Imagine watching a video of a candidate saying something that could completely change your perception of an election. The image seems real, the voice is familiar, the movements are convincing, and within minutes, the content begins to circulate on social networks. The problem is that that person may never have said those words. With the evolution of generative artificial intelligence, producing images, videos, and audios capable of mimicking real people has become more accessible, turning deepfakes into an issue of digital security and, during an election period, into a potential threat to the very integrity of the democratic process. In the 2026 Elections, the Superior Electoral Court established specific rules for synthetic content and banned the use of deepfakes to benefit or harm candidacies. But there is a question as important as the regulation itself: when a forgery manages to look real to a person, how can technology find out that it is fake?

What is a deepfake and why has it become so difficult to identify?

Not all manipulated content is a deepfake. A photograph can be edited in Photoshop or a video can be cut and taken out of context, for example. Deepfakes differ because they use artificial intelligence to create or modify images, videos, and audios in an automated way. The technology can replace faces, reproduce voices, alter expressions and movements, or even create a person and a situation that never existed.

This capability has completely changed the challenge of identification. In the first generations of synthetic content, it was common to find noticeable signs, such as unnatural expressions, lighting issues, strange movements, or distortions in areas of the face. In images generated by AI, errors in hands, teeth, and eyes also became known as signs of forgery. These signs, however, are becoming less reliable as generative models evolve.

Today, artificial intelligence systems can produce much more convincing content, and this has created a kind of technological race between those who generate and those who detect the fakes. Recent research shows that detectors can perform significantly worse when they encounter content produced by models or techniques that were not present in their training data. This issue, known as generalization, is one of the main challenges of deepfake detection.

Therefore, identifying a deepfake does not simply mean looking for an error in the image or video. Analysis can involve statistical characteristics of pixels, compression patterns, facial movements, synchronization between voice and mouth, and other evidence that goes unnoticed by the human eye. NIST, for example, maintains specific evaluations to measure the ability of systems to detect manipulations in images and videos.

The more artificial intelligence becomes capable of mimicking reality, the less we can rely on human perception alone to decide what is true.

And that is precisely why the next stage of deepfake security is not just looking for imperfections, but finding out where that content came from and what happened to it before it reached us.

Before looking for the error, it is necessary to discover where the content came from

One of the strategies to combat deepfakes does not consist of looking for an error in the image or video, but in verifying its origin. This digital provenance allows recording information about how content was created, which tool was used, and what changes it underwent along the way.

One of the main standards for this is the C2PA, which stands for Coalition for Content Provenance and Authenticity. Through so-called Content Credentials, content can carry provenance information secured by cryptographic signatures, allowing verification of whether those records were subsequently altered.

In practice, it is like creating a verifiable history for an image, video, or audio. A photograph can record its creation, editing, and publication, for example. In an electoral scenario, this information can help differentiate content with a verifiable origin from a file whose provenance cannot be confirmed.

This does not mean that Content Credentials are a deepfake detector. The absence of this information does not prove that content is fake, just as its presence does not guarantee that everything presented in the content is true. Provenance is just another layer of evidence. In an increasingly synthetic internet, finding out where content came from can be just as important as finding out if it looks real.

Invisible watermarks can reveal content produced by AI

Another strategy to identify content generated by artificial intelligence uses digital watermarks that cannot be perceived by the human eye. SynthID, developed by Google DeepMind, for example, embeds this information directly into images, videos, and audios produced by AI systems.

The idea is simple: even if the content looks completely real, a specialized system can look for this signature and indicate that it was created or altered by artificial intelligence. The technology was also developed to remain detectable after common modifications such as cropping, filters, and compression.

But there is an important limitation. The absence of a watermark does not mean the content is real. Different tools use different methods, and not all AI-generated content has this type of identification. Therefore, the watermark acts as additional evidence and not as a universal deepfake detector.

In practice, the more tools adopt this type of identification, the easier it might be to discover when content has gone through artificial intelligence. But to verify what does not have a digital signature, it will still be necessary to resort to other forms of analysis.

