What Happens When Reality Becomes Harder to Verify Than to Fake?

A video from the library committee meeting appears in a neighborhood group chat on Thursday morning. The chair sits in the familiar room, beside the familiar clock, and says the Saturday closure was settled before anyone voted.

The clip lasts nineteen seconds. It is sharp enough to show a twitch in the chair’s face and quiet enough to hear the air conditioner. People who attended the meeting recognize the room. One member forwards it with the words: I knew something was wrong.

The chair says the recording is fabricated. Someone else says that is exactly what a guilty person would say. A third person finds a longer recording, but nobody knows who made it. The families who use the reading room want an answer before Saturday, when the closure takes effect.

This is a fictional continuation of the library committee in the previous two investigations. There is no hidden real recording to uncover. The difficulty it raises is genuine: what happens when a convincing image or voice can be produced in minutes, while establishing whether an event occurred may require witnesses, source files, technical knowledge, and time?

The dangerous moment arrives when the ability to doubt evidence spreads faster than the ability to examine it.

The Hidden Assumption

We have long treated a recording as a relatively direct route from an event to our judgment. A witness could misremember; a person could lie; a written account could omit inconvenient details. A camera seemed to preserve something outside the narrator’s control.

That expectation was never absolute. Photographs could be staged or edited, audio could be spliced, and even perfectly genuine footage could mislead through framing. What is changing is the cost of creating a plausible record of an event that never happened.

Underneath the question about fakes lies a second assumption: that the person who knows what happened can usually demonstrate it. But knowledge, evidence, and the public ability to verify evidence are different things. The committee might know the vote was honest and still fail to persuade its neighbors; a dishonest committee might publish authentic footage of a carefully rehearsed meeting.

So what exactly should we try to verify? The pixels? The file’s history? The spoken claim? Or the event itself?

Digital forensics and philosophy of evidence

Four Questions Hidden Inside the Word ‘Real’

Suppose an original video file is available. We can ask whether its bytes have changed since a particular checkpoint. A secure hash can help answer that narrow question. We can ask which device or account recorded or exported it, and whether the claimed origin can be cryptographically authenticated. We can ask what the visible scene actually shows. Finally, we can ask whether the conclusion drawn from the scene is true.

These checks can disagree without contradiction. A recording may be authentic and show a staged confession. An unaltered clip may omit the next sentence, in which a speaker makes clear that the apparent confession was a joke. An edited montage may contain entirely genuine moments arranged into a false chronology.

In courtrooms, laboratories, and investigations, the same basic problem appears in different clothing: a chain of custody supports the account of how evidence was handled, but relevance and meaning still require argument. The chain itself is only as dependable as its procedures and participants. This is an analogy to evidentiary practice, not a claim that online platforms follow legal standards.

The first improvement is to stop asking a single yes-or-no question of a complicated object. We can verify the integrity of a file and remain uncertain about the truth of the claim attached to it.

Perceptual science

Your Eyes Are Not a Provenance System

When the library clip looks convincing, viewers may treat their confidence as a measurement. Yet experiments put limits on that intuition.

In a study published in Communications of the ACM in 2025, 1,276 participants judged authentic and synthetic material across images, audio, video, and audiovisual examples. Mean performance was close to chance in the study conditions. That does not mean every person faces fifty-fifty odds with every clip. Stimulus selection, format, and the tests’ online-like presentation all matter. It means visual or auditory familiarity alone can fail in important settings.

A 2025 Scientific Reports study of AI-generated voice clones found that participants often heard a cloned voice as the real speaker. On average, they correctly identified a voice as AI-generated only about sixty percent of the time in the reported tests. A voice that sounds like someone we know is therefore evidence of resemblance, with no automatic guarantee of origin.

There is a useful complication. Five preregistered experiments on political speech deepfakes, published in Nature Communications in 2024, found that audiovisual information sometimes helped people distinguish genuine speech from fabrications better than text alone. The studies were not a verdict that video is useless; they measured particular stimuli and decisions.

The conclusion is neither ‘trust your eyes’ nor ‘your eyes tell you nothing.’ Perception remains part of inquiry. It cannot safely carry the entire burden.

Cooke and colleagues (2025), As Good as a Coin Toss · Communications of the ACM

Barrington, Cooper & Farid (2025), People Are Poorly Equipped to Detect AI-Powered Voice Clones · Scientific Reports

Groh and colleagues (2024), Human Detection of Political Speech Deepfakes · Nature Communications

Political science and strategic doubt

The Advantage of Denying What Really Happened

A convincing fabrication can falsely accuse someone. The same technology can also help a person deny an authentic record. Researchers call this second opportunity the liar’s dividend: when people know media can be manipulated, a person confronted with damaging evidence may describe it as fake.

An experiment published in The Journal of Politics in 2025 found that political deepfakes in the tested setting were about as credible as misinformation delivered through other formats. In a subsequent task, participants sometimes misidentified authentic scandal videos as deepfakes, especially when the politician belonged to their own party. That finding supports a specific risk of motivated doubt. It does not establish that most genuine videos will be rejected.

