
Full analysis
Dr Nicola Harding, Financial Crime Lab
On 8 October 2026, Certos, the fraud prevention business of Early Warning Services, published a survey of 230 senior finance leaders at Fortune 500 companies, venture-backed firms and large non-profits. Eighty-one per cent said their company had experienced an attempted fraud involving AI-generated or AI-enhanced content in the past year. Eighty-four per cent said AI-enabled fraud was harder to detect than traditional fraud, yet only half were very or extremely confident they could identify an attempt before money left the organisation (Certos, 2026). Earlier in the year, the Association of Certified Fraud Examiners (ACFE) and SAS surveyed 713 anti-fraud professionals. Seventy-seven per cent reported an increase in deepfake social engineering over the past two years, and only 7 per cent said their organisation was more than moderately prepared to detect or prevent AI-powered fraud (ACFE, 2026).
Both surveys measure perceptions rather than losses, and both come from organisations with an interest in the answer. Even so, they capture a real anxiety. Deepfakes have become the defining image of contemporary fraud. I want to argue that the image is misleading. The deepfake is a prop. The crime is persuasion, and persuasion is something criminology has understood for a long time.
What actually happened at Arup
The case most often cited in discussions of deepfake fraud is the attack on the engineering firm Arup in early 2024. A finance worker in its Hong Kong office received a message, apparently from the company's UK-based chief financial officer, requesting a secret transaction. The worker initially suspected a phishing attempt. Those doubts were set aside after a video call with people who looked and sounded like the chief financial officer and other colleagues he recognised. All of them were deepfakes. He went on to make 15 transfers totalling HK$200 million, about US$25.6 million (CNN, 2024).
Arup's Chief Information Officer, Rob Greig, has been careful about how he describes it. He called the incident "technology-enhanced social engineering" and noted that none of the firm's systems were compromised. His most revealing comment was a question: "Deepfake sounds very glamorous but what does it actually mean? It means someone successfully pretended to be somebody else" (World Economic Forum, 2025).
Two features of the case matter more than the quality of the video. First, the employee was not careless. He was suspicious, and his suspicion was the reason a video call took place at all. The fraudsters used the video call to answer his doubt, so the deepfake functioned as a counterfeit verification step. Second, the scheme relied on familiar levers. The request came from the most senior finance figure in the organisation, it was framed as secret, and it was confirmed by several apparent colleagues at once.
Old principles, new props
Stajano and Wilson (2011) distilled seven principles from hundreds of scams documented for the BBC programme The Real Hustle. They include the social compliance principle, that people are trained not to question authority; the herd principle, that people let their guard down when others around them appear to share the same risks; the deception principle, that things and people are not what they seem; and the time principle, that under pressure people shift to decision strategies involving less reasoning. They argued that these principles were exploited by fraudsters for centuries before computers existed, and that "Users fall prey to these principles not because they are gullible but because they are human" (Stajano and Wilson, 2011).
Read against these principles, the Arup case is not technologically novel. The fake chief financial officