A convincing video was once considered powerful evidence that an event had actually happened. Seeing appeared to be believing.
That assumption is becoming increasingly dangerous.
Modern artificial intelligence systems can generate realistic footage of politicians, celebrities, wars, protests, natural disasters, financial announcements, and supposed breaking-news events. A fabricated clip may contain believable lighting, natural facial movements, synchronized speech, realistic camera shake, and environmental details that would have been difficult to reproduce only a few years ago.
At the same time, misleading video does not always involve a completely artificial scene. A genuine recording can be edited, relabeled, shortened, translated inaccurately, paired with fabricated audio, or presented as evidence of an event that happened in another country or another year.
This means the central verification question is no longer simply:
“Does this video look fake?”
A better set of questions is:
Where did the video come from?
Who first published it?
Was it generated or edited using artificial intelligence?
Does it contain trustworthy information about its creation history?
Do independent sources confirm the event shown in the footage?
Could a real video be circulating with a false description?
Hidden watermarks and Content Credentials can help answer some of these questions. However, they are not universal truth detectors. Professional verification teams, including BBC Verify, use them as part of a much broader investigation involving source tracing, geolocation, metadata analysis, contextual research, reverse searches, and comparison with independent evidence.
Why AI-generated videos are becoming harder to recognize
Early deepfakes frequently contained visible mistakes. Faces flickered between frames. Mouth movements did not match the words being spoken. Skin tones changed abruptly. Hairlines appeared blurred. Reflections behaved incorrectly, and hands or teeth were often malformed.
These problems made many synthetic videos detectable through close visual inspection.
AI video systems have since improved substantially. Newer models can generate coherent movement, realistic shadows, plausible camera angles, changing facial expressions, and detailed backgrounds. Short clips are particularly difficult to judge because viewers have less time to notice inconsistencies.
Social media also creates ideal conditions for deception. Videos are often:
-
compressed into lower quality;
-
cropped into vertical formats;
-
covered with captions or logos;
-
viewed on small mobile screens;
-
shared without their original source;
-
reposted by accounts unrelated to the creator;
-
presented during emotionally charged breaking-news events.
Compression can conceal technical imperfections that might have been visible in the original file. A clip lasting only five or ten seconds may therefore appear completely credible to someone scrolling quickly through a news feed.
Visual clues remain useful, but they should now be treated as preliminary warning signs rather than definitive proof.
What makes an AI video misleading?
The phrase “AI video” covers several different types of content, and not all of them are deceptive.
A filmmaker may openly use generative AI to create an artistic sequence. A company may use a synthetic presenter in an advertisement. A teacher may produce an animated historical reconstruction clearly labeled as a simulation.
The danger arises when viewers are led to believe that artificial, altered, or misrepresented footage documents a real event.
Misleading videos generally fall into several categories.
Fully generated scenes
The people, location, dialogue, movement, and event are created by an AI model. No original recording exists.
Examples could include a fabricated explosion, a nonexistent protest, or a synthetic press conference involving a public figure.
Face or identity manipulation
A real person’s face may be placed onto another person’s body, or their facial movements may be altered to make them appear to say or do something that never occurred.
Synthetic or cloned audio
The visual footage may be genuine while the voice has been generated or replaced. A real interview can therefore be transformed into a false statement without changing most of the images.
AI-assisted editing
Generative tools can add or remove objects, replace backgrounds, extend scenes, alter crowds, modify signs, or create transitions between genuine and synthetic footage.
Genuine footage with false context
The video itself may be authentic, but the description is false. Old hurricane footage may be presented as a storm happening today. A military exercise may be labeled as an active attack. A crowd celebrating a sporting victory may be described as a political demonstration.
This final category is especially important because no AI detector will solve it. The pixels may be genuine even though the claim attached to them is completely misleading.
Visible watermarks are not strong evidence
Most people are familiar with visible watermarks: a company logo, username, broadcaster symbol, or translucent text placed over a video.
These markings can help identify who published or owns a clip, but they provide weak evidence of authenticity.
A visible logo can easily be:
-
cropped from the frame;
-
blurred or painted over;
-
covered with another graphic;
-
copied onto unrelated footage;
-
added by an account that did not create the video.
A BBC, CNN, Reuters, government, or police logo inside a video does not prove that the organization produced it. Impersonators frequently copy the branding of trusted institutions.
