For generations, a photograph carried a simple kind of persuasive power.
A witness could lie.
A person's memory could fail.
But a photograph seemed different.
It showed something that was actually in front of a camera.
Audio and video became even more compelling. A jury could hear someone's voice or watch an event unfold rather than relying entirely on another person's description of it.
Artificial intelligence is weakening that assumption.
A realistic photograph can now depict an event that never occurred.
A person's voice can be recreated saying words the person never spoke.
Video can place a recognizable face on another person's body or generate a scene from scratch.
And the technology required to do it is increasingly available to ordinary users rather than sophisticated visual-effects studios.
That does not mean photographs, audio recordings and videos have suddenly become useless as evidence.
It means courts increasingly have to confront a question that once arose much less often:
How do we know this recording is real?
The problem has become significant enough that the federal judiciary's Advisory Committee on Evidence Rules is actively considering a new rule specifically addressing evidence alleged to have been fabricated using generative artificial intelligence.
Seeing Is No Longer Enough
Imagine prosecutors possess a video that appears to show a defendant meeting another person in a parking lot.
The defendant says:
“That isn't me. The video was generated with artificial intelligence.”
Now reverse the situation.
The defense has an audio recording that appears to contain another person confessing to the crime.
Prosecutors respond:
“That recording is fake.”
A generation ago, creating a convincing fabrication of either recording could require substantial expertise.
Today, the allegation is at least technologically plausible.
That creates two different problems.
The first is obvious: fake evidence can be presented as real.
The second is more subtle: real evidence can be dismissed as fake.
Courts have to deal with both.
Evidence Has Always Needed Authentication
Deepfakes did not invent the problem of fraudulent evidence.
People have altered documents for centuries.
Photographs can be edited.
Recordings can be cut.
Signatures can be forged.
Physical evidence can be mislabeled.
That is one reason courts have authentication rules.
Under Federal Rule of Evidence 901, a party offering an item generally must produce evidence sufficient to support a finding that the item is what the party claims it is.
Suppose someone offers a photograph of a damaged vehicle.
A witness who saw the vehicle might testify that the photograph accurately depicts it.
A security recording may be authenticated through testimony about the camera system and how the recording was obtained.
Electronic information can sometimes be authenticated through distinctive characteristics, metadata, system evidence or surrounding circumstances.
Authentication does not ordinarily mean proving beyond all doubt that an item is genuine.
It means establishing the evidentiary foundation required by the applicable rules.
Deepfakes complicate that process because a fabricated item can increasingly contain many of the characteristics people traditionally associate with authentic evidence.
A Fake Recording Can Look Completely Ordinary
Poor fakes can be easy to recognize.
A face moves unnaturally.
A hand has the wrong number of fingers.
Lighting changes strangely.
The voice sounds artificial.
The background contains impossible objects.
Those errors can create a false sense of security.
Generative technology keeps improving.
The more important legal problem is not the fake everyone immediately recognizes.
It is the one that looks ordinary.
A convincing fabricated recording may contain realistic facial movement, natural speech, plausible shadows and believable surroundings.
If jurors cannot reliably identify the fabrication simply by watching it, the authenticity dispute must be resolved through other evidence.
That may include where the file originated, who possessed it, metadata, other recordings, witness testimony and technical analysis.
The Federal Courts Are Already Preparing for This Problem
The federal judiciary has not treated deepfakes as a distant hypothetical.
In 2026, the Advisory Committee on Evidence Rules was considering proposed Rule 901(c), specifically addressing evidence challenged as fabricated through generative AI.
The working proposal establishes a two-stage process.
First, the party claiming an item is an AI fabrication would need to produce enough evidence to justify an authenticity inquiry. A bare assertion that something might be a deepfake would not automatically be enough.
If that initial showing is made, the burden would shift to the party offering the evidence to demonstrate to the judge that the item is more likely than not authentic.
As of September 2026, this remains part of the federal rulemaking process rather than an already-effective nationwide Federal Rule of Evidence. The judiciary's current published Federal Rules of Evidence were last amended in 2024.
The fact that a special rule is being considered nevertheless shows how seriously the judiciary is taking the problem.
Courts Cannot Require Proof Every Time Someone Says “AI”
There is an obvious danger in going too far in the other direction.
