Nowadays, most topics related to artificial intelligence are highly relevant and widely debated. One of the most controversial and feared is undoubtedly that of deepfakes, tools that immediately and tangibly demonstrate the power of new technologies, as well as how easily we can be deceived by them.

First of all, the term “deepfake” is a combination of the words “deep”, referring to deep learning techniques, and “fake”, meaning something that is not genuine. The term refers to the generation or modification of images, videos, or audio recordings through the use of deep learning techniques (E. Altuncu, V. Franqueira and S. Li, “Deepfake: definitions, performance metrics and standards, datasets, and a meta-review”, Frontiers in Big Data, 2024).

Deep learning is an artificial intelligence technique that uses neural networks composed of a huge number of layers dedicated to data processing, enabling a system to learn and reproduce complex patterns derived from computationally expensive training (see also H as in Hidden Layers, https://mirai.systems/h-come-hidden-layers/).

When we refer to the concept of deep learning, we are considering a vast range of techniques that can give rise to systems designed for very different purposes. Indeed, deep learning forms the basis of algorithms used for facial recognition and the recognition of complex images, the analysis and production of natural-language texts and, as already mentioned, the synthesis of deepfakes (I. Sarker, “Deep Learning: A Comprehensive Overview on Techniques, Taxonomy, Applications and Research Directions”, SN Computer Science, vol. 2, no. 6, 2021).

Deepfake technology offers a wide range of possibilities, including the automatic translation of videos or audio recordings into foreign languages and the digital recreation of a person. With regard to the latter function, it is possible to generate images of a deceased person to simulate their presence, as happened at the Dalí Museum in St. Petersburg, Florida. Through the “Dalí Lives” project, the museum’s management presented the public with a synthetic reproduction of the painter Salvador Dalí, capable of introducing himself to visitors and interacting with them (T. Flattery, C. Miller, “Deepfakes and Dishonesty”, Philosophy & Technology, vol. 37, 2024).

Other well-known visual deepfake techniques include digital rejuvenation and aging, which are widely used in the film industry and make it possible to modify a person’s face, making them appear younger or older. Digital de-agingtechnology attracted considerable public attention following its use in the film “The Irishman”, directed by Martin Scorsese. In the film, actors Robert De Niro, Al Pacino, and Joe Pesci, all over seventy-five years old at the time of filming, were digitally rejuvenated by several decades (K. Loock, “On the realist aesthetics of digital de-aging in contemporary Hollywood cinema”, Orbis Litterarium, vol. 76, 2021).

In general, we can say that deep learning systems designed to create deepfakes, by learning a subject’s characteristics, are then able to reproduce them in a highly realistic manner.

In the case of visual deepfakes, these systems can use technologies to modify faces, such as face swap, which exchanges them, and face blend, which merges them, or simply generate synthetic images. Audio deepfaketechnologies, on the other hand, are based primarily on voice cloning or speech synthesis, which makes it possible to convert written text into an audio recording (A. Mohammed, “Deep Fake Detection and Mitigation: Securing Against AI-Generated Manipulation”, Journal of Computational Innovation, vol. 4, no. 1, 2024).

Useful Applications of Deepfake Technology

The potentially deceptive nature of deepfakes tends to cause considerable concern. Consequently, there is a risk of forgetting that these tools offer an extremely wide variety of positive uses.

First of all, deepfakes have proven extremely useful in the entertainment industry, making it possible to significantly improve films, television series, video games, and other media. Using this technology can make the production of these works less expensive by simplifying processes that are usually complex, such as creating extremely realistic characters, refining an actor’s accent or, as mentioned earlier, modifying their appearance, thereby improving the authenticity of the performance.

Education is another sector that has benefited greatly from these tools, and the examples to consider involve a wide range of activities. A system capable of generating synthetic audio can simulate a conversation with a user in a foreign language, while an algorithm designed to produce realistic videos can, for example, create an immersive scenario set in another historical period to make learning more engaging. The long-term benefits of these tools in the field of learning have yet to be demonstrated, but the innovations they offer are certainly attracting considerable interest (E. Lundberg, P. Mozelius, “The potential effects of deepfakes on news media and entertainment”, AI & Society, vol. 40, 2024).

Another field in which the use of deepfakes can bring significant benefits is medicine and medical research, which make use of both audio and visual technologies. For example, the creation of synthetic audio can once again enable a person who has lost the ability to speak to communicate using their own voice, by reproducing their original voice through a text-to-speech system. This system converts written text into audio output characterized by the user’s vocal timbre, rhythm, and other nuances (S. Bratieres, G. Donvito, M. Filosto, E. Frontoni, M. Kostovic, F. Minazzi, R. Pugliese, S. Regondi, “Artificial intelligence empowered voice generation for amyotrophic lateral sclerosis patients”, Scientific Reports, vol. 15, 2025).