The account publishing the video, the page hosting it, and the original source must still be checked independently.
What is an invisible AI watermark?
An invisible watermark is a machine-detectable signal embedded into digital content. Unlike a logo placed in the corner, the signal is not intended to be noticed by a human viewer.
Depending on the system, the watermark may be incorporated into:
-
image pixels;
-
individual video frames;
-
audio frequencies;
-
generated text patterns;
-
file metadata;
-
a combination of the content and an external verification record.
The goal is to make AI-generated material identifiable even after ordinary transformations such as compression, resizing, filtering, or limited cropping.
Google DeepMind’s SynthID is one of the best-known examples. Google says SynthID embeds imperceptible watermarks into AI-generated images, video, audio, and text. For video, the signal is placed into generated frames and is designed to remain detectable after common modifications, including cropping, filters, frame-rate changes, and lossy compression.
Google also uses SynthID with content created through its Veo video technology.
What a detected watermark can prove
When a recognized watermark is reliably detected, it can provide meaningful evidence that content was created or modified by a particular family of AI tools.
For example, the detection of a valid SynthID signal may indicate that all or part of a video was generated or altered using supported Google AI technology.
That is useful evidence, but it does not automatically reveal:
-
who published the clip;
-
why it was created;
-
whether the accompanying caption is true;
-
whether genuine footage was mixed with generated footage;
-
whether the video is satire, art, advertising, or deliberate disinformation;
-
whether later editing changed its meaning.
A watermark identifies something about the technical origin of the media. It does not independently determine the truthfulness of the story being told.
Why the absence of a watermark proves very little
One of the most common verification mistakes is assuming that a video must be genuine because no AI watermark was detected.
That conclusion is unsafe.
A video may lack a detectable watermark because:
-
the generating tool never embedded one;
-
the model was open source or independently modified;
-
the watermarking feature was disabled;
-
extensive processing weakened the signal;
-
only a small section of the video was generated;
-
the detector does not support the tool used;
-
the file was recorded from another screen;
-
the video was reconstructed through multiple exports;
-
the footage is real but still falsely captioned.
Google itself describes SynthID as an important component rather than a complete solution for identifying AI-generated content.
A positive watermark result can therefore be informative. A negative result should normally be treated as inconclusive.
Watermarks and metadata are not the same thing
Watermarks and metadata are often discussed together, but they work differently.
A watermark is embedded into the content or associated with it through a detection mechanism. Metadata is information stored alongside the file.
Traditional metadata may include:
-
creation time;
-
device model;
-
camera settings;
-
editing software;
-
geographic coordinates;
-
file format;
-
export history.
Metadata can be extremely helpful, but it is fragile. Social networks, messaging apps, editing programs, and publishing systems may remove it automatically. It can also be edited or fabricated.
This is one reason the technology industry has developed cryptographically signed provenance systems such as C2PA Content Credentials.
What are Content Credentials?
Content Credentials are designed to provide a verifiable history of a piece of digital media.
They are based on technical standards developed by the Coalition for Content Provenance and Authenticity, commonly known as C2PA. The coalition develops open standards intended to help people examine the source and history—or provenance—of images, videos, audio files, and documents.
A Content Credential can contain signed statements describing:
-
how an asset was created;
-
which device or application created it;
-
whether generative AI was used;
-
which editing steps were recorded;
-
which pieces of media were combined;
-
who or what signed the credential;
-
whether the associated provenance information has been altered.
C2PA describes a Content Credential as a cryptographically bound structure containing assertions about an asset’s origin, modification history, and use of AI.
The system is sometimes compared to a nutrition label for digital media. The comparison is useful because the credential provides information that helps a person evaluate the content. It does not make the decision for them.
How cryptographic signing helps
A Content Credential can be digitally signed by a device, application, publisher, or organization.
Cryptographic signing helps establish that the credential is associated with the media and that its recorded information has not been changed without detection.
For example, a compatible camera could sign footage at the moment of capture. Editing software could then add another signed entry recording an adjustment, crop, color correction, or AI-assisted modification. A publisher might later add information about its editorial process.
The viewer could potentially inspect that chain of activity.
This provides stronger provenance evidence than ordinary metadata because silent changes to the credential should be detectable.
