A convincing video clip is no longer proof that an event happened as shown. The same problem applies to photographs and increasingly realistic voice recordings. For journalists, that changes a basic newsroom task: deciding whether media is trustworthy enough to report.
Deepfakes don't make verification impossible, but they do make visual inspection alone less reliable. Newsrooms need repeatable methods for checking suspicious media, protecting sources, and avoiding both false publication and false accusations.
Deepfakes turn media verification into a source problem
The term deepfake usually refers to synthetic or manipulated audio, video, or imagery created with machine learning. A person's face can be altered, speech can be generated in their voice, or an entirely artificial image can be presented as documentary evidence.
Journalists should also distinguish deepfakes from simpler manipulation. A genuine video can be cropped, slowed down, paired with unrelated audio, or reposted with a false date or location. These "cheapfakes" may require less technical skill but can be just as misleading in a breaking-news cycle.
That distinction matters because searching for obvious signs of artificial generation can distract from a more basic question: Where did the material come from?
Consider a 20-second video appearing to show a public official making a controversial statement. Several explanations are possible. The video could be synthetic. The speech might be genuine but edited. The footage could be several years old. It might show another person, or the accompanying caption might simply misrepresent what happened.
A newsroom therefore shouldn't begin with "Is this a deepfake?" The better starting questions are:
- Who first published or supplied the file?
- Can the newsroom obtain an earlier or original version?
- When and where was it supposedly recorded?
- Who else witnessed the event?
- Is there independent evidence that the event occurred?
- Has the person or organisation depicted been contacted for verification?
These questions remain useful even when the manipulation technique changes. They also prevent a common mistake: assuming that strange-looking media must be artificially generated.
Detection tools should support verification, not replace it
Deepfake detectors can add another layer of evidence when a newsroom receives questionable media. Depending on the system, detection may examine faces, frames, audio characteristics, compression patterns, or other signals associated with synthetic generation.
For organisations handling media verification at scale, a deepfake detection API can be incorporated into an intake or review process so suspicious files receive automated analysis before publication. That can be useful when journalists are dealing with large volumes of submitted video or audio, particularly during elections, conflicts, natural disasters, or other fast-moving stories.
The important part is what happens after the result arrives. A detector's score should be treated as evidence, not a verdict.
A low-confidence result doesn't establish authenticity. Media may have been generated by a method the system handles poorly, or later editing and compression may obscure useful signals. Conversely, a detector may flag genuine footage after it has been resized, transcoded, filtered, or repeatedly downloaded and uploaded.
Newsrooms should also avoid telling readers that something is a deepfake solely because software labelled it that way. Making a public accusation raises a higher evidentiary bar than deciding internally that a file requires more investigation.
A stronger workflow combines technical analysis with traditional reporting. Contact the apparent speaker. Locate other recordings of the same event. Compare environmental details with known locations. Examine earlier versions of the file. Ask the source how it was obtained.
The question isn't whether a detector can replace a reporter. It can't. The useful question is whether it can reveal enough uncertainty to prevent an unverified clip from becoming tomorrow's correction.
Verify the source, content, and surrounding context separately
Deepfake investigations become easier to manage when journalists divide the problem into three parts: provenance, media analysis, and corroboration.
Provenance concerns the history of the file. A video forwarded through several messaging apps has a weaker chain of custody than footage obtained directly from the person who recorded it. Reporters should preserve the original file whenever possible rather than relying on a screen recording or social-media download.
Media analysis examines the recording itself. Reporters can look for discontinuities in lighting, reflections, facial movement, lip synchronization, background sound, frame transitions, or speech cadence. None of these observations proves manipulation by itself. They are reasons to investigate further.
The FBI's guidance on synthetic content notes that visual distortions, unnatural movement, irregular pauses, mismatched background noise, and unusual lighting can be warning signs. The agency also makes an important broader point: synthetic-content tools have become easier for more people to access.
Corroboration is often the strongest part of the process. If a video supposedly shows an explosion outside a government building, a newsroom can check whether local reporters, emergency agencies, eyewitnesses, traffic cameras, satellite imagery, or other independent evidence supports the claim.