Visual deepfakes, however, are currently the ones providing the greatest benefits in the medical sector, both by directly helping patients and by improving the tools currently available to medical personnel. One example of the use of synthetic content for therapeutic purposes is the generation of familiar faces for a person suffering from Alzheimer’s disease. These faces are created to look as though they come from a period that the patient remembers more clearly, with the aim of reducing their sense of isolation and improving communication with family and friends.

Deepfakes can also prove extremely useful in medical fields beyond therapy, for example, in producing images indistinguishable from authentic photographs, which can be used to train AI systems employed by medical personnel to identify specific diseases (S. O’Brien, Positive use cases of “deepfakes”, https://5g.wilsoncenter.org/article/positive-use-cases-deepfakes, 2025).

Harmful Uses of Deepfakes

Unfortunately, deepfakes also present various risks because, like any technological tool, their malicious use can cause considerable harm to individuals or society as a whole.

The most notoriously harmful example concerns pornographic deepfakes: artificial reconstructions in which a person, in almost all cases without having given their consent, is depicted completely naked. In this case, the extreme realism of the images plays a key role, as they can easily be perceived as authentic by third parties, thereby damaging the reputation of the person depicted.

Another well-known example of the malicious use of these technologies concerns fraud in which a scammer manages to deceive a victim by faithfully reproducing the voice of someone they know, inducing them to trust the instructions communicated and disclose confidential information.

Furthermore, reproducing an individual’s face and voice can be exploited to bypass digital security checks, thereby allowing the privacy of the person concerned to be violated or, in more serious cases, private data or money to be stolen (A. Bauerle, A. Diel, T. Lalgi, F. S. Mellis, M. Teufel, “The harm of deepfakes: A scoping review of deepfakes’ negative effects on human mind and behavior”, AI & Society, 2025).

As already mentioned, deepfakes can harm not only individuals but also society as a whole when shared interests are involved.

The clearest example can be found in the spread of fake news, a problem that affects society as a whole. False news was already being effectively disseminated in the past without access to deepfakes. However, there is no doubt that when information is supported by credible-looking, albeit fake, visual evidence, it is considered much more reliable and spreads more rapidly.

The convincing nature of deepfakes, particularly in the age of social networks, where any news item can easily reach millions of users, can have potentially disastrous consequences for society. Audio recordings or videos synthetically produced during an election period to undermine the credibility of a political candidate can manipulate and falsify election results. More generally, these artificial contents can damage the credibility of institutions (R. Chesney, D. Citron, “Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security”, California Law Review, 2019).

Technologies for Generating and Recognizing Deepfakes

The first technologies for generating deepfakes were rather crude, and it was therefore not difficult to identify anomalies that made the artificial nature of the content evident. Today, however, creators of synthetic media can rely on much more robust and powerful technologies, among which the range of techniques that exploit Generative Adversarial Networks (GANs) certainly stands out (H. L. Gururaj, B. C. Soundarya, “Deepfake detection: critical review of state-of-the-art approaches and future perspectives”, Discover Applied Sciences, 2026).

A GAN is a neural network composed of two sub-networks that compete with each other in order to create increasingly realistic data. The first component is the “generator”, which, starting from an input, aims to generate synthetic content that is as realistic as possible. The second is called the “discriminator” and, after examining both the input consisting of real data and the generator’s output, must determine whether the content it examines is authentic or fake.

The power of this technology lies in its so-called “adversarial” approach: the generator attempts to produce content that the discriminator might mistakenly consider real, while the latter tries to consistently distinguish authentic outputs from synthetic ones. This competition leads the system to generate extremely realistic deepfakes (M. Ding, B. Liu, B. Liu, T. Zhu, “A Review of Deepfake and Its Detection: From Generative Adversarial Networks to Diffusion Models”, International Journal of Intelligent Systems, 2025).

In general, deep learning technologies that enable the generation of deepfakes are constantly evolving and are making it possible to create synthetic material that is increasingly difficult to distinguish from authentic content. Clearly, not everyone has the same intuition when it comes to recognizing synthetic content, and, similarly, some systems produce more realistic deepfakes than others. In general, however, artificially generated visual content is easier to recognize than audio recordings because it contains inconsistencies and anomalies that can be identified more easily (M. Holmes, D. Miller, K. Somoray, Human Performance in Deepfake Detection: A Systematic Review, https://ssrn.com/abstract=4955104, 2024).

Nowadays, relying on human beings to recognize deepfakes is not a reliable option, and this task isdelegated to dedicated artificial intelligence systems. These detection techniques, however, have a major drawback: although they can deliver potentially excellent results, beyond the reach of any human being, they become obsolete very quickly. Indeed, generative models are advancing at an incredibly rapid pace, and both those developing generative techniques and those developing systems to identify deepfakes are trying to improve their tools in order to gain an advantage over the other side. This creates a genuine competition known as an “arms race” (D. Camacho, K. A. P. Costa, J. Del Ser, D. Jodas, J. P. Papa, L. A. Passos, D. Rodrigues, L. A. Souza Junior, “A Review of Deep Learning-based Approaches for Deepfake Content Detection”, Expert Systems, 2024).

