Deepfakes, artificial images and the disturbing psychological consequences of a world in which even genuine evidence can be dismissed as fake.
There was a time, not very long ago, when a photograph still possessed a peculiar authority over us. We knew that photographs could be staged, edited or removed from context, and anyone familiar with the history of propaganda knew that visual manipulation was hardly a modern invention, yet an image still seemed to carry something that words did not. It appeared to say that, whatever interpretation might later be imposed upon it, something had existed in front of a camera at a particular moment in time. Video strengthened that feeling even further because movement, voice, hesitation, facial expression and the small imperfections of ordinary human behavior seemed much harder to fabricate than a written account.
Generative artificial intelligence is quietly dismantling that old psychological arrangement.
It is now possible to manufacture faces belonging to people who never existed, clone voices from relatively small samples, reconstruct scenes that never occurred and produce increasingly convincing video in which real or invented people appear to move through events that happened only inside a computational system. None of this means that every artificial image is indistinguishable from reality, and the popular claim that nobody can tell the difference anymore would be an exaggeration. What has changed is subtler and, in some respects, more troubling: the mere knowledge that convincing fabrication is possible has entered our everyday understanding of the world, which means that suspicion no longer requires evidence of manipulation before it begins.
A strange reversal is taking place. For most of the history of misinformation, the principal fear was that fabricated evidence might be mistaken for reality, yet researchers studying deepfakes have increasingly drawn attention to another possibility, in which genuine material loses some of its authority simply because people know that falsification has become technically plausible. The danger, therefore, is not confined to believing things that never happened. We may also be approaching an information environment in which events that really did happen can be pushed into a fog of uncertainty simply by suggesting that the evidence might have been manufactured.
That is where the story becomes considerably darker, because a society does not need to lose reality in order to become confused. It only needs to lose agreement about how reality can be established.
The New Problem Is Not Simply That Images Can Lie
Human beings have always lied with images. Long before digital editing, photographs were cropped, retouched, staged and selectively published, while governments, advertisers, newspapers and private individuals learned very quickly that a camera could be used to persuade as effectively as it could be used to record. What generative AI changes is not the existence of visual deception but its accessibility, its speed and, increasingly, its realism.
An experiment published by the Harvard Kennedy School’s Misinformation Review in November 2025 examined how realistic AI-generated images influenced belief in false headlines. Researchers found that participants were more likely to believe false headlines when accompanying synthetic images appeared realistic and seemed to provide strong visual evidence for what the headline claimed. The study did not suggest that people indiscriminately accepted every artificial image, which would be a much stronger claim than the evidence supports, but it did demonstrate something psychologically important: when an image looks plausible and appears to confirm a narrative, its visual authority can increase the credibility of information that is false.
This is hardly surprising when we consider how people encounter information online. We do not normally sit in front of each photograph like forensic investigators, enlarging reflections, examining compression patterns and tracing the history of every file before allowing ourselves to form an impression. We scroll through hundreds of fragments of reality every day, making rapid judgments about faces, accidents, wars, celebrities, disasters, scientific discoveries, crimes and apparently ordinary moments in the lives of strangers. Context arrives in the form of captions written by people we may not know, while the emotional impression created by an image often reaches us before careful reasoning has had time to intervene.
The artificial image therefore does not need to be perfect. It needs to survive long enough to enter the mind as a possibility.
Correction may come ten minutes later, or ten hours later, but by then the original image may already have been copied, cropped, reposted, stripped of its context and absorbed into countless conversations. Even where a correction works, the larger lesson remains behind: photographs can no longer be approached with the same instinctive confidence they once enjoyed.
At first, that sounds healthy. Skepticism is useful, particularly in an online world saturated with manipulation, sensationalism and commercial incentives designed to capture attention. Yet skepticism is beneficial only while it helps us distinguish reliable information from unreliable information; once it expands until every form of evidence appears equally suspect, it stops functioning as a filter and begins dissolving the distinction it was meant to protect.