The same principle works for audio. Imagine that a reporter receives a recording in which a company chief executive apparently announces layoffs before an official statement. Instead of concentrating only on whether the voice sounds artificial, the newsroom can ask whether the source attended the meeting, whether employees heard the same announcement, whether the company confirms it, and whether internal documents support the recording.
Deepfake verification works best when no single clue carries the whole decision.
Newsrooms need a process before suspicious media arrives
Verification becomes much harder when editors are designing the procedure while a questionable clip is already spreading.
A practical newsroom policy should define who evaluates suspicious media and what happens while that review takes place. Reporters need to know whether they should send questionable files to a specialist, security team, visual investigations desk, or designated editor.
The process should also preserve evidence. Keep the file originally supplied to the newsroom, along with available metadata, source messages, timestamps, and correspondence. Avoid repeatedly modifying the only copy during analysis.
For high-impact material, use a simple escalation threshold. Media deserves additional scrutiny when it could materially affect an election, public safety, financial markets, someone's reputation, or an ongoing conflict. The same applies when the source is anonymous, newly created, unwilling to explain the file's origin, or pressing journalists to publish immediately.
Speed still matters in journalism, but "we haven't verified this footage" is often a better editorial decision than racing competitors with material that later proves false.
There is another reason to formalise the process: genuine media can now be dismissed as artificial.
A politician, executive, or other public figure may claim that an authentic recording is a deepfake. This is sometimes called the liar's dividend, where the existence of convincing synthetic media creates plausible deniability around real evidence. Journalists therefore need enough documentation to defend an authenticity finding as carefully as a manipulation finding.
Deepfakes are also a newsroom security risk
Synthetic media isn't only something journalists investigate. It can be used against journalists themselves.
An attacker could clone an editor's voice and call a reporter. Someone could impersonate a source during a remote interview. A fake executive could request confidential information, while a fabricated audio message could appear to authorise an unusual payment or account change.
That makes deepfake awareness part of operational security.
Newsrooms should establish secondary verification for unusual requests. If a familiar voice suddenly asks for unpublished documents, credentials, money, or sensitive source information, staff should confirm the request through another trusted channel. A known telephone number, established messaging account, or previously agreed verification phrase can be more useful than asking whether the voice "sounds right."
Journalists covering sensitive subjects should consider similar safeguards with important sources. Agreeing beforehand on how identity will be confirmed can reduce confusion if an impersonator later attempts contact.
Public-facing newsroom material deserves attention too. Interviews, podcasts, conference appearances, and social-media videos provide large amounts of voice and visual material that can potentially be reused to construct impersonations. Removing all such material is neither realistic nor desirable for most journalists. Awareness and verification procedures are more practical defenses.
Report uncertainty as carefully as you report manipulation
There will be cases where a newsroom cannot conclusively determine whether material is genuine.
The publication language should reflect that uncertainty. "We have not independently verified the recording" is different from "the recording is fake." Likewise, evidence that a file has been edited does not necessarily show that the central event depicted never occurred.
If manipulated media becomes part of a story, explain what the newsroom actually established. Was the audio generated? Was the video altered? Was authentic footage paired with a fabricated description? Was the original source impossible to identify?
Precision matters because "deepfake" can easily become a catch-all label for anything doubtful online. That weakens readers' understanding of both the technology and the evidence.
Corrections also need care. If synthetic or misleading media was published before verification was complete, removing it without explanation may leave copies circulating elsewhere. A clear correction can document what was initially reported, what subsequent verification established, and which part of the original claim was wrong.
Verification matters more as synthetic media improves
Journalists don't need to become machine-learning engineers to report responsibly on deepfakes. They do need to stop treating a convincing recording as self-authenticating evidence.
The strongest newsroom response combines source verification, technical analysis, corroborating reporting, and clear editorial thresholds. Detection software can contribute to that process, but it works best as one signal among several.
As artificial media becomes harder to judge by sight or sound alone, the old reporting discipline becomes more valuable: establish where the evidence came from, verify it independently, and don't claim more certainty than the facts support.