A separate American Political Science Review study used five survey experiments involving more than 15,000 American adults. Politicians’ claims that scandal reports were misinformation increased support across partisan subgroups in its experimental scenarios. Crucially, the researchers found this tactic worked better against text-based reports than against video evidence. The liar’s dividend is a documented possibility with limits, not a universal solvent for accountability.

Back in the neighborhood, one group treats the video as conclusive because it confirms what they suspected. Another group dismisses it because they like the chair. Neither has yet inspected the origin. Two competing certainties can be produced from the same absence of verification.

Barari, Lucas & Munger (2025), Political Deepfakes Are as Credible as Other Fake Media · The Journal of Politics

The Liar’s Dividend: Can Politicians Claim Misinformation to Evade Accountability? (2025) · American Political Science Review

Information security

What a Digital Signature Can—and Cannot—Settle

There are practical ways to give media a more traceable history. Standards such as C2PA allow participating tools to attach cryptographically protected records about a file’s origin and edits. A verifier can check whether the signed assertions match the asset and whether a trusted signing credential was used. The Coalition for Content Provenance and Authenticity describes the trust model and the checks in its official technical FAQ.

This matters. A photographer, newsroom, or institution that keeps an accountable record from capture through publication may be able to answer questions that a viral screenshot cannot. If a copy has been altered outside the recorded process, a signature check can reveal a mismatch.

But a signature does not certify that the scene was spontaneous, that the camera operator was honest, or that the accompanying caption is correct. Microsoft’s provenance guidance explicitly distinguishes origin and processing history from truth, authorship, and intent. A perfectly signed file could record an actor pretending to be the committee chair. An honest witness could also circulate an unsigned recording.

Metadata may disappear when material is re-encoded, cropped, or turned into a screenshot. A missing credential therefore cannot be read as a finding of fabrication. Inversely, a valid credential answers questions about the claimed provenance within a particular trust system; it is no substitute for corroborating what happened.

The metaphor of a sealed envelope is helpful only to a point. A seal can indicate that someone has not opened the envelope since it was sealed. It cannot tell us whether the letter inside was truthful when written.

Coalition for Content Provenance and Authenticity, official FAQ

Microsoft Learn, Content Provenance: what provenance does and does not establish

Technical measurement and uncertainty

Why a Detector Is Not a Judge

Another proposal is to run every suspicious clip through an AI detector and publish a score. Some detectors find real signals in particular kinds of synthetic content, but their performance depends on the generator, the media type, the transformations applied, and the distribution of examples used for testing.

The US National Institute of Standards and Technology examined provenance tracking, watermarking, synthetic-content detection, and evaluation in its November 2024 report on digital content transparency. Its treatment of several technical approaches is itself a warning against reducing the problem to one instrument.

An output such as ‘eighty percent likely synthetic’ is not a universal probability of deception. To interpret it responsibly, we need to know the tested population, the false-positive rate, and what the tool labels as synthetic. A legitimate photograph may have had an AI-assisted enhancement. A wholly staged recording may contain no AI-generation artifacts at all.

A detector may be useful as one lead in an investigation. It is risky as an automatic verdict on a person's credibility or a justification for suppressing an allegation. Technical uncertainty becomes a question of fairness as soon as a score changes what a person is allowed to say or whom an audience will believe.

NIST (2024), Reducing Risks Posed by Synthetic Content, AI 100-4

Journalism and institutions

Trust Is a Public Resource—and Publics Disagree

In the Reuters Institute’s 2026 Digital News Report, sixty-two percent of surveyed respondents across its markets expressed concern about what is real and fake in online news, up four percentage points from 2025. The report also recorded overall trust in news at thirty-seven percent, its lowest level in the survey series since 2015.

Those are survey measures of concern and reported trust. They do not show what percentage of news is fabricated, and they cannot establish that generative AI caused the decline. Institutional trust, partisan conflict, changing news habits, and other pressures are entangled.

Nevertheless, the figures describe a difficult environment for verification. A credible correction may never reach the people who saw the original clip. A newsroom can establish where footage came from and still face an audience that distrusts the institution. A claim can remain popular because it expresses a group's fears even after its factual premise weakens.

Some people respond by calling for a single authoritative source to certify every important recording. That could reduce confusion in well-governed contexts, but it also concentrates power. What happens when the authority makes an error, faces political pressure, or excludes people without access to compliant devices?

The alternative is not universal suspicion. It is a network of methods and accountable institutions that can be examined, challenged, and corrected.

Reuters Institute (2026), Digital News Report — overview and key findings

Access and responsibility

The Cost of Proving an Ordinary Event

Consider two witnesses to the same event. One works for a large media organization with source-file retention, lawyers, trained investigators, and signing equipment. The other is a commuter with an inexpensive phone who sends a compressed clip before a platform deletes the account.