However, the presence of a valid signature does not mean that every statement or scene in the video is objectively true. C2PA explicitly explains that Content Credentials do not make value judgments about the truth of provenance claims. They help verify that provenance information is associated with an asset, properly formed, and tamper-evident.
What Content Credentials can tell you
Depending on how the media was produced, a credential may help answer questions such as:
-
Was this file captured by a credential-enabled camera?
-
Was generative AI declared during its creation?
-
Was the background replaced?
-
Was the file edited in compatible software?
-
Which organization signed the media?
-
Has the credential remained valid?
-
Were multiple source files combined?
-
Is there a recorded chain from capture to publication?
These answers can provide valuable context, particularly when the source is a recognized newsroom, photographer, camera, or software provider.
What Content Credentials cannot guarantee
Content Credentials are not a universal certificate of truth.
They cannot necessarily prove that:
-
a person’s statement is factually correct;
-
the scene was staged before filming;
-
the person recording the footage understood the event correctly;
-
important activity outside the frame was not omitted;
-
the caption accurately explains what happened;
-
an apparently real object was not a screen, projection, or physical recreation;
-
a publisher acted honestly;
-
the absence of credentials means the media is fake.
A genuine video can document a staged event. A properly signed photograph can still be selected in a misleading way. A truthful video can also lose its credentials when uploaded to a platform that removes metadata.
Provenance supports verification, but it does not eliminate the need for journalism and human judgment.
Why durable credentials matter
Ordinary metadata may disappear when a file is uploaded, compressed, copied, or screenshot. This creates a practical problem: a trustworthy provenance record is not useful if it vanishes during normal distribution.
Developers are therefore working on more durable approaches that combine:
-
secure metadata;
-
invisible watermarks;
-
media fingerprinting;
-
externally stored credential records.
In such a system, even when embedded metadata is removed, a watermark or fingerprint may help a verification tool find the associated credential elsewhere.
The Content Authenticity Initiative describes durable Content Credentials as a combination of secure metadata, watermarking, and fingerprinting intended to preserve access to provenance information through media transformations.
This remains an evolving area rather than a universally deployed solution.
How BBC Verify investigates suspicious footage
BBC Verify is a specialist BBC News operation focused on examining claims, imagery, data, eyewitness material, and videos circulating online.
Its work illustrates an important principle: verification rarely depends on one technical test.
A newsroom investigating a suspicious video may carry out several processes at the same time.
Finding the earliest available version
Journalists attempt to identify where the clip first appeared.
A repost with millions of views may be several generations removed from the original upload. Earlier versions may contain:
-
a longer sequence;
-
better image quality;
-
the original audio;
-
an uncropped landmark;
-
a username or caption;
-
useful comments from witnesses;
-
an upload time closer to the event.
Finding the earliest source can also reveal whether the footage existed before the event it supposedly shows.
Examining the publishing account
The account itself is investigated.
Relevant questions include:
-
When was the account created?
-
Does it regularly publish reliable original footage?
-
Has it changed its name?
-
Does it impersonate a news outlet?
-
Does it upload unrelated viral clips?
-
Is its location consistent with the claimed event?
-
Has it previously distributed fabricated material?
An established account can still publish false information, but its history provides useful context.
Inspecting watermarks and Content Credentials
Verification teams can check whether the file contains:
-
a known AI watermark;
-
C2PA Content Credentials;
-
conventional metadata;
-
software or device information;
-
signs of an interrupted editing history.
A valid credential may reveal that part of the video was generated, that an image was edited, or that a compatible camera signed the original footage.
The Content Credentials verification service allows users to upload supported files and inspect available credentials and changes. It also warns that credentials are still being adopted, meaning many files will contain no inspectable information.
Geolocating the footage
Geolocation attempts to determine where a video was recorded.
Investigators compare visible details with maps, satellite imagery, street-level photographs, architectural records, and previously published media.
Useful clues include:
-
road layouts;
-
building shapes;
-
mountains;
-
coastlines;
-
shop signs;
-
traffic direction;
-
utility poles;
-
road markings;
-
public monuments;
-
vegetation;
-
license-plate formats;
-
language on signs.
A single storefront, tower, intersection, or skyline feature can sometimes identify the location.
Determining when it was recorded
Chronolocation tries to establish when the footage was captured.