Suppose authentic video clearly shows a person committing a crime.
The defendant could simply say:
“It's a deepfake.”
If that unsupported statement automatically required an expensive technical investigation, almost any digital recording could become difficult to use.
That is one reason the proposed federal approach requires the challenger to make an initial showing suggesting fabrication before imposing a higher authentication burden on the party offering the evidence.
The system has to account for both possibilities.
A deepfake accusation can be legitimate.
It can also be a convenient way to attack genuine evidence.
Real Evidence Can Now Be Attacked Because Fake Evidence Exists
This phenomenon is sometimes called the liar's dividend.
Once people know convincing fake recordings exist, the existence of the technology itself can create doubt about authentic material.
Imagine a genuine voicemail.
The speaker admits something damaging.
Years ago, the dispute might have focused on what the speaker meant.
Now the speaker can potentially argue that the voice was artificially generated.
The same problem applies to video.
A recording can be authentic and still face an accusation of manipulation.
The existence of deepfake technology therefore affects more than fabricated evidence.
It changes the environment in which genuine evidence is evaluated.
The Original File Can Become Extremely Important
A video forwarded through several messaging applications may be harder to investigate than the original recording.
The original file can contain technical information that copies do not preserve.
Metadata may indicate when the file was created.
File structure may provide information about the recording device or software.
A device may contain related photographs or videos captured immediately before and afterward.
Cloud records may help establish when the file was uploaded.
A continuous series of recordings may make fabrication less plausible.
None of these factors automatically proves authenticity.
Metadata can itself be manipulated.
But provenance—the history of where evidence came from and how it traveled—can become increasingly important when the content itself can be convincingly manufactured.
A Camera Can Have a Chain of Evidence
Consider a security camera outside a store.
It records continuously.
At 8:13 p.m., a vehicle arrives.
At 8:15, someone enters the building.
At 8:19, the person leaves.
At 8:20, the vehicle drives away.
Investigators obtain the recording directly from the camera system.
Other cameras show portions of the same event.
A witness remembers the vehicle.
A transaction inside the store occurred at 8:16.
Those independent pieces of evidence make the recording very different from an isolated video file anonymously emailed to investigators.
The more independently verifiable connections an item has to the surrounding world, the harder it can be to fabricate the entire evidentiary picture.
Deepfake analysis therefore may involve more than inspecting pixels.
It can involve reconstructing the history of the evidence.
Multiple Cameras Can Authenticate One Another
Modern surveillance creates an interesting defense against modern fabrication.
Suppose one camera records a person crossing a parking lot.
A second camera records the person entering a building.
A doorbell camera across the street records the person's vehicle arriving.
A traffic camera captures the vehicle several minutes earlier.
Fabricating one recording may be possible.
Creating several independently sourced recordings that fit together in time and space is a different challenge.
That does not make coordinated fabrication impossible.
But corroboration can become powerful evidence of authenticity.
The digital world that makes deepfakes possible also produces enormous amounts of independent information capable of testing them.
Audio May Be Especially Difficult
A voice can be powerful evidence because people recognize familiar speakers instinctively.
AI voice cloning complicates that intuition.
With enough source material, synthetic systems can imitate a person's speech characteristics.
A fabricated voice can potentially be used to create a confession, threat, agreement or apparently incriminating conversation.
Audio presents an additional problem because listeners cannot inspect facial movement or visual inconsistencies.
Investigators may instead examine the original file, recording circumstances, background sounds, metadata and other technical characteristics.
They may also compare the alleged conversation with surrounding events.
Did the call actually occur?
Do telephone records support it?
Did another participant possess the recording?
Was the speaker physically capable of participating at that time?
The content is only one part of the analysis.
Deepfakes Can Affect Civil Cases Too
The issue extends far beyond criminal prosecutions.
Imagine a business lawsuit involving a disputed oral agreement.
One side produces an audio recording in which an executive appears to accept particular terms.
Or consider a divorce case involving an allegedly compromising video.
An employment case might include an audio recording of a supervisor making discriminatory remarks.
A defamation lawsuit could turn on whether someone actually made a recorded statement.
An insurance dispute could involve images purporting to show damaged property.
As synthetic media improves, authenticity disputes can arise anywhere digital evidence matters.
