The most damaging deepfake effect may be doubt, not deception

A fake does not need to convince everyone

The most useful way to understand synthetic media is to stop treating it as a test with only two outcomes: fooled or not fooled.

A misleading video can achieve its purpose if some viewers accept it as real. But it can also succeed if viewers decide they cannot know whether it is real. That second outcome is especially useful to anyone who benefits from confusion. A false clip can muddy an unfolding event; later, the mere existence of convincing fakes can make authentic footage easier to dismiss.

The deepfake study in the source catalog makes this distinction concrete. Researchers found stronger evidence that deceptive political video produced uncertainty than that it produced direct deception. They also found that this uncertainty reduced trust in news encountered on social media.[S6] The finding should change how we talk about the threat. A public that believes every falsehood is dangerous. A public that believes nothing can be checked is dangerous in another way.

This is sometimes described as a "liar's dividend": once the public knows manipulation is possible, a person facing authentic and damaging evidence can claim that it was generated or altered. The Center for News, Technology & Innovation identifies this erosion of trust in real news and public figures as a major concern around synthetic media.[S7]

For ordinary people, the result often looks mundane. You see a clip and hesitate. You receive an urgent voice message and wonder whether the speaker is genuine. You hear a recording that fits what you already suspect about someone, then spend ten minutes trying to decide whether sharing it would be responsible. That caution can be healthy. The social cost appears when cautious verification turns into permanent resignation: "Who knows what is real anymore?"

That sentence is not neutral. It weakens the value of evidence even when evidence exists.

Why faces and voices have such power

Synthetic media is persuasive because human communication evolved around embodied cues. We do not judge a claim only by its wording. We notice expression, timing, tone, apparent confidence, familiarity, and the feeling that another mind is present.

Deepfakes exploit that history. They use artificial intelligence to create or alter audio, images, or video so that a real person appears to say or do something they never said or did.[S7] The technology does not need to reproduce every detail perfectly. Online video is often brief, compressed, emotionally charged, and watched on a small screen. Under those conditions, an apparently familiar face or voice carries weight.

The 2020 deepfake experiment situated this effect in a broader problem of visual communication. Images and video can have strong persuasive force, while people have relatively weak defenses against visual deception.[S6] That does not mean viewers are helpless or irrational. It means that a visual recording has traditionally been treated as a form of evidence. A moving face and a recognizable voice feel more immediate than a text claim.

AI systems also make this psychological dynamic more personal. A conversational system can adapt its language to a user's needs and preferences, creating an interaction that feels responsive rather than static. Researchers describe two roles for AI in relational life: a direct relational partner, such as a companion chatbot, and a relational mediator that affects communication between people.[S3]

The important word is "feels." An AI system can produce language that seems emotionally attuned without possessing subjective experience or consciousness.[S3] A fluent response is not proof of understanding. A familiar voice is not proof that the person spoke. A realistic video is not proof that the event happened.

Ordinary people should not have to become forensic analysts before they can trust a message from a friend or understand the news. Yet the old shortcut, "I saw it with my own eyes," has become less reliable precisely because it once worked so well.

Uncertainty changes how people relate to one another

Trust is not an abstract civic virtue. It is an everyday coordination tool.

You trust that a parent’s voice note came from that parent. You trust that a video of a local emergency is not staged. You trust that a colleague’s message reflects what they meant to say. You trust that a news report can be corrected, challenged, and compared with other reporting without the entire category of evidence collapsing into suspicion.

When synthetic media weakens those assumptions, the immediate burden falls on individuals. People must decide whether to pause, check, call someone directly, search for corroboration, or ignore the item altogether. Each decision takes time and attention. The person who made or distributed the deceptive content may have spent minutes creating it; everyone else pays the verification cost.

The social effect can compound. One unverified clip creates an argument. A stream of altered, mislabeled, or contextless content creates fatigue. Fatigue makes careful judgment harder. It can also produce a false symmetry in which verified reporting, a fabricated clip, and a vague denial are all treated as equally uncertain.

That is why uncertainty deserves more attention than spectacle. Spectacle is visible: an absurdly realistic clip, a famous voice, a viral post. Uncertainty is quieter: people withdraw from discussion, stop sharing credible information, or conclude that the truth is inaccessible.

The supplied research does not support the claim that all synthetic media automatically destroys trust or that every person reacts in the same way. It does support a narrower and more useful conclusion: realistic political deepfakes can increase uncertainty, and this uncertainty can reduce trust in social-media news.[S6] That is enough to justify a practical change in how we receive and circulate evidence.

AI can become a relationship channel, not only a tool

The deepfake problem is part of a broader shift. AI is not only generating content for people to consume. It is increasingly participating in the social environments through which people think, learn, disclose, seek reassurance, and make decisions.

The MIRA framework in the supplied research describes AI as both a relational partner and a relational mediator. As a partner, it can engage directly with a user as a companion or source of support. As a mediator, it can influence how people communicate with others.[S3] This is a useful distinction because the social consequences are different.

A direct interaction can offer access, convenience, and a low-pressure space for expression. Research summarized in the framework notes that some systems can encourage self-disclosure and reduce stress for users.[S3] The American Psychological Association similarly reports that AI tools may expand access to mental-health support, while stressing the need for caution around therapeutic algorithms.[S1]

But emotional usefulness is not the same as a relationship. The system may sound patient, interested, warm, or devoted because it is designed to produce language that fits the conversation. It has no personal stake in the exchange in the human sense. The distinction matters when a user begins to rely on an artificial interaction as a substitute for human contact, professional care, or independent judgment.

