The Image That Cannot Be Trusted

For most of human history, a video of a person doing something was reasonable evidence that the person did it. Not perfect evidence — video could be edited, staged, selectively framed. But the baseline assumption was that what you saw had some relationship to what happened. That assumption is now under serious pressure. Deepfakes — synthetic media generated by AI systems that can plausibly replicate a person's face, voice, or both — have moved from expensive research projects to tools available to anyone with a consumer laptop and a few hours. The quality has improved faster than most researchers anticipated. Videos that would have taken a Hollywood visual effects team weeks to produce can now be generated in hours by software that does not require any expertise to operate. The consequences are not theoretical. Political figures have been depicted saying things they never said. Executives have had their voices cloned in audio calls used to authorize fraudulent wire transfers. Ordinary people — mostly women — have had their faces inserted into explicit content without their consent. The scale of this last category is particularly troubling: a 2023 report estimated that over ninety percent of deepfakes online were nonconsensual sexual images, with the vast majority targeting private individuals rather than public figures.

What Makes Deepfakes Difficult

The challenge is not that deepfakes are always convincing on careful examination. Many can still be identified by artifacts — inconsistent blinking, unnatural skin texture, audio that doesn't quite sync with mouth movement. The challenge is that most content is not viewed on careful examination. It is encountered in a scroll, shared before verification, consumed in a moment of emotional engagement when critical analysis is at a low ebb. The cognitive work required to identify a deepfake is real and effortful. The speed at which misinformation travels is not. This asymmetry is the core problem, and no amount of media literacy training fully closes it.

The Liar's Dividend

One consequence that received less early attention than it deserved is what researchers call the liar's dividend: the ability to dismiss authentic evidence as fabricated. If deepfakes are known to exist, then any inconvenient video can be claimed to be one — and enough doubt can be introduced to neutralize it. A politician caught on camera making a damaging statement can now credibly seed uncertainty about its authenticity, even if the video is real. This creates an environment where the mere existence of deepfake technology erodes the evidentiary value of all video, authentic or not. The tangent: this dynamic predates AI. Long before deepfakes existed, powerful people claimed that unflattering audio recordings were doctored, that photographs were manipulated, that witnesses were lying. Deepfakes did not invent the technique of challenging evidence. They industrialized it and made it more credible.

Institutions Working on This

Researchers at the MIT Media Lab have been developing detection tools that analyze deepfake videos at the pixel level, looking for statistical signatures that remain even after sophisticated generation. Their detection systems have shown high accuracy in controlled settings, though the accuracy degrades as generation technology improves — an ongoing race between creation and detection that has no obvious endpoint. The Stanford Internet Observatory has studied the spread of deepfake content in political contexts, tracking how synthetic media travels through social platforms and what interventions — labeling, reduced amplification, removal — have measurable effects on belief updating. Their findings suggest that labeling synthetic media is less effective than commonly assumed, but that timing matters: labels applied before initial viewing reduce credibility significantly more than labels applied after a person has already seen the content once.

Living With the Uncertainty

There is no fully satisfying resolution here. Detection tools help but are not infallible. Platform policies help but are not consistently enforced and cannot keep pace with creation volume. Legal frameworks are developing slowly and vary enormously across jurisdictions. What individuals can do is limited but not nothing. Slowing down before sharing emotionally charged video content, checking whether multiple credible sources have verified a claim, and developing a general prior that surprising viral content should be treated with extra skepticism — these habits are not foolproof but they are not worthless either. The harder adaptation may be psychological: learning to hold uncertainty about evidence without becoming paralyzed by it, to neither accept everything nor reject everything, and to stay functional in a media environment where the baseline rules have genuinely changed.