In early 2024, Hong Kong police reported a fraud case: A finance staff member at a multinational company’s Hong Kong office received an email from the "CFO" instructing him to wire a large sum overseas. His first reaction was correct—he suspected phishing. To dispel his doubts, he was invited into a video conference: On screen appeared the "CFO" and several "colleagues" he recognized, all looking and sounding authentic. He complied, transferring HK$200 million (approximately $25.6 million) in 15 installments across five accounts. Only days later, upon verifying with headquarters, did he realize he had been scammed. As a senior Hong Kong police superintendent told the media, "Everyone he saw in that multi-person video conference was fake."
The most bitter aspect of this case is not the amount involved, but this detail: The victim performed what we call "verification"—he did not blindly trust text; he demanded face-to-face confirmation. Yet his verification relied solely on his eyes and ears. And in 2024, eyes and ears have already failed.
## First Line of Defense: The Naked Eye—Empirically Ineffective
The naked eye cannot distinguish truth from falsehood—not as a lament, but as an empirical result. In 2022, PNAS published a series of perceptual experiments: Researchers used StyleGAN2 to generate 400 synthetic faces covering diverse genders, ages, and ethnicities; each was paired with a real face of similar appearance. Among 315 participants tasked with distinguishing real from fake, the average accuracy rate was 48.2%—close to coin-flip level.
A second experiment was even harsher: 219 new participants underwent training with immediate feedback per question, yet accuracy rose only to 59%, and did not improve further with practice. The authors’ explanation: Some synthetic faces contain no perceptible flaws whatsoever. A third experiment was most counterintuitive: 223 participants rated faces for "credibility" on a scale of 1 to 7; synthetic faces averaged 4.82, significantly higher than real faces’ 4.48; among the four most "credible" faces, the top three were all AI-generated. The paper’s conclusion: Synthetic faces have crossed the uncanny valley—"not only difficult to distinguish, but more credible than real faces."
Within the same paper lies an easily overlooked sentence—a verdict on the entire "look more closely" approach: In a digital world where any image or video can be forged, "the authenticity of any untimely or unwelcome recording can be questioned."
## Second Line of Defense: Detection Tools—A Losing Arms Race
Can algorithms do better? Also difficult. Detection models and generative models are locked in an inherent arms race; detectors trained in labs collapse in the real world when confronted with unseen generators, compression artifacts, screenshots, and forwarding—all without warning. This is not media hyperbole—the PNAS paper itself states: Existing detection technologies are neither efficient nor accurate enough to "handle the massive volume of uploads arriving daily." Tech media Ars Technica put it bluntly: Tools claiming to directly authenticate AI content have been proven "highly unreliable in real-world scenarios."
The industry’s response is a strategic pivot: from "guessing after the fact whether something is fake" to "declaring at birth." Google’s SynthID, launched in August 2023, embeds imperceptible watermarks into AI-generated images, text, audio, and video. Officially, over 10 billion pieces of content have already been watermarked. In May 2025, Google launched a verification portal allowing users to upload content and check for watermarks. But boundaries remain clear: It recognizes only content generated by Google’s own tools (and a few partners); Google itself admits watermarks can be evaded—especially in text; the portal opened first to early testers, with journalists, media professionals, and researchers able to apply via waitlist. In 2023, seven AI companies pledged to the White House to develop watermarking technology; yet as Ars Technica assessed in May 2025, a unified, mutually recognized standard remains absent.
## Third Line of Defense: Provenance—Being Deployed, Yet Wavering
A more systemic solution is provenance. The C2PA Alliance—whose steering committee includes Adobe, Amazon, BBC, Google, Meta, Microsoft, OpenAI, Sony, TikTok, and Truepic—maintains an open standard called Content Credentials: like attaching a "nutrition label" to content, recording who created it, when, with which tools, and what edits were made—anyone can verify this anytime.
Platforms are already acting. In February 2024, Meta announced it would read C2PA and other industry-standard metadata embedded in image files, automatically labeling AI-generated images from tools such as Google, OpenAI, Adobe, and Midjourney; for highly realistic AI videos and audio, uploaders must proactively declare their origin—or face penalties up to post removal.
But real-world deployment is far messier than official announcements. Meta, shifting toward "more labels, fewer removals," ceased deleting posts solely under its old manipulated-media policy starting July 2024; that same month, it erroneously labeled numerous authentic photos—lightly edited with AI—as "Made with AI," provoking strong backlash from photographers; by September, Meta retreated further—labels for "AI-edited only" content were moved into the post’s three-dot menu, while only fully AI-generated content retained prominent labeling. Meta’s official blog candidly admitted at the time: It cannot identify all AI-generated content, and invisible markers can also be stripped away.