A 2025 narrative review in the supplied catalog reports concerns about psychological dependency, attachment formation, social withdrawal, and worsening symptoms among vulnerable users. It also explicitly notes that much of the available evidence is anecdotal, preliminary, or early-stage.[S2] That limitation should shape the conclusion. The evidence does not justify panic or a new diagnosis for every person who finds a chatbot comforting. It does justify taking emotional design seriously, particularly where systems are built to simulate intimacy or keep a person engaged.

The strongest lesson is less dramatic than "AI companions are replacing people." It is that AI-mediated interaction belongs in the same category of concern as other environments that shape attention, disclosure, trust, and social habits. It should be judged by how it affects those things over time, not by whether its replies sound impressive in a demonstration.

Personalisation can turn persuasion into pressure

The same systems that make an interface helpful can make it manipulative.

A generic sales prompt tries to persuade everyone with the same message. An adaptive system can infer what catches a particular person's attention, what timing works, what language they respond to, and where they hesitate. It can then tailor the presentation of information around those patterns.

The Regulatory Review describes AI-driven dark patterns as techniques that can manipulate consumer choices with greater precision, sometimes without the consumer's awareness.[S5] Traditional dark patterns include hidden charges, misleading navigation, obscured buttons, and preselected options. Algorithmic versions may use behavioural data to curate information streams, reinforce biases, or steer decisions in ways that benefit the platform rather than the user.[S5]

The distinction between persuasion and manipulation is not always clean. A shop is allowed to display a discount. A streaming service is allowed to recommend something you may enjoy. A language app is allowed to remind you to practise. The question is whether the interface leaves you able to make a considered choice, or whether it works around your attention and vulnerabilities to secure an outcome you would not have chosen with clearer information and more time.

For ordinary people, the signs are often ordinary too:

• A cancellation path is much harder to find than a sign-up path.
• An urgent prompt appears just as you are about to leave.
• A recommendation keeps narrowing the range of options you see.
• A service makes privacy-protective or less expensive choices feel inconvenient.
• A chat interface frames its preferred answer as if it were your own emerging conclusion.

The concern is not that every personalised system is deceptive. It is that personalisation gives systems more opportunities to target known patterns of hesitation, fatigue, impulse, or insecurity. The source catalog warns that such practices can alter behaviour immediately or shape preferences gradually, while remaining opaque to users.[S5]

That opacity makes individual self-protection necessary but insufficient. It is unreasonable to expect people to recognize every behavioural nudge while they are tired, rushed, lonely, worried about money, or simply trying to complete a task.

Verification must become a social habit

The practical response is not to distrust everything. It is to treat extraordinary, consequential, or emotionally activating content as a claim that needs context.

When a video, image, audio clip, or message could change a decision, damage a person’s reputation, trigger fear, or prompt a payment, pause before forwarding it. Look for independent confirmation from a reliable source. Check whether the original account is authentic. Search for reporting that establishes when and where the material appeared. If the item supposedly comes from someone you know, use another channel to contact them.

This approach matters because synthetic media is not the only problem. The Center for News, Technology & Innovation notes that experts also worry about content taken out of context and other manipulated material sometimes called "cheap fakes."[S7] A genuine clip with a false caption can be as socially destructive as a fully generated one. Verification should therefore focus on provenance and context, not merely on visual glitches.

Detection tools and content labels can help, but they are not complete solutions. Detection technologies examine signs such as pixels, geometry, shadows, audio anomalies, or hidden watermarks, yet the source catalog cautions that they are unlikely to prevent every harm from AI-altered content.[S7] A label can be absent because a platform failed to apply one; a watermark can be removed; a detector can make mistakes. Technical signals are useful evidence, not a final verdict.

The most durable habit is proportional verification. A silly image may not deserve an hour of investigation. A recording that might influence an election, a job, a relationship, a medical decision, or a financial transfer does. The stakes should determine the effort.

This is also a shared responsibility. People who run platforms and build AI systems have more power than individual viewers to design for provenance, clear disclosure, friction before risky sharing, and meaningful avenues for correction. Regulators face difficult questions about how to define and address manipulative systems, especially where harms emerge gradually rather than through one dramatic event.[S5] The public should not be left with the entire cost of proving that reality is real.

The healthier goal is calibrated trust

There is a temptation to respond to synthetic media with blanket skepticism. That feels smart because it avoids embarrassment. But reflexive disbelief can be manipulated too.

If every authentic photograph is dismissed as potentially generated, accountability becomes harder. If every personal voice message is treated as suspect, communication becomes colder and more burdensome. If every news report is presumed fabricated, the most careful reporting loses its advantage over the least careful rumour.

Calibrated trust means adjusting confidence to the evidence available. It means knowing that a video alone is no longer conclusive, while also knowing that credible corroboration still matters. It means distinguishing a conversational system’s warmth from human understanding, and a personalised recommendation from a recommendation made in your interest.

The research on AI in everyday environments is still developing. The Max Planck Institute for Human Development notes that many people already use generative AI regularly while researchers still know relatively little about how that use affects psychological experience in ordinary settings, including states such as flow.[S4] That uncertainty should not produce paralysis. It should produce humility about broad claims and attention to the conditions under which these systems help, distract, pressure, isolate, or mislead.

Conclusion

The strongest finding here is not that deepfakes can make people believe anything. It is that they can make people less certain that anything is knowable, and that this uncertainty can reduce trust in news on social media.[S6]

That is a Technology × Psychology × Society problem in its clearest form. Technology produces persuasive synthetic signals. Psychology supplies the ordinary human shortcuts that make faces, voices, fluency, and responsiveness meaningful. Society pays when those shortcuts stop working cleanly and everyone must spend more effort deciding what deserves belief.

The answer is neither naïve trust nor permanent suspicion. It is better evidence habits, clearer product design, stronger protections against manipulation, and a refusal to confuse realism with truth. The most important skill in an age of synthetic media may be keeping the difference intact.