And when there is no watermark? Digital forensics steps in

When content does not have a watermark or provenance credential, the analysis must look for other traces left behind during its creation or manipulation. This is where digital forensics comes in, which can examine pixel characteristics, compression patterns, textures, frame-to-frame movements, and even specific characteristics of an audio recording.

Artificial intelligence systems can also be trained to recognize these patterns. Instead of looking for an obvious error, they analyze statistical characteristics that may indicate that the image, voice, or video sequence went through a process of generation or manipulation. In videos, for example, the analysis can look for inconsistencies between different frames, while in audio, abnormal patterns in frequencies or in the way the voice was synthesized can be identified.

The problem is that these traces can disappear or become less obvious after the content undergoes compression, cropping, filters, or other common changes on social networks. Recent research shows that detectors experience a performance drop when they encounter different manipulations from those used during their training.

Therefore, digital forensics does not work as a simple tool that answers "true" or "false". It gathers different pieces of evidence to estimate whether a piece of content shows signs of manipulation, making the analysis more complex, but also more reliable.

AI can analyze an image, a voice, and a video at the same time

One of the most promising avenues for detecting deepfakes is combining different types of evidence. Instead of analyzing just the image or just the audio, multimodal systems can compare visual and sound information to look for inconsistencies.

Image and movement

In video, AI can analyze faces, expressions, movements, and visual patterns frame by frame. Small differences between regions of the image or changes that do not naturally follow the rest of the video can indicate manipulation.

Voice and synchronization

Audio can also be analyzed separately. Voice characteristics, frequency, rhythm, and other patterns can reveal signs of synthesis. In a video of a person speaking, the system can still compare lip movement with what is being said.

Crossing the evidence

It is precisely this combination that makes multimodal analysis interesting. A video may look convincing visually, but feature a synthetic voice or unusual synchronization between speech and movement. By crossing these signals, the system can build a more comprehensive analysis of the content.

This type of approach still faces limitations, mainly because AI generators also evolve rapidly. Even so, the trend is for detection to stop looking for a single "proof" of forgery and start combining different pieces of evidence to evaluate the authenticity of a piece of content.

The problem: no detector is a detector of truth

It is tempting to imagine that there is a tool capable of analyzing any video and simply answering: "it is fake." In practice, it doesn't work that way. Deepfake detectors work with probabilities and evidence, and their performance can vary depending on the type of content and the technology used to produce it.

A detector trained to recognize a certain type of manipulation may struggle with a newer technique. Research and evaluations from NIST show that accuracy can drop when systems encounter content different from what was used in their training, especially after the file undergoes compression, filters, or other changes.

Therefore, deepfake identification tends to work best when different pieces of evidence are combined: provenance, watermarks, forensic analysis, visual characteristics, audio, and context. Technology does not deliver a magical answer. It helps build a more reliable conclusion.

This boundary is especially important during an election. A detector may indicate that a certain piece of content shows signs of manipulation, but this does not replace verifying the information being presented.

Deepfakes in the 2026 Elections: what changes for candidates, platforms, and voters

In the 2026 Elections, the use of artificial intelligence is not banned in general, but it has started to follow specific rules. The TSE determines that synthetic content used in electoral propaganda must be explicitly and accessibly identified. However, the use of deepfakes to benefit or harm candidacies is prohibited.

For candidates and campaigns, this means that using AI requires more care with the identification, context, and purpose of the content. For platforms, the rules also establish measures to reduce the circulation of content that could compromise the integrity of the electoral process.

And the voter?

This is precisely where technology meets its greatest limit. No detector alone can guarantee that information is true. The best protection remains combining different signals: verifying the source of the content, looking for reliable sources, observing possible manipulations, and being suspicious of materials that trigger an immediate reaction before presenting any context.

The fight against deepfakes, therefore, does not depend on a single tool. It involves artificial intelligence, digital forensics, platforms, the Electoral Justice system, and, above all, the voter's ability to verify what they receive before sharing. In an election where anyone can produce a convincing fake, knowing how to be skeptical has also become a digital security skill.