By Javier Surasky
A Spanish version
(ES) will be published next Thursday
.png)
The scene
captures the problem: images, voices, and digital presence can now be manufactured with enough realism to deceive individuals and institutions. A
deepfake, a word that combines deep learning and fake, is “a
fictitious video, audio, or image file that realistically imitates someone’s
appearance, voice, or gestures using artificial intelligence” (INCIBE, 2024).
This forces us to revisit a basic biological assumption: that what we see and
hear, when we are not intentionally immersing ourselves in the world of fantasy
(movie, theater, TV) is real.
Three fronts of a real threat
The
expansion of generative AI has accelerated the spread of deepfakes by putting
increasingly accessible and lower-cost tools for creating them within reach of
more and more people.
The risk
posed by deepfakes is especially evident in fraud and cybersecurity, through
practices such as identity scams, vishing, voice-based deception, and executive
impersonation. According to the Voice Intelligence and Security Report 2025,
“deepfake fraud attempts increased by more than 1,300% in 2024 compared with
2023” (Pindrop, 2025, p. 3). These attacks rely on a logic that combines
technical realism with psychological pressure and the creation of a sense of
“urgency.”
No less
important, however, is the risk deepfakes pose by producing political
disinformation. In January 2024, in New Hampshire, an automated call using Joe
Biden’s voice circulated with the aim of discouraging people from voting,
leading to a six-million-dollar fine for the consultant responsible (Reuters,
2024a). In Slovakia, two days before the 2023 elections, “a fake 40-second
audio clip portraying a member of the Progressive Slovakia party allegedly
discussing vote buying circulated during the election silence period, a moment
when debunking the content is almost impossible” (Wired, 2023).
There is
also a high-risk area involving personal and reputational harm, where a strong
gender bias is also evident, as the Stimson Center points out:
Image-based abuse represents one of the most alarming manifestations of
AI-driven harm. Advances in generative AI have dramatically lowered the
barriers to producing hyper-realistic deepfake images and videos, most notably
non-consensual pornographic content. Recent studies estimate that 98% of all
deepfake content online is non-consensual and pornographic, and that 99% of
those depicted are women (Mingeirou, Osman, & Rafin, 2026).
It should
also be added that producing harmful deepfakes does not require great technical
expertise, because a fake can cause serious harm even when it has obvious
technical flaws. It is enough for it to circulate in the victim’s environment,
or in a socially sensitive situation, to trigger persistent doubts that are
difficult to dispel.
The crisis of authenticity
The deepest
threat posed by deepfakes lies in the erosion of trust in what is true: a
person accused of real conduct can now claim that the material was manipulated,
even when it was not, adopting a dodge-and-weave strategy, a phenomenon known
as the “liar’s dividend” (Schiff, Schiff, & Bueno, 2025, p. 71). Nina
Schick captures this when she writes that, in our time, “anyone can be
targeted, and everyone can deny everything” (Schick, 2020, p. 8).
How to protect ourselves: from reaction to protocols
For a long
time, internet verification guides advised people to pay attention to physical
details: out-of-sync lip movements, odd blinking, uneven lighting, or unnatural
intonation. But technological advances have made that advice obsolete, making
detection with the naked eye practically impossible in the case of the most
sophisticated deepfakes. Today, the only real defense, though not a complete
one, lies in contextual verification.
What do we
mean by “contextual verification”? Something as basic as this: the question of
truthfulness should no longer be whether the video, image, or audio looks real,
but where it comes from.
Deepfake
detection cannot rest on a single capacity, whether human or technical.
Pehlivanoglu et al. (2026) found that machine-learning algorithms classify
static images more accurately, while humans perform better with dynamic videos.
They also showed that analytical thinking and digital skills are associated
with better human detection of deepfake videos.
The result?
It is essential to combine contextual verification with digital literacy and
the use of technical tools, rather than relying on individual perception or
automated systems alone.
More
specifically, for protection against deepfakes to be more effective, action is
needed at three levels:
1. The
individual and procedural level: the rule here is to slow down before acting,
confirm the source, and look for alternative sources before reacting to a
stimulus designed to provoke urgency or outrage.
2. The
corporate-industrial level: this includes implementing protocols under which a
voice or video call is not enough to authorize critical operations, creating
double-confirmation systems, independent channels, and tiered controls.