Social Media Makes Provenance Harder
Evidence discovered on social media can present special problems.
A video may have been downloaded and reposted repeatedly.
The person presenting it may not know who originally recorded it.
Compression may remove metadata.
Editing tools may alter the file.
The original account may disappear.
Hundreds of copies can circulate while the source becomes increasingly difficult to identify.
Now add generative AI.
A realistic fabricated video can spread through the same network.
By the time someone tries to use it in court, the most basic question—where did this come from?—may be surprisingly difficult to answer.
That makes early preservation of original material increasingly important.
A Screenshot of a Deepfake Is Still a Deepfake
Copying fabricated evidence does not make it authentic.
Suppose someone creates a fake social-media post and then takes a screenshot of it.
The screenshot may genuinely be a screenshot.
But that does not prove the underlying post was genuine.
Likewise, a person can make a screen recording of fabricated video.
The screen recording accurately depicts what appeared on the screen.
The underlying content can still be false.
Courts therefore sometimes have to distinguish between authenticating the container and authenticating what the container purports to show.
A real photograph of a fake document does not transform the document into a real one.
Expert Analysis May Become More Common
Some authenticity disputes can be resolved through ordinary witnesses and surrounding circumstances.
Others may require technical expertise.
A forensic examiner might analyze file structure, compression, metadata or signs of manipulation.
Experts may disagree.
Detection technology itself can have error rates.
And there is an inevitable technological race involved.
As deepfake generators improve, detection tools attempt to identify them.
As detection improves, generation techniques adapt.
That makes a courtroom rule based solely on “run it through a deepfake detector” unlikely to solve the problem permanently.
The federal evidence committee itself has discussed the difficulty of detecting sophisticated deepfakes and whether judges should take a stronger gatekeeping role when a credible fabrication challenge is raised.
Jurors May Be Particularly Vulnerable to Convincing Video
Evidence does not affect people equally.
A spreadsheet may require explanation.
A witness's recollection can be questioned.
Video feels immediate.
Jurors can watch the event themselves.
That persuasive force is exactly what makes fabricated video dangerous.
The federal evidence committee has discussed the concern that realistic AI-generated audio and visual evidence can have an unusually powerful effect on jurors even when authenticity is disputed.
Once someone has watched a convincing video, simply being told later that it might be artificial may not erase the impression.
That is one reason authenticity questions can matter before evidence reaches a jury.
Deepfakes Will Not Make Digital Evidence Meaningless
It is easy to imagine an extreme future in which nobody can believe any photograph, recording or video.
That conclusion goes too far.
Digital evidence does not exist in isolation.
A genuine photograph may have an original file.
A camera system may document its creation.
Other cameras may capture the same event.
Witnesses may confirm what happened.
Location information may correspond with the recording.
Communications may discuss the event immediately afterward.
Technical examination may find no evidence of manipulation.
Authenticity can be built from multiple independent facts.
In many cases, the surrounding evidence may be far harder to fabricate than a single image.
The Strongest Evidence May Become Evidence With a History
For decades, people often asked:
“What does the video show?”
Increasingly, another set of questions may come first.
Who recorded it?
Where is the original?
What device created it?
When was it created?
Who possessed it afterward?
Was it edited?
Does another recording show the same event?
Do independent records agree with it?
Is there evidence that generative AI was involved?
Those questions describe something larger than the content of the file.
They describe its history.
And that history may become one of the strongest ways of distinguishing genuine digital evidence from a convincing fabrication.
Courts Are Entering an Era in Which Reality Itself May Be Contested
For most of the history of photographic evidence, fabrication required effort.
That provided a practical safeguard even when the legal rules did not change.
Generative AI is reducing that safeguard.
The federal judiciary has already recognized the significance of the change. Its evidence committee has said that deepfakes are becoming easier to create and harder for ordinary jurors to detect, while debating whether the existing authentication framework needs a special rule for AI fabrication.
The challenge is not merely determining whether artificial intelligence can create a convincing fake.
It can.
The more important question is how courts distinguish that fake from genuine evidence without making every authentic photograph, recording and video presumptively suspicious.
For the legal system, that may become one of the defining evidence problems of the AI era.
The courtroom of the future may not only have to determine what happened. It may first have to determine whether the evidence showing what happened ever existed in reality at all.