Investigators may examine:
-
weather records;
-
sunrise and sunset times;
-
shadow direction;
-
construction progress;
-
seasonal vegetation;
-
clothing;
-
known event schedules;
-
traffic conditions;
-
visible advertisements;
-
previous versions of the same location.
A clip supposedly recorded in winter may show trees, shadows, or weather conditions inconsistent with the claim.
Comparing independent footage
A major public event is rarely captured by only one person.
Journalists look for:
-
other camera angles;
-
livestreams;
-
security footage;
-
photographs;
-
emergency-service statements;
-
satellite images;
-
local media reports;
-
eyewitness accounts;
-
official records.
If a video shows a major explosion in a populated city but there are no emergency reports, witness videos, local alerts, traffic disruptions, or physical damage, the absence of corroboration becomes significant.
Analyzing frames and audio
Frame-by-frame examination remains valuable, even though visual artifacts are no longer decisive.
Investigators may look for:
-
disappearing objects;
-
inconsistent reflections;
-
warped text;
-
changing jewelry or clothing;
-
unnatural body movement;
-
incorrect lip synchronization;
-
inconsistent depth;
-
repeated crowd patterns;
-
physically impossible shadows;
-
sudden changes in background detail.
Audio can provide separate clues:
-
mismatched room acoustics;
-
abrupt background-noise changes;
-
repeated breathing patterns;
-
unnatural pronunciation;
-
missing environmental sounds;
-
speech that does not match visible mouth movement.
These signs can justify further investigation, but they should not be used alone to declare a video fake.
Contacting the source
Whenever possible, journalists contact the person who filmed or uploaded the material.
They may request:
-
the original file;
-
additional footage;
-
photographs from the location;
-
an explanation of how the clip was recorded;
-
permission and ownership information;
-
confirmation of time and place;
-
details of other witnesses.
The original file may contain information lost during social-media processing.
BBC involvement in content provenance
The BBC has participated in broader industry work concerning trustworthy digital media and content provenance.
The C2PA ecosystem brings together efforts from the Content Authenticity Initiative and Project Origin, a Microsoft- and BBC-led initiative concerned with disinformation in digital news.
This work reflects a larger goal: audiences should be able to see not only the finished video but also meaningful information about where it came from and how it changed before publication.
Professional camera manufacturers are also moving provenance closer to the moment of capture. Sony has developed camera-authenticity technology that can place digital signatures into media and support the C2PA format.
Sony’s PXW-Z300 professional video camera has been presented as part of an end-to-end workflow in which Content Credentials can move from camera capture into cloud-based production systems.
The significance is not that one camera can solve misinformation. It is that trustworthy provenance can begin before footage reaches an editor or social platform.
How ordinary viewers can check a suspicious AI video
You do not need access to a newsroom laboratory to investigate a questionable clip.
A practical verification process can be completed using publicly available information and tools.
Step 1: Stop before sharing
Urgency is often part of the manipulation.
Misleading clips frequently include phrases such as:
“Share before it is deleted.”
“The media will not show you this.”
“This just happened.”
“Breaking: confirmed.”
“They are hiding the truth.”
Pause before forwarding the video. A few minutes of checking can prevent a false claim from reaching hundreds of people.
Step 2: Separate the video from the caption
Write down exactly what is being claimed.
For example:
“This video shows flooding in London today.”
That claim contains several separate elements:
-
the video shows flooding;
-
the location is London;
-
the event happened today;
-
the footage documents the claimed incident.
Each element needs evidence. The images may show real flooding but in another city, or they may be old footage from a previous year.
Step 3: Find the earliest upload
Search for distinctive phrases from the caption. Look for earlier posts on multiple platforms.
Pay attention to timestamps, but remember that screenshots of timestamps can be fabricated and platforms may display times differently by location.
An earlier upload may reveal the real event or show that the video predates the current claim.
Step 4: Search key frames
Capture several clear frames from different points in the video.
Choose images showing:
-
faces;
-
landmarks;
-
signs;
-
vehicles;
-
buildings;
-
unusual objects;
-
wide views of the scene.
Use image-search tools to see whether the same frames appeared in older reports.
Do not rely on only one frame. A video may combine material from several sources.
Step 5: Inspect Content Credentials
Look for a Content Credentials symbol or use a compatible inspection service.
