application.

The Content Authenticity Initiative's explanation is careful: credentials describe available provenance, but do not determine whether a claim in an image is true. The C2PA explainer also distinguishes evidence of content history from factual accuracy.

A genuine photograph of an empty street can still be captioned “the entire city has evacuated”, which the photograph cannot establish. Conversely, an image without Content Credentials may simply have been produced using a device or platform that does not support them, or the information may have been lost through ordinary sharing.

Use verified provenance as one piece of evidence and always test the story attached to it.

Step 6: examine what AI detectors can—and cannot—tell you

Automated detectors look for technical patterns associated with synthetic media. Some can provide useful evidence under specific conditions, but their results depend on the source, compression, manipulation and models involved. New generation methods can make previous telltale signs less dependable.

Full Fact's published methodology says it does not routinely rely on AI detectors as definitive tools, citing problems with their reliability. Its investigators instead use contextual research, reverse-image searches, visual comparison and expert input.

If one service labels a photo “99% AI”, do not announce that result as proof of fabrication. Ask what the detector was evaluated on and whether independent evidence supports the conclusion. Nor should a low synthetic-media score make you trust a clip whose location or date is demonstrably false.

Step 7: seek evidence outside the file

This is the stage that often resolves the question. Does the alleged speaker's organisation publish the same announcement? Are there independent eyewitness recordings from different positions? Do credible journalists verify the event? Does an emergency service or relevant utility confirm the operational facts?

Beware false corroboration: hundreds of accounts can all reproduce one original post. Five screenshots of the same message are still one source. Genuine independent corroboration provides information from different routes, not just repetition.

For an immediate threat, the safest choice may be to follow confirmed safety instructions while leaving the video unverified. You can take precautions without claiming a doubtful clip is authentic.

A full example: the fake emergency broadcast

This is a fictional teaching example.

An online video appears to show a national news presenter announcing a sudden banking shutdown. The footage contains a convincing studio and voice. Comments claim it was broadcast ten minutes earlier and urge people to withdraw cash.

The verification process begins with the claim, not facial analysis. The publisher is a previously unknown account with no link to the broadcaster. A search finds the same presenter footage in a two-year-old segment about a different topic. The broadcaster's official website and programme stream contain no such announcement. Reposts all trace to the original anonymous upload.

Even if you cannot prove which editing technique was used, the available evidence contradicts the purported broadcast. The rational conclusion is that the claim is unsupported and likely misleading. There is no basis for treating it as an emergency instruction.

That example demonstrates why proving a particular AI model made a clip is often less urgent than disproving the dangerous claim attached to it.

When the video concerns a real person

Deepfake abuse can target private people as well as celebrities and politicians. If a manipulated clip impersonates you, threatens you or exposes intimate content without consent, preserving necessary evidence and reporting it to the platform or relevant authorities may matter more than public debate over technical artefacts.

Do not amplify harmful imagery to ask the internet whether it is fake. Share only what is necessary with appropriate support services, investigators or trusted contacts. In a financial impersonation case, notify your bank or the relevant provider promptly.

Our family impersonation defence plan helps households prepare for this kind of misuse before a manipulated call or video reaches them.

How to record a defensible conclusion

Words such as confirmed, contradicted and unverified are useful only when they correspond to what your evidence actually shows.

A twenty-minute deepfake verification drill

1. Choose a well-documented historical example from a reputable fact-checker; do not circulate an unverified live emergency clip.
2. Write down the exact claim, date and alleged location.
3. Find the earliest credible version or independent source using search and reverse-image tools.
4. Identify at least two visual details you can compare with outside records.
5. Check whether reliable independent reporting supports the event.
6. Write a two-sentence conclusion: what you established and what remains uncertain.

Do not make the exercise dependent on paying for an AI detector. The valuable skill is developing and testing a chain of evidence, not collecting a score from a proprietary tool.

The bottom line

Detecting synthetic media is becoming harder. But many false claims still collapse under straightforward checks of source, time, place, provenance and independent corroboration. You need not identify the precise model or editing tool behind a clip to protect yourself from its misleading instructions.

Verify the claim first. Treat technical clues as supporting evidence. And when the stakes involve safety or money, act on trusted channels rather than a viral file.

Frequently asked questions

How can I spot a deepfake video?

Start with the original source, date, location and independent confirmation. Visual glitches can be clues but should not be treated as a reliable detection method.

Can an AI detector prove an image is fake?

No detector is universally reliable. A high score may justify further investigation but is not an adequate substitute for verifying context and provenance.

Do Content Credentials prove a photograph is real?

They can supply signed information about origin and edits where supported, but do not establish that the depicted scene or its caption is factually accurate.

What if the image has no metadata?

Missing metadata is common after sharing or editing and is not proof of AI generation.

Should I share a possible deepfake to ask others to check?

Not automatically. Avoid amplifying harmful or private imagery. Where appropriate, seek help from reliable fact-checkers, the platform or relevant authorities and share only the minimum necessary.

Sources and further reading

• Full Fact — A Fact-Checker’s Guide to Spotting AI Content — practical limits of detection and verification.
• Bellingcat — Online Investigation Toolkit — tools for source tracing, images, videos and geolocation.
• Content Authenticity Initiative — How Content Credentials Work — provenance, origin and editing history.
• C2PA — Content Credentials Explainer — why provenance and truth are distinct.
• Content Authenticity Open Source — FAQ — absent credentials and supported claims.

Research reviewed 6 November 2026. Examples are fictional demonstrations, not claims about specific current emergencies.