Hive, C2PA, and Intel: How Deepfake Verification Services Work and Where They Fail

Services for recognizing deepfakes are rapidly turning into a separate market, but none of them yet provide an iron-clad guarantee. A test of four popular solutions showed: the best results come not from universal promises, but from tools with clear specialization and confirmed file origin.

How the test was conducted

To verify, we gathered a set of 100 files in three categories. It included regular photographs and videos from cameras, modern synthetic videos and images, as well as hybrid content — for example, frames after neural network retouching, upscaling and other processing. Audio forgeries were also tested separately to understand how systems behave not just on images, but on voice as well.

This set made it possible to compare services not on sterile demos, but on typical use cases that newsrooms and ordinary users encounter. An important detail of the methodology — files were submitted in the form they actually circulate on the internet. This means: without some EXIF data, after compression by social networks and messengers, sometimes in a re-saved form.

In practice, this is exactly where many services start making mistakes. In laboratory conditions, a detector can look convincing, but after Telegram, WhatsApp or Instagram its confidence noticeably drops. This is why the test results are closer to real-world use than to marketing presentations.

Who showed results

Comparison quickly showed that there is no universal winner. Each tool works well in one scenario and noticeably underperforms in another, so the result depends not on brand loudness, but on what exactly you're checking: photos, conversational video, audio, or a file with a preserved chain of origin. Because of this, comparing them by a single overall accuracy percentage is simply meaningless. For one class of tasks, a leader easily becomes an outsider in the next one.

• Hive confidently recognizes many generated images, but can mistake artifacts of heavy compression in old video for traces of AI.
• RealityGuard from Sensity handles video better than others, where face, facial expressions and lip sync with voice matter, but sharply loses accuracy on landscapes and images without people.
• Content Credentials based on the C2PA standard doesn't so much search for forgery as confirm file origin if the camera and software preserved the chain of signatures.
• Intel FakeCatcher shows the best results on quality close-up video, analyzing physiological signals of the face, but is almost useless on static images and low-resolution videos.

Against this background, the C2PA standard stands out especially. If a file was originally shot on a compatible device and did not lose signatures during editing or transmission, this is the strongest argument for authenticity. The problem is that in real life, such files are still a minority: old cameras, messengers and simple re-uploading easily break this chain. For news organizations and photographers, this is already turning from an option into a working standard, not an experiment.

Where detectors break

The first reason is the race between generators and detectors. As soon as verification systems learn to look for one set of artifacts, new models remove precisely these weak points. Previously, forgeries were given away by strange fingers, unnatural eyes or jerky lip sync. Now these markers are less common, and detectors have to rely on more subtle signs that are easily destroyed by compression. Essentially, verification systems are almost always catching up, not ahead of generators.

The second problem is the content distribution environment itself. Photos and videos almost never reach the user in their original form. Social networks cut quality, messengers re-compress files, and platforms remove some metadata. Because of this, even powerful tools begin to confuse real content with synthetic or, conversely, miss a quality deepfake. Checks work especially poorly where there is no close-up face, clear audio, or original file at all.

The third problem is the person on the other side of the screen. Even if a service shows the probability of forgery, many perceive the result as a final verdict. But a score of 60 or 70 percent is not a court expert opinion, but merely a signal that the content needs to be checked more deeply. The reverse error causes no less harm, when a user completely ignores a warning because the interface is unclear or the service has made a mistake once before.

«In 2026, the most reliable detector is still a combination of

technologies, independent investigation and common sense.»

What this means

For newsrooms, fact-checkers and ordinary users, the conclusion is simple: there is no single button yet for exposing deepfakes. The working scheme today is to combine automatic verification, look at metadata and file origin, take into account the quality of the source and, if possible, enable Content Credentials on your own devices. The more important the decision, the more dangerous it is to rely on one service verdict. Especially if it involves money, reputation or publication of controversial material.