That transformation has a name.
The Liar’s Dividend: When the Existence of Fake Evidence Protects the Truthful-Looking Lie
In their influential work on deepfakes, legal scholars Robert Chesney and Danielle Citron described a phenomenon they called the liar’s dividend, a concept that deserves far more attention than it usually receives outside specialist discussions of misinformation.
Their argument was unsettling because it reversed the familiar problem of the fake.
Suppose a genuine recording emerges showing a person saying or doing something damaging. In an earlier technological environment, that person might have attempted to explain the recording, challenge its context or claim that the interpretation was misleading. In an age familiar with deepfakes, another possibility becomes available: deny that the recording is genuine at all.
The person making that denial does not necessarily have to prove that the recording was fabricated. Public awareness that deepfakes exist may itself introduce enough uncertainty to weaken the force of authentic evidence. Chesney and Citron described this as a dividend enjoyed by liars because improvements in our awareness of digital manipulation can paradoxically provide additional protection to people confronted with genuine material.
This is a very different informational problem from ordinary deception. A traditional lie attempts to replace one version of reality with another, whereas the liar’s dividend can operate simply by contaminating certainty. Instead of persuading everyone that an alternative explanation is true, it may be sufficient to persuade enough people that nobody can know which explanation is true.
The distinction matters enormously because uncertainty is much easier to manufacture than an entirely convincing alternative reality.
Imagine that an authentic video appears online. Within minutes, comments begin claiming that the mouth movement looks wrong, that the voice sounds artificial, that a shadow is inconsistent, that the camera motion resembles generative video or that an unnamed expert has supposedly identified synthetic artifacts. Other users insist that the video is genuine, while still others upload altered copies as jokes, demonstrations or deliberate attempts to create confusion. Someone produces screenshots purporting to prove authenticity; another person claims those screenshots were also fabricated. The original evidence has not disappeared and may remain entirely genuine, yet it now exists inside an ecosystem of competing claims that an ordinary viewer may have neither the time nor the expertise to resolve.
At some point the psychological question changes from Is this real? to How could I possibly know whether this is real?
That difference is not trivial.
A Deepfake Can Fail as a Lie and Still Succeed
One of the most interesting findings in deepfake research appeared before the present explosion of generative AI.
In 2020, researchers Cristian Vaccari and Andrew Chadwick studied how a large representative sample of people in the United Kingdom responded to synthetic political video. Their results did not show a population helplessly believing every deepfake placed in front of them. Instead, participants exposed to deceptive deepfakes were more likely to become uncertain than simply to accept the false material as true, and that uncertainty was associated with reduced trust in news encountered through social media.
This finding changes the way the problem should be understood. A deepfake does not necessarily need to convince somebody of its literal content in order to have an effect; if it makes that person less certain about the reliability of subsequent information, then the deception has produced a secondary consequence even after failing in its primary purpose.
There is something strangely efficient about this form of manipulation. Persuading millions of people to believe the same elaborate falsehood is difficult because people possess different values, experiences, political loyalties, cultural assumptions and levels of knowledge, whereas creating doubt requires much less agreement. One person may suspect the media, another may distrust governments, another may distrust corporations, another may distrust artificial intelligence itself, yet all can arrive at the same psychological destination from entirely different directions: I don’t know what to believe anymore.
A systematic review published in Frontiers in Political Science in June 2026 examined 74 empirical studies conducted between 2018 and 2025 and concluded that synthetic media frequently create what the authors describe as epistemic uncertainty, rather than producing universal persuasion. The review uses the expression skepticism tax to describe the additional mental burden imposed when people must constantly evaluate whether apparently authentic information might have been artificially created or manipulated.
The idea of a skepticism tax deserves to be taken literally as a psychological cost. Every questionable photograph requires attention, every suspicious video invites verification, every source must be compared against another source, and every correction can itself become the object of another round of suspicion. For professional investigators that work is part of the job, but ordinary life was never designed to require forensic authentication before we decide whether a video of an event deserves to be believed.