One example
of good practice is the technology industry’s progress in applying the C2PA
cryptographic standard, which seeks to certify the provenance and authenticity
of files from the moment they are captured by a camera. This standard makes it
possible to include a verifiable digital credential in photos, videos, audio
files, and documents, providing information about their origin and
modifications. It does not tell us whether something is true or false, but it
does help us know where it came from, whether it was edited, and whether the
information accompanying the file has been altered at any point.
We want to
be very clear on this point: although applying the C2PA standard is useful and
helps identify edited content, it functions more as a transparency tool than as
proof that the content is real.
3. The state
level: countries are responding to deepfakes by creating regulatory frameworks
that follow different approaches. The European Union’s Artificial Intelligence
Act requires synthetic content to be labeled, with platforms facing sanctions
for noncompliance, while South Korea has chosen a strict punitive path,
imposing prison sentences for the creation and distribution of explicit or
malicious deepfakes.
Apart from
those two cases, there are no general AI laws in any other country in the
world, although there are sectoral or specific laws, as in the case of Brazil,
which banned the use of AI to spread disinformation in election campaigns
(Reuters, 2024b).
Seeing is believing: an obsolete rule
Generative
AI is making it increasingly difficult to know what is real.
For decades, images and voices served as strong evidence of reality, but “seeing and hearing are no longer believing” (Schick, 2020, p. 26), but today, digital trust must be built through traceability and corroboration.
Trust requires cooperation among the actors involved, basic levels of digital literacy among individuals, and an investment of time and effort, because verification is no longer an exceptional practice or one reserved for certain professions; it has become a basic comprehension skill.
In the age
of deepfakes, distrust is no longer paranoia but a last refuge for defending
reality: technology, which once called on us to expand our capacity for
imagination, now forces us to doubt our own eyes and ears.
References
INCIBE. (2024). Deepfakes. Instituto Nacional de Ciberseguridad. https://www.incibe.es/aprendeciberseguridad/deepfakes
Mingeirou, K., Osman, Y., & Rafin, R. (2026). The impact of artificial
intelligence on violence against women and girls. Stimson Center. https://www.stimson.org/2026/the-impact-of-artificial-intelligence-on-violence-against-women-and-girls
Pehlivanoglu,
D., Zhu, M., Zhen, J., Gagnon-Roberge, A. A., Kern, R. K., Woodard, D., Cahill,
B. S., & Ebner, N. C. (2026). Is this real? Susceptibility to deepfakes in
machines and humans. Cognitive Research: Principles and Implications, 11,
Article 3. https://doi.org/10.1186/s41235-025-00700-y
Pindrop.
(2025). 2025 Voice Intelligence and Security Report. Pindrop Security. https://www.pindrop.com/resources/reports/2025-voice-intelligence-security-report
Reuters.
(2024a, September 26). Consultant fined $6 million for using AI to fake
Biden's voice in robocalls. https://www.reuters.com/world/us/fcc-finalizes-6-million-fine-over-ai-generated-biden-robocalls-2024-09-26/
Reuters.
(2024b, February 29). Brazil Justice Moraes warns political candidates not
to use AI against opponents. https://www.reuters.com/world/americas/brazil-justice-moraes-warns-political-candidates-not-use-ai-against-opponents-2024-02-29/
Schick, N.
(2020). Deepfakes: The coming infocalypse. Twelve.
Schiff, K.,
Schiff, D., & Bueno, N. (2025). The liar’s dividend: Can politicians claim
misinformation to evade accountability? American Political Science Review,
119(1), 71–90. https://www.cambridge.org/core/services/aop-cambridge-core/content/view/687FEE54DBD7ED0C96D72B26606AA073/S0003055423001454a.pdf/the-liars-dividend-can-politicians-claim-misinformation-to-evade-accountability.pdf
The
Guardian. (2024, February 5). Company worker in Hong Kong pays out £20m in
deepfake video call scam. https://www.theguardian.com/world/2024/feb/05/hong-kong-company-deepfake-video-conference-call-scam
Wired. (2023, October 3). Slovakia's election deepfakes show AI is a danger to democracy. https://www.wired.com/story/slovakias-election-deepfakes-show-ai-is-a-danger-to-democracy/