Real or AI? Alonso Hurtado is cloned in front of ASCOM experts to demonstrate that they can no longer distinguish him

For a few seconds, no one in the auditorium of the Madrid Press Association doubted that the Alonso Hurtado appearing on the screen was the real one. He was thanking the organizers, addressing the auditors and present investigators, and announcing that he would speak about how artificial intelligence is changing the rules of the game. But the real Alonso, at that very moment, was seated at the table. "It's a fake," were his first words.

With this dramatic opening, he began his presentation, «AI, deepfakes and digital evidence as new challenges for independent experts», which concluded the ASCOM Current Issues Conference on Regulatory Compliance Experts Register. The session was presented and moderated by Elena Ballesteros Tejado, who admitted that for a few seconds everyone had believed it was true.

Ten minutes and a one-minute recording

Alonso explained how he did it. The night before, he had recorded himself for just over a minute, wearing a t-shirt and shorts, in front of his laptop. Using a very affordable commercial tool with a monthly subscription, he had created his double in about ten minutes.

He did not hide the flaws: hand gestures were stiff, the voice could have been better, and the neck moved unnaturally due to overly quick rendering. With more time, the result would have been “much more perfect.” But the lesson was a different one: “Anyone can do this. It’s not that difficult.”

The technology, he emphasized, is not bad in itself. At ECIJA, we use it with clients to create corporate training videos with avatars that can be translated into another language with a single click.

How much does it cost to deceive?

The jump to sophisticated forgery is, according to Alonso, a matter of money: the investment depends on what is at stake and, with sufficient resources, the result can be practically undetectable. He recalled the case of Arup, the engineering multinational that lost about 25 million dollars in Hong Kong due to a video call in which their financial director was impersonated.

He also described a plausible scenario: an attacker who has already accessed the systems of a large company creates a fictitious user account and assigns it an avatar driven by generative AI, capable of conversing naturally during a routine meeting. It is in these trusted environments where we let our guard down the most.

He was cautious when discussing the possible repercussions of the subject in court. He acknowledged that, so far, he knows of no real case where this type of evidence has been manipulated in an expert report, even though he does not rule out that one will appear soon. And the risk goes beyond video: it also affects audio recordings, signatures, documents, and photographs. A still image, he noted, is 'much, much easier' to forge.

Machines against machines

To explain how to detect this type of manipulation, Alonso referred to Alan Turing and the deciphering of the Enigma code: to beat a machine, you need machines. Applied to our day, AI can only be detected with another AI specifically trained for that purpose, not with a generic tool.

He also warned of the risk of false positives, citing a personal anecdote. He passed a professional document he had written through an AI-generated text detector, and the tool identified it as artificial. “The reason was that the document was very well organized, numbered, and full of connectors - common features of legal language,” he explained.

For this reason, in his opinion, the independent expert “must be aware that these tools exist, must know how to use them, and, above all, must know how to explain them to decision-makers.”

Systematic skepticism

The moderator raised the central question: while until recently an image or an email inspired trust, nowadays an audio recording or a video is no longer considered conclusive evidence. What should an expert do when a company only provides a video?

The answer was unequivocal: question everything. Alonso explained that this summer he was almost the victim of a fraud attempt during a moment of complacency, and he was frustrated by verifying the information through an independent channel. The second step is to familiarize oneself with the tools available on the market, which now have prices he considers subsidized. As an example, he cited the case of a financial entity supervised by the Bank of Spain, where a large-scale AI analysis matched one hundred percent with the transactions that the same entity had already identified.

“The AI said it” is not enough

An expert must defend his report in court, facing another expert. Alonso pointed out that European regulations impose obligations regarding explainability and transparency, and that the systems allow for a step-by-step breakdown of the reasoning and retain a record of the queries that have been made. The correct approach would be: “I used artificial intelligence, but this is the logical process I followed to reach this conclusion.” The professional must understand it and make it his own.

He also recommended running each report through the AI and asking it to act as the opposing party's lawyer, questioning the report itself. “Never settle for the first thing an artificial intelligence system tells you,” he warned. These systems tend to be “vague and complacent,” which is why key conclusions should always be contrasted with original evidence. The professional will not be replaced, but he becomes a bottleneck compared to the speed of the machine.

What can be input to AI

Regarding the information that can be input into these systems, Alonso highlighted what needs protection: trade secrets, personal data, and intellectual and industrial property. He advised distinguishing between local processing and cloud-based processing and warned that, in the United States, authorities may request conversations stored on servers. The key, he explained, lies in non-retention: “If they don't retain data, they cannot train the model.”

Beyond expert testimony

The warning extends to social networks, where false videos and audios can trigger trends with serious consequences, ranging from protests to destabilizing electoral processes. Alonso explained that Instagram automatically tagged a vacation photo he had edited as “edited with AI.” Platforms, he emphasized, do not do this out of their own will, but because European regulations require them to proactively identify manipulated or AI-generated content.

He concluded with a reflection that summed up the day: “AI is what email was 25 years ago.” It has come to stay, and the old adage “seeing is believing” no longer applies.

You can read the full report at Confilegal.