Deepfakes are no longer science fiction. Avatars in video calls and voices cloned from a single public speech sample can create personalised fraud produced at scale. Yet the solution to this increasing threat may lie less in detecting what is fake than in redesigning trust itself, say Sopra Steria cyber defence experts François Grime and Stéphane Grousseau.
This morning, Charles looks in good shape. Your colleague is present for the video meeting he scheduled via email yesterday. He asks you to confirm the transfer needed to pay the invoice for a recent assignment. Everything seems normal, so you do it. The problem is that the person on screen does not exist. It is an avatar generated and animated in real time by AI.
“We’ve entered an era where we interact with a digital avatar that takes on the appearance of someone we know… and it isn’t them,” says Stéphane Grousseau, Deputy Director of the Cyber Defence and Intelligence Agency at Sopra Steria. “This is no longer science fiction: it’s technology that is becoming accessible.”
In 2025, Sopra Steria measured the economic impact of disinformation, particularly in the spread of deepfakes. The benchmark case was a fraud carried out in 2024 using a recording.
Eighteen months later, the question we put to two of our specialists in cyber defence and information warfare is no longer whether the technology works. The question is how quickly it is becoming industrialised – and which defences actually hold up.
A regime change or an acceleration?
For François Grime, a cyber-defence consultant to Sopra Steria and contributor to the Cercle Pégase think tank dedicated to information warfare, the shift is structural. “For me, this is a regime change. Models keep improving, and this will only continue to expand,” he says.
He points less to the technology than to its raw material, noting that a business leader who speaks at a single roundtable leaves hours of usable voice and video online. Thanks to today’s professional culture, which encourages leaders to embody their companies on social media and in podcasts, this has turned everyone’s digital footprint into training data.
Grousseau is cautious about the language of revolution. “For me, it’s an acceleration, not a regime change,” he says. According to him, the most powerful factor is not that audio samples are getting shorter, but the collapse in production time, cost, and reaction speed, not to mention the arrival of live deepfakes.
The practical consequence remains the same. Fraud moves from craftsmanship to industry and what once required a skilled operator, and time can now be produced at scale – and personalised at the same time.
A multiplier that is hard to detect
Our 2025 study estimated that deepfakes could increase online fraud by 15–20%. Today, both Grousseau and Grime consider that an outdated range. “Calling it 15 to 20% is more than conservative,” says Grousseau. “We’re closer to a multiplication factor of two or three.”
Grime attributes this leap to orchestration, AI agents and automated workflows that multiply attacks while tailoring each one individually. The result is not only more fraud, but more believable fraud – moving beyond the classic CEO fraud scenario into everyday life. Examples include the panicked phone call of “Mum, it’s me, I need a bank transfer”; non-consensual intimate deepfakes; and “sextortion” targeting ordinary people who have no legal department to defend them.
The instinctive response is to detect fakes. Both experts support that effort but warn that it is not enough. “It’s a real cat-and-mouse game,” says Grime. He says that detection is not simply identifying generated content; it also requires an entire pipeline of collection, pre-processing, and analysis. And there is an uncomfortable asymmetry because collecting and qualifying data costs much more than generating it.
The most promising path, he suggests, is monitoring behaviour rather than content. “When we rely more on behaviour and less on content, we reduce the volume of analysis,” says Grime. The idea is to track inauthentic patterns across a network rather than judging each message individually.
Rebuilding an architecture of trust
If content can no longer be taken at face value, the response is to rethink how trust is granted. Grousseau refers to a concept well known in cybersecurity: zero trust. “Zero trust is a concept, not a technology,” he says. “It is based on trusting no one at any moment. So, there is a cultural and human dimension, not just a technical one.”
Applied to communication, this means a return to multi-factor verification – even in very human forms. This could be a password agreed upon in advance with relatives or executives, or a question only the real person could answer.
Grime’s field experience means that he also believes in the power of agreeing on internal recognition phrases. “It costs nothing; it’s discipline – just like cybersecurity. When you don’t have technical means, you put procedures in place,” he says.
That requires operational changes. Grime encourages organisations to act proactively, monitoring their own information environment, mapping vulnerabilities as they would in cybersecurity, and preparing playbooks, legal contacts, and press distribution lists in advance, as well as appointing someone to coordinate communications, security, cyber, and legal teams that rarely work together. “Responsiveness is the enemy of security,” he says.
A collective response
The most striking reversal, however, is generational. Phishing taught us to protect older users who were less comfortable with digital tools, but deepfake fraud, argues Grousseau, will hit young people first. These are social media natives who expose much of their lives online, providing attackers with rich and accessible human intelligence, and who often extend trust from the very first interaction.
“Cross-checking simply isn’t part of their culture,” he says. He expects collective resilience to grow over time, just as we learned to distrust images online. But he warns about an intermediate “period of uncertainty” – the period in which we will be most exposed.
The honest conclusion is that no firewall solves this problem. Awareness comes first, verification habits second, technology third. “Imagining that the technological solution will be a real firewall” means framing the problem incorrectly, warns Grousseau.
Instead, the deepfake era will be navigated by deliberately rebuilding trust – an exercise that is as human as it is technical. Organisations that begin now – mapping their exposure, training their teams, and treating speed as a risk – will likely be the least affected when that transitional period ends.