The public verification tool supports several common image, video, audio, and document formats. When credentials are present, it may show creation and editing information.
Interpret the result carefully:
“Credential found” does not mean every associated claim is true.
“No Content Credential” does not mean the file is fake or genuine. It simply means that no supported credential was available to the inspector.
Step 6: Check for supported AI watermarks
When you suspect that a clip was created with Google AI, SynthID detection may be useful. Google says users can upload supported media to Gemini and ask whether SynthID indicates that it was created or altered using Google AI.
A positive result is evidence of supported Google AI involvement.
A negative result does not exclude other AI systems.
Step 7: Verify the location
Search for distinctive landmarks or text visible in the scene.
Compare the video with:
-
maps;
-
photographs;
-
satellite imagery;
-
local business listings;
-
street layouts;
-
official building images.
Check whether traffic moves on the expected side of the road and whether license plates, signs, architecture, and public infrastructure fit the claimed country.
Step 8: Check the date and weather
Compare the visible conditions with historical weather data.
Ask:
-
Was it raining at the claimed time?
-
Was there snow?
-
Was the sky clear?
-
Does the direction of sunlight make sense?
-
Were trees in season?
-
Was the building shown already constructed?
-
Did the advertised event occur on that date?
These details are particularly useful when old footage is being recycled.
Step 9: Look for independent confirmation
A dramatic event should usually create evidence beyond one viral video.
Search for confirmation from:
-
local authorities;
-
emergency services;
-
established local journalists;
-
residents;
-
transport agencies;
-
weather services;
-
hospitals;
-
nearby businesses;
-
multiple eyewitnesses.
Be cautious when dozens of websites repeat the same unsupported social-media post. Repetition is not independent confirmation.
Step 10: Examine the full context
Watch the complete video rather than the shortest edited version.
A clip may end immediately before an important explanation, crop out a sign identifying the real location, or remove the moment showing that a supposed emergency was staged.
Context can be more important than detecting pixel manipulation.
Warning signs that still deserve attention
Although visual inspection is insufficient by itself, several signs should raise suspicion.
Unreadable or unstable text
AI-generated signs, banners, labels, and screens may contain letters that change, merge, or lack linguistic meaning.
Objects that transform between frames
Watch glasses, earrings, microphones, fingers, buttons, vehicles, and background faces.
Unnatural movement
People may glide, merge into crowds, move with unusual weight, or interact incorrectly with objects.
Incorrect reflections and shadows
Mirrors, windows, water, and polished surfaces may not match the main action.
Overly smooth skin or facial motion
Some synthetic faces appear polished, symmetrical, or strangely detached from the lighting around them.
Repeated background patterns
Crowds, windows, trees, or objects may be duplicated.
Audio without environmental consistency
Speech may sound unusually clean for a noisy location, or background sounds may repeat in loops.
None of these signs is conclusive. Real video can contain compression artifacts, stabilization errors, poor synchronization, or unusual lighting.
Deepfake video calls and impersonation scams
AI video deception is not limited to public misinformation.
Criminals can use cloned voices, stolen photographs, synthetic video, and compromised accounts to impersonate:
-
relatives;
-
company executives;
-
bank employees;
-
lawyers;
-
police officers;
-
government officials.
A victim may receive a call in which a familiar-looking or familiar-sounding person claims to need urgent money.
In these cases, do not rely on appearance or voice alone.
End the call and contact the person through a known number. Ask a question based on private information that is not easily available online. Confirm payment requests through a separate communication channel.
Technical provenance systems may eventually become more common in calling and messaging platforms, but basic identity-verification habits remain essential.
The privacy problem with provenance
Detailed provenance creates its own risks.
A record that shows where, when, and by whom media was captured could expose:
-
journalists working in dangerous areas;
-
whistleblowers;
-
political dissidents;
-
victims of violence;
-
confidential sources;
-
military or humanitarian operations.
A successful provenance system must therefore balance transparency with privacy and safety.
Not every data field should always be public. Some information may need to be removed, concealed, generalized, or selectively disclosed while still preserving evidence that the media passed through a trustworthy process.
Why universal labeling remains difficult
A global system for identifying AI-generated video faces several obstacles.
Not every company participates
A voluntary standard cannot identify content produced by tools that do not support it.