This competition is evolving at an extremely rapid pace. Consequently, it is advisable to find solutions that address the problem at its root by seeking to implement measures that discourage the creation of harmful deepfakes, rather than focusing on techniques for identifying them (F. Abbas, A. Taeihagh, “Unmasking deepfakes: A systematic review of deepfake detection and generation techniques using artificial intelligence”, Expert Systems with Applications, 2024).

The solution that currently appears to be the most reliable is watermarking, that is, the insertion of an imperceptible watermark into AI-generated content. Ideally, this would be made mandatory by law for every generative system, so that any inauthentic content would be identified as such. The watermark in question must be imperceptible but, at the same time, easily detectable. This is certainly the main limitation of this technique today, since watermarks can be easily compromised through minor alterations to the generated content. Simply resizing an image, for example, can compromise or eliminate the information that makes it possible to detect the watermark.

It should be specified that watermarking is not intended to make the generation of a deepfake more difficult. Rather, by highlighting its artificial nature, it would make malicious use of the tool considerably more difficult (J. Li, L. Li, H. Luo, “Digital Watermarking Technology for AI-Generated Images: A Survey”, Mathematics, 2025).

Deepfakes in Relation to Traditional Methods of Media Manipulation

The fact that deepfakes frighten us because of their potential should not make us forget that the ability to modify audio or photographic content has always been within our reach, ever since we became capable of producing it. Indeed, if it is possible to modify a photograph using a deep learning system, the same can be done using software such as Photoshop or, to give an example that now seems outdated, by manually retouching a photographic negative (H. Farid, “Creating, Using, Misusing and Detecting Deep Fakes”, Journal of Online Trust and Safety, 2022).

The use of modified media for political purposes or to influence public opinion is usually associated with the contemporary era, but this practice was already widespread many decades before the advent of deepfakes. One example can be found during the period of Iosif Stalin’s dictatorial government in the Soviet Union. Some photographs dating back to 1920 were tampered with. They depicted Lenin giving a conference in Moscow alongside the revolutionary Leon Trotsky, who, as a person disliked by the Stalinist regime, was removed from the photograph.

The image alteration, of course, was not an end in itself, but was intended to reinforce the regime’s historical narrative by removing an individual who had fallen out of political favor (M. Rossner, K. M. Yamada, “Digital Images Are Data: And Should Be Treated as Such”, The Journal of Cell Biology, 2014).

The manipulation of photographs, videos, and audio recordings is certainly nothing new. At the same time, however, it is clear that deepfake technology frightens people much more than any other means serving the same purpose. Instinctively, it is easy to think that the fear people feel toward deepfakes is linked exclusively to their hyperrealistic nature, which, as already stated, makes this type of content increasingly difficult to distinguish from authentic material (M. Dosaikina, A. M. Proverbio, “Neural signatures of hyper-realistic AI-generated faces: dissociating behavioral indistinguishability from implicit neural evaluation”, Scientific Reports, 2026). However, the truth is that the issue is much more complex than that.

Indeed, the more sophisticated traditional forms of media manipulation, such as the aforementioned Photoshop software, require study, practice, and even talent. Generally, modifying content so that it appears genuinely authentic requires meticulous and lengthy work, accessible only to experts. By contrast, generating a deepfake requires neither time nor skill and is within anyone’s reach, including users who are not even aware of how the generative algorithm works (M. Pawelec, “Deepfakes and Democracy (Theory): How Synthetic Audio-Visual Media for Disinformation and Hate Speech Threaten Core Democratic Functions”, Digital Society, 2022).

The fact that nowadays the ability to create hyperrealistic images or audio recordings has such a low barrier to entry inevitably leads to an extremely large quantity of deepfakes circulating, many of them produced with malicious intentions. As mentioned earlier, this phenomenon not only facilitates the spread of false news but also risks causing people to distrust authentic information, leading them to think that the media supporting it have been artificially generated (D. Fallis, “The Epistemic Threat of Deepfakes”, Philosophy & Technology, 2020).

Conclusions

In conclusion, we have seen that deepfake technology can have various positive uses, but people usually focus their attention on the risks associated with it, thereby overshadowing the potential benefits of this technology.

Artificial intelligence techniques, including those that make it possible to generate deepfakes, have potential that we currently struggle even to quantify, and consequently we feel intimidated by them. What we must not forget, however, is that throughout history, all technological innovations have initially frightened people, who have nevertheless always gradually managed to become familiar with them and use them appropriately.

For this reason, when examining the current situation concerning deepfakes and other AI technologies, it is extremely important to understand the potential dangers associated with these techniques. At the same time, however, it is equally essential not to panic and to learn to become familiar with what will inevitably be the technologies of the future. Studying this field will allow us to ensure the responsible development of these systems which, together with appropriate regulation, will enable us to use them and benefit from them in an informed manner.

Giorgio Fumagalli

(translated by MIRAI)