The ability to tell what is real from what is fake on the internet has never been more fragile. A recent study examining how well people can detect deepfakes finds that most individuals have a hard time distinguishing AI-generated video from authentic footage. And as artificial intelligence tools continue to improve, that gap is likely to widen.

The findings arrive at a moment when synthetic media has moved from a curiosity to an everyday threat. Anyone with a modestly powered computer and an afternoon of free time can now produce a convincing fake video of a politician saying something they never said, a celebrity endorsing a product they never touched, or a CEO announcing a fictional bankruptcy. The technology has advanced faster than the public's ability to recognize its artifacts.

The study, conducted by an unnamed research group, tested participants on a set of videos that included both real clips and AI-generated fakes. The results were not encouraging. A majority of participants misidentified the deepfakes, often expressing high confidence in incorrect judgments. Even when viewers were told in advance that they would be shown deepfakes, performance barely improved. The implication is stark: the default human ability to spot these videos is unreliable, and consciousness-raising alone will not solve the problem.

Part of the difficulty is that modern deepfakes have become exceptionally good. Earlier generations of AI-generated video suffered from obvious tells: unnatural blinking, warped facial boundaries, inconsistent lighting, or lip movements that did not quite match the audio. Those cues are largely gone. The latest models, trained on enormous datasets of human faces and voices, produce footage that is often indistinguishable from reality when viewed under normal conditions. A slow-motion frame-by-frame analysis might reveal anomalies, but that is not how most people watch video online. They scroll, they react, they share.

This creates a fertile environment for misinformation. Deepfakes have already been used to smear political opponents, impersonate corporate executives in fraud schemes, and produce non-consensual explicit content. As the technology becomes more accessible and harder to detect, the potential for harm grows. The study underscores a painful reality: the burden of verification currently falls on the viewer, and most viewers are not equipped for the job.

There are, of course, counterarguments. Some researchers argue that deepfake detection algorithms can catch what the human eye misses. Automated systems that analyze facial micro-expressions, evaluate consistency of lighting gradients, or check metadata can flag suspicious content with high accuracy. The problem is that these tools are not widely deployed where they are needed most: on social media platforms where the content spreads fastest. Even when they are, bad actors can tweak their generation methods to evade detection. An arms race between creators and detectors is unfolding, and the creators currently have the advantage of iteration speed.

Another counterpoint holds that deepfakes are not actually more dangerous than older forms of disinformation like cheap fakes or out-of-context clips. A video taken out of its original setting can be just as misleading as a synthetic one. That is true, but it misses the scale and plausibility that deepfakes introduce. A cheap fake requires real footage that can be misrepresented. A deepfake invents the footage entirely, allowing lies that have no anchor in reality. The combination of realism and limitless inventiveness is what makes the new generation of fakes uniquely corrosive.

Where does this leave the average person? The study suggests that simply telling people to be more skeptical is not enough. Skepticism without tools leads to paralysis or cynicism: either people stop believing anything they see online, or they dismiss the risk altogether. Neither outcome is healthy for public discourse.

What is needed is a layered approach. First, social media platforms must invest aggressively in automated detection and in labeling known deepfakes. The technology exists, and its deployment should be transparent and consistent. Second, media literacy education should include practical modules on how deepfakes are made and how to verify a video source. Third, legal frameworks that impose consequences on malicious deepfake creators must be updated. Several countries have begun moving in this direction, but enforcement remains patchy.

The study makes one thing clear: the problem is not going to solve itself. AI video generation is getting cheaper, faster, and more realistic with each passing quarter. The gap between the ease of creation and the ease of detection is growing, not shrinking. If the public cannot reliably tell what is real, then the concept of shared reality itself starts to erode. The findings of this study should act as a warning, not a verdict. The tools to fight back exist. The question is whether we will use them in time.