Open-source models can be modified
Developers may remove labeling or watermarking features.
Platforms may strip information
Uploading and re-encoding can remove attached metadata.
Adversaries actively attack detection systems
Someone creating disinformation has an incentive to alter, crop, record, or reconstruct content until a detector fails.
Authentic media can be falsely contextualized
No watermark can determine whether a truthful recording has been paired with a dishonest caption.
Labels can be misunderstood
Some viewers may treat “AI used” as equivalent to “false,” even when AI was used only for harmless editing, translation, accessibility, or background cleanup.
Others may treat “no AI detected” as proof of authenticity.
Both conclusions can be wrong.
The liar’s dividend
The growth of synthetic media creates another problem: real footage can be dismissed as fake.
A public figure confronted with genuine evidence may claim that the recording was generated by artificial intelligence. This phenomenon is sometimes called the liar’s dividend.
As awareness of deepfakes grows, people may become more willing to reject authentic evidence that conflicts with their beliefs.
Provenance systems can help counter this problem by providing stronger evidence of origin and editing history. However, they will work best when adoption is broad, verification interfaces are understandable, and trusted institutions explain their conclusions transparently.
A better way to think about authenticity
Authenticity should not be treated as a simple choice between “real” and “fake.”
A video can be:
-
genuinely captured and accurately described;
-
genuine but edited;
-
genuine but falsely captioned;
-
partly generated and clearly disclosed;
-
partly generated and deceptively presented;
-
completely synthetic;
-
technically authentic but staged;
-
authentic in origin but misleading through omission.
The correct conclusion may therefore be more specific than “fake.”
Examples include:
“The video is genuine, but it was filmed in 2022 rather than today.”
“The footage is authentic, but the audio was replaced.”
“The clip contains an AI-generated background.”
“The original source cannot be established.”
“No supported watermark was found, but that does not verify the footage.”
“The location is correct, although the caption misidentifies the event.”
Precise language is one of the most important parts of responsible verification.
Frequently asked questions
Can an AI watermark be removed?
Some watermarks are designed to survive common editing, but no watermark should be assumed to withstand every possible transformation. Heavy cropping, reconstruction, repeated compression, screen recording, frame replacement, or deliberate adversarial processing may reduce detectability.
Does every AI video contain a hidden watermark?
No. Watermarking depends on the tool and how it was configured. Many systems do not add a universally detectable mark.
Does a missing watermark mean a video is real?
No. It means only that the detector did not find the watermark it was designed to recognize.
Can Content Credentials be faked?
C2PA uses cryptographic signatures to make unauthorized changes detectable, but the system does not guarantee that the person or organization creating the signed assertions was truthful. Trust also depends on who signed the credential and how their identity was validated.
Are Content Credentials the same as copyright information?
No. They may include information about a creator or editing history, but their primary purpose is provenance and transparency rather than serving as a complete copyright registry.
Can a genuine video still be misinformation?
Yes. Genuine footage can be mislabeled, taken out of context, edited selectively, or presented with a false claim.
Can AI detectors prove that a video is fake?
Most detectors provide signals or probability assessments rather than absolute proof. Their results should be combined with source, context, provenance, and corroborating evidence.
What is the easiest check an ordinary viewer can perform?
Try to locate the earliest version and determine whether reputable, independent sources confirm the event. This often reveals more than searching for visual defects.
The most reliable defense is a layered one
There is unlikely to be a single tool capable of examining every video and delivering a flawless “real” or “fake” verdict.
Hidden watermarks can identify content produced by participating AI systems. Content Credentials can reveal a signed creation and editing history. Metadata can offer clues about devices, software, dates, and locations. Frame analysis can expose inconsistencies. Reverse searching can reveal older versions. Geolocation can test whether the scene matches the claimed place. Independent reporting can establish whether the event actually occurred.
Each method has limitations.
Used together, however, they create a far stronger verification process.
That is the central lesson behind the work of BBC Verify and other professional fact-checking teams: authenticity is not established by staring at a video until something looks wrong. It is established by reconstructing the media’s history, testing its context, examining available technical evidence, and comparing the claim with the world outside the frame.
As AI-generated video becomes more convincing, viewers will need to change their habits. The most important question will not be whether a clip looks real.
It will be whether its origin, history, context, and supporting evidence can be verified.
