Deepfakes are synthetic or manipulated images, video, and audio produced with artificial intelligence. Modern generative models can reproduce facial appearance, facial movements, speech patterns, and voices. Their development has created a verification problem because human observers cannot consistently distinguish synthetic media from authentic material.
A 2024 systematic review and meta-analysis examined 56 studies involving 86,155 participants. Across the studies included in one pooled analysis, total deepfake detection accuracy was 55.54%. Detection accuracy was 53.16% for images and 57.31% for videos. The researchers concluded that overall human sensitivity to deepfakes was not significantly above chance.
Why Visual Evidence Is Psychologically Powerful
Human perception does not operate as a forensic authentication system. Visual and auditory information is processed through cognitive mechanisms that use previous experience, contextual information, expectations, and learned associations.
A preregistered 2021 experiment involving 210 participants examined people’s ability to identify deepfake videos. Participants could not reliably detect the manipulated videos and tended to classify deepfakes as authentic rather than classify authentic videos as deepfakes. Participants also overestimated their ability to identify manipulated material.
These findings identify two measurable problems:
- Exposure to audiovisual information does not guarantee accurate judgments about authenticity.
- Confidence in a judgment can exceed actual detection performance.
- Awareness that deepfakes exist does not automatically produce reliable detection skills.
- Financial incentives for correct identification did not significantly improve detection accuracy in the 2021 experiment.
The mismatch between confidence and detection performance matters because confidence can influence whether information is accepted, rejected, or shared.
Deepfakes Can Produce Uncertainty Without Producing Belief
A deepfake does not have to convince every viewer that its content is authentic to affect information processing. Synthetic media can also introduce uncertainty about whether genuine material is real.
Research describes one consequence as the “liar’s dividend.” The term refers to a situation in which the existence of convincing synthetic media makes it easier to claim that authentic evidence has been fabricated.
Five survey experiments involving more than 15,000 American adults tested this effect using reports of real political scandals. False claims that reports were misinformation increased support for politicians under several experimental conditions. The effect was stronger for text-based reports than for video evidence.
The psychological problem therefore operates in two directions: fabricated evidence can be accepted as authentic, while authentic evidence can be dismissed as fabricated.
Verification Shifts From Appearance to Provenance
When visual appearance cannot establish authenticity, verification requires information independent of the image or video itself. Relevant evidence can include the original publisher, publication history, metadata, independent reporting, and the identity or ownership of the source distributing the material.
Domain information can provide one part of that verification process. A whois lookup can be used to retrieve available registration information associated with a domain. Registration data does not prove whether a particular video is authentic, but domain age, registrar information, registration status, and available ownership details can contribute to an assessment of the website distributing the material.
This approach changes the verification question from “Does this look real?” to questions that can be investigated using external evidence.
Deepfakes Intersect With Phishing and Impersonation
Synthetic media also affects fraud because generative AI can reproduce characteristics associated with identifiable individuals. Voice cloning can imitate speech, while generated images or videos can support false identities.
Phishing attacks rely on impersonation, deceptive messages, fraudulent websites, or other methods intended to induce recipients to disclose information or perform an action. AI-generated content can increase the amount and personalization of material available for impersonation.
Relevant verification measures include:
- Confirming unusual requests through a separate communication channel.
- Inspecting the actual domain used by a website or email sender.
- Avoiding authentication through links supplied in unexpected messages.
- Treating urgent requests for credentials, money, or account access as requiring independent verification.
- Using multifactor authentication to reduce the consequences of stolen passwords.
Additional procedures for identifying and responding to AI-assisted scams are described in this guide to protecting yourself from phishing in the AI era.
Training Can Improve Deepfake Detection
Human detection performance is not completely fixed. The 2024 meta-analysis found that interventions including feedback training, AI assistance, and deepfake caricaturization improved detection performance. Across the intervention studies included in that analysis, pooled detection accuracy increased to 65.14%.
However, there is no universal visual artifact that identifies every deepfake. Manipulation techniques differ, and detection clues associated with one generation method may not occur in material created by another method.
Consequently, verification based exclusively on facial irregularities, unnatural blinking, lighting errors, or synchronization problems cannot establish authenticity in every case.
The Psychological Consequence Is Broader Than Deception
Research has documented effects extending beyond believing an individual fabricated video. A 2026 systematic review of political deepfake research reported effects involving altered perceptions, confusion, reduced trust in online news, and increased distrust in government. The same review found that deepfakes were not consistently more persuasive than misinformation presented through other formats.
A separate scoping review of empirical deepfake research identified reported harms including false memories, attitude changes, sharing intentions, false investment choices, anxiety, distress, reduced self-efficacy, and distrust in media.
The documented psychological effect of deepfakes is therefore not limited to successful deception. Synthetic media also changes the conditions under which authentic information is evaluated. When images, recordings, and videos can be generated or manipulated convincingly, judgments of authenticity increasingly depend on provenance, corroboration, source verification, and independent evidence rather than appearance alone.