Eventually fatigue becomes part of the information environment.
When verifying everything becomes impossible, people do not necessarily become more careful. They may simply choose whom they trust.
When Evidence Becomes a Matter of Identity
This is where the problem begins to intersect with conspiracy culture in a way that is more complicated than simply saying that people who believe conspiracies are easily fooled.
Many conspiracy narratives begin with suspicion toward official explanations, and sometimes skepticism toward authority is justified; institutions have lied, corporations have concealed damaging information, intelligence agencies have conducted secret programs, governments have withheld documents, newspapers have made serious errors, and historical archives contain enough genuine deception to make unquestioning faith in authority intellectually indefensible. The difficulty begins when healthy skepticism loses any standard by which it might eventually be satisfied.
If every official document can be dismissed as fabricated, every photograph as artificial, every witness as compromised, every recording as synthetic and every contradictory piece of evidence as part of the deception, then a theory has become almost impossible to disprove. Evidence against it no longer weakens the theory because the evidence itself can be absorbed into the theory.
Generative AI adds a remarkable new instrument to this older psychological tendency because it provides a technically plausible explanation for almost any inconvenient piece of digital evidence.
This does not mean that accusations of manipulation are necessarily false. Some images really are generated, some videos really are altered and some recordings really are fraudulent, which is precisely what gives the problem its strength. The most effective form of uncertainty is rarely produced by an impossible claim; it grows from something that genuinely could have happened.
We therefore enter a peculiar informational landscape in which the existence of real fakes can help false accusations of fakery survive.
The distinction is subtle enough to be missed while scrolling through a feed, but its implications are enormous. If people become accustomed to treating authenticity as merely one interpretation among many, evidence gradually stops functioning as a common language. Two individuals can watch the same recording and inhabit entirely different versions of the event, not because they disagree about what the recording means but because one believes that the recording itself belongs to reality while the other believes it belongs to simulation.
At that point disagreement moves beneath interpretation and into ontology itself.
We are no longer arguing about what happened. We are arguing about whether the thing in front of us ever existed.
The Matrix Problem Has Changed
For decades, popular culture has returned obsessively to the same unsettling possibility: what if the world we experience is not the world that actually exists?
The Matrix became perhaps the most famous modern expression of that anxiety because its central horror was not simply that human beings were being deceived; they inhabited an apparently coherent reality so complete that the deception became indistinguishable from ordinary life.
The emerging problem of synthetic media turns that fear inside out.
The frightening question may no longer be whether artificial intelligence can construct a false reality convincing enough for us to believe. It may be whether repeated exposure to artificial realities can make us incapable of recognizing the genuine one when it appears.
That reversal is more psychologically interesting than the familiar science-fiction scenario because it requires no secret machine controlling civilization and no hidden architect designing an illusion around us. All that is required is an environment containing enough convincing fabrications that every authentic image arrives already carrying the shadow of possible falsification.
A generation growing up inside such an environment may develop a relationship with visual evidence quite different from that of previous generations. The photograph may cease to mean this happened and begin to mean only someone is presenting this as something that happened. Video may become less a record than a claim, while voice recordings, screenshots and digital documents acquire the same provisional status.
There are reasonable technological responses to this problem. The Coalition for Content Provenance and Authenticity, known as C2PA, has developed standards designed to attach cryptographically verifiable provenance information to digital media, allowing systems to preserve information about the source of an asset and changes made to it. The idea behind Content Credentials is not to create an infallible machine for declaring truth but to make the history of digital content more transparent and tamper-evident.
This is important, although provenance alone cannot solve the deeper psychological problem. A technically authenticated photograph can establish information about where a file came from and how it was modified, but technology cannot force someone to trust the organization providing the authentication. Every system of verification ultimately reaches a human question about which institutions, standards and authorities deserve confidence.
That may become the real battlefield of the synthetic age.