A verification system that recognizes only the first person's evidence may protect against some fabrications while creating a new problem of unequal credibility. The absence of perfect provenance often reflects mundane technical conditions, not dishonesty. This is a foreseeable design trade-off, not an experimentally measured inequality asserted for every platform.

Asking a person to produce more proof can be appropriate when they make a serious accusation. Yet we should distinguish proportional scrutiny from an impossible demand. Some events leave witnesses but no preserved file. Others leave digital traces whose collection would expose vulnerable people.

A sensible process therefore needs more than technical authenticity. It should permit independent corroboration, preserve uncertainty without treating it as guilt, explain how a finding was reached, and provide a way to correct mistakes. The quality of the process matters especially when a finding affects reputation, safety, or public decisions.

It may soon be cheap to generate a clip that interrupts someone's life. The hours required to investigate it still have to be paid for by somebody.

Where the Fields Collide

Computer security asks whether a record has an authenticated history. Perceptual science asks how people judge what they see and hear. Political science asks who benefits when a record can be accepted or denied according to allegiance. Journalism asks how witnesses and independent documents can be brought into a public account. Each field supplies a necessary part of the answer, and each is capable of overclaiming when used alone.

A cryptographic signature may make a file's history easier to inspect but cannot resolve a dispute about meaning. A psychological experiment may explain why people overtrust a familiar voice without proving any particular viral recording false. A survey can demonstrate concern without quantifying how much deception has actually occurred. A detector's performance in a benchmark does not establish how it will behave on the next unknown generator.

Two serious positions remain. One says we should make authenticated capture, provenance, and visible verification routine, because ordinary audiences cannot investigate every file. The other warns that a world requiring machine-readable credentials for belief will let well-resourced institutions define reality while others struggle to be heard.

Both identify real risks. The strongest response may combine easier provenance with older, slower practices: multiple witnesses, original records, independent reporting, preserved corrections, and explicit statements about what has not been established. This is a proposal drawn from the evidence and trade-offs, not a trial-proven recipe for restoring trust.

The collision reveals the problem beneath the spectacle of deepfakes. Evidence is valuable because it allows an assertion to be checked beyond the authority of the speaker. If the checking procedure itself becomes unavailable, expensive, or distrusted, the record loses part of its social power even when it is true.

What We Know — and What We Don't

Evidence: Controlled studies show that participants can struggle to identify synthetic audiovisual material, although performance varies by task and modality. Experiments also show that claims of misinformation can sometimes protect accused political figures, with important differences between text and video. Technical standards can preserve and verify selected provenance assertions. The 2026 Reuters Institute survey measures high public concern about online falsehoods.

Causation: None of those findings proves that synthetic media caused a particular institution to lose trust, that the video in our invented library is fabricated, or that a signature would prevent a false allegation. The studies measure separate mechanisms in separate settings.

Interpretation: It is reasonable to infer that making plausible fabrications cheaper can increase the burden of verification and can offer cover for denying authentic evidence. The size and distribution of those effects, across countries, platforms, and communities, remain uncertain.

Analogy: A signed file resembles an evidence envelope with a traceable seal. The resemblance is useful for explaining integrity; it fails if taken to mean a signature verifies the truth of everything depicted.

Speculation: We may move toward a world in which important public claims are accompanied by richer provenance and chain-of-custody records. We may instead fragment into communities that trust different authorities. These are possible futures, not conclusions established by the cited studies.

What we do know is that verification is a process with multiple questions. A confident answer to only one of them must not quietly become certainty about all the others.

Back to the Reading Room

A member of the library committee offers the group-chat moderators the original file from her phone. It includes a longer conversation. The chair appears to say the disputed words while describing a rumor circulating before the vote. Another person supplies the written agenda, and the meeting secretary finds a contemporaneous note.

These are invented pieces of corroboration. They might clear the chair, expose a misleading edit, or reveal a genuine attempt to manipulate the vote. What matters is that the inquiry can compare records that were created for different purposes, by different people, before the allegation went viral.

The neighborhood may still disagree about whether the committee behaved well. That is a question of judgment. It need not remain confused with whether a sentence was said, when it was said, and what the full exchange meant.

If making a credible lie becomes easier, we should improve the ways a true claim can be tested. But there is no reason to pretend the test will always be quick, impartial, or conclusive. Reality does not owe us a convenient proof.

The hard question is whether our institutions will make room for careful uncertainty before our impatience turns a plausible clip into a verdict.

The Next Question

The last question in this sequence concerns a responsibility we cannot hand entirely to a camera, a detector, a platform, or a trusted authority.

When evidence is contested, someone must decide which records to preserve, which doubts are reasonable, how much corroboration is enough, and when to admit that the answer is still unknown. Those choices distribute the cost of uncertainty among real people.

So after asking whether reality can still be verified, one question remains for the reader—and for the institutions that shape what we see: who should bear the cost of proving what happened?