This wavering in the provenance approach exposes a structural contradiction: Labels too strict cause false positives for ordinary editing; too lenient let true deepfakes slip through. Moreover, it is creating a new dividing line (a judgment made in this article): Content bearing provenance—produced by major institutions, big platforms, or new-model cameras—enjoys a presumption of authenticity; content lacking provenance—old photos, screenshots, or images taken by ordinary people—is presumed suspect by default. Authenticity is becoming a resource distributed unequally.
## The Liar’s Dividend
This term was coined in 2019 by legal scholars Chesney and Citron: Deepfakes do not merely cause "false things to be believed true"—they also enable "true things to be dismissed as false." When any recording can plausibly be claimed forged, those caught on camera now possess a ready-made alibi: lying gains dividends.
This is not a thought experiment. Political opponents of Gabon’s president once claimed a public video showing the president’s good health was a deepfake; that claim later served as partial justification for an attempted coup. Experts commented: "In a sense, whether that video was real or fake no longer matters—it has already been used to erode credibility and sow doubt." On September 28, 2023—two days before Slovakia’s parliamentary elections—a fabricated audio clip purporting to show opposition leader Šimečka conspiring with a journalist to manipulate the election went viral on social media; he ultimately lost the election. Some dubbed it the first election "influenced by deepfakes"; Harvard Kennedy School’s Misinformation Review cautioned that attributing the entire election outcome to one audio clip oversimplifies complex electoral dynamics—but equally undeniable is this: fabrication costs nearly zero, while correction cannot outpace dissemination. In January 2024, two days before the U.S. New Hampshire primary, thousands of voters received phone calls featuring a voice impersonating President Biden urging them not to vote; the orchestrator later admitted spending only $150 to hire someone to generate the AI voice, claiming his purpose was to alert the public to AI dangers. In February 2024, the U.S. Federal Communications Commission (FCC) issued a preliminary ruling: AI-generated voices fall under the federal Telephone Consumer Protection Act’s definition of "artificial or prerecorded voice," making their use in automated calls illegal; in September, the FCC levied a $6 million fine—the agency’s first penalty involving generative AI—and the orchestrator also faces 13 felony and 13 misdemeanor charges at the state level.
Punishment arrived—but only after the election. And the liar’s dividend carries an even more awkward property, as Citron wrote in congressional testimony: The more people learn about deepfakes, the more likely they are to distrust authentic recordings—anti-deepfake public education simultaneously funds liars’ defenses. As early as 2017, media documented politicians in over a dozen countries—including Russia, Brazil, Turkey, and Poland—adopting the "fake news" rhetoric; the American Political Science Review has published empirical research confirming the strategy of politicians labeling genuine information as forged to evade accountability.
## Three Judgments
First, authenticity is migrating from "perceptual judgment" to "provenance verification." "Looks real" has empirically failed—accuracy stands at 48.2%, equivalent to flipping a coin; "a detector says it’s real" is neither reliable nor universal; what remains is checking provenance: watermarks, signed records, source chains. For ordinary people, this means: discerning truth from falsehood is no longer an instantaneous act—it is becoming a task requiring tools, source checks, and cross-verification.
Second, provenance proves "origin," not "truth." Content Credentials can tell you who created a piece of media, when, with what tool, and whether it was edited—but it cannot tell you whether the content itself is truthful. Real footage paired with false context, or staged "documentary" scenes, can carry flawless provenance. Industry builds infrastructure for verifying "origin authenticity," but the public will treat it as infrastructure for verifying "content authenticity"—and this gap will be exploited. The final mile of verification still resides in multi-source corroboration and institutional fact-checking—not in any single label.
Third, deepfakes’ greatest harm may not lie in how many people false content fools, but in how all true evidence is devalued. Recordings were once ironclad evidence for accountability—in courtrooms and in public discourse; now they have been downgraded to "evidence pending verification." This devaluation amounts to a tax on the entire information ecosystem—and the tax revenue flows straight into the pockets of those most adept at evading accountability.
## Four Takeaways
First, stop trusting intuition. "Looks real" does not mean real; "looks fake" does not mean fake—you scored at random on this test, and training recovers only ten percentage points.
Second, replace intuition with process: Check platform AI labels and provenance markers; reverse-image search to find earliest origins; seek a second independent source; for critical content, return to the original posting—do not stop at forwarded screenshots.
Third, beware the "liar’s dividend" rhetoric. When someone—especially the subject—labels damaging imagery as "deepfake" without evidence, treat that denial itself as information requiring verification. Denials, too, require evidence.
Fourth, create your own provenance: Preserve original files and capture metadata for important moments; when publishing content, include provenance signatures where feasible. In the near future, "being able to prove you’re real" may matter just as much as "being real."
Seeing is no longer believing. But don’t let "seeing is no longer believing" become a new excuse for laziness—the replacement for eyes should not be universal suspicion, but more patient verification. Have you ever had a moment where you couldn’t tell real from fake at a glance? Share in the comments.