The Most Dangerous Fake May Be the One Nobody Believes
Much of the public conversation about generative AI still concentrates on improving detection. Researchers search for patterns invisible to ordinary viewers, platforms develop labeling systems, digital provenance standards are being refined, and users are repeatedly instructed to look for anomalies in hands, text, reflections, lighting, lip synchronization and movement.
Those tools are necessary, but they address only one side of the problem.
Detection assumes that our central task is separating real objects from fabricated ones, yet the liar’s dividend suggests that the larger problem may eventually involve restoring confidence in authentic material after a culture of suspicion has already taken hold. An AI detector that occasionally makes mistakes can itself generate new uncertainty, while competing detectors may produce conflicting results, and ordinary users may have little idea which system deserves greater trust.
The situation becomes almost paradoxical: as tools for manufacturing reality improve, tools for verifying reality become increasingly necessary, yet the very need for constant verification teaches us that unverified perception can no longer be trusted.
This is why a badly made deepfake can still contribute to the problem.It does not need to fool you personally. It only needs to remind you that better ones exist.
After enough reminders, every surprising video contains a small hesitation before belief. Every extraordinary photograph invites suspicion before wonder. Every recording of something consequential arrives alongside the possibility that somebody made it yesterday on a laptop.
That hesitation is rational, but a civilization cannot operate entirely through hesitation.
Courts, journalism, history, science and ordinary interpersonal trust all depend upon methods for deciding that, after reasonable investigation, some things are sufficiently established to be treated as true. Absolute certainty has never been available, but practical certainty is indispensable. We cross bridges because engineers certify them, take medicine because evidence supports it, reconstruct crimes from testimony and physical traces, and understand historical events because different forms of evidence converge strongly enough to make endless doubt unreasonable.
A culture that loses the ability to reach that point does not become more intelligent merely because it distrusts everything.
It becomes easier to manipulate.
The Pageant of the Unreal
There is an old assumption buried deep inside many theories of propaganda: whoever controls the story controls the audience. Artificial intelligence may be revealing a more sophisticated possibility, because control does not always require constructing one story powerful enough to dominate all the others. Sometimes the easier strategy is to produce so many competing versions that the audience eventually gives up attempting to establish which one corresponds to reality.
The result would not necessarily resemble an authoritarian world in which everyone is forced to believe the same official narrative. It could look almost like the opposite: millions of people surrounded by unlimited information, each apparently free to choose what they believe, while shared standards of evidence quietly disintegrate beneath them.
That world would be loud rather than silent, crowded rather than censored and filled with images rather than deprived of them. There would be more recordings, more photographs, more documents, more commentary and more apparent witnesses than any previous civilization possessed, yet the abundance itself could make certainty harder to achieve because every authentic fragment would coexist with countless imitations.
This is why the coming crisis of artificial media may ultimately have less to do with whether computers can perfectly imitate reality than with whether humans can preserve a workable relationship with truth after imitation becomes ordinary.
The greatest success of a fabricated world would not be making everyone believe the same lie. Such uniform deception is probably impossible, and human beings have always been too argumentative for that.
A much more realistic possibility is that enough contradictory images, synthetic voices, manipulated documents, confident accusations and competing explanations accumulate until truth begins to feel like something that belongs to personal allegiance rather than evidence.
When that threshold is crossed, a photograph will no longer settle an argument, a recording will no longer end a denial, and the existence of proof may become only the beginning of another dispute about whether the proof itself is real.
The uncomfortable question raised by deepfakes is therefore not simply whether artificial intelligence will learn to deceive us.It is whether we are constructing an information environment in which deception no longer needs to succeed because certainty itself has become difficult to defend.
Perhaps that is the real pageant of the unreal: not a world in which everything is fake, but one in which reality continues to exist exactly as before while our confidence in recognizing it slowly disappears.And once a society reaches that point, the oldest question in philosophy acquires an unexpectedly modern form.
We will still ask what is true, but before we can answer it we may increasingly have to answer something even more fundamental: what kind of evidence are we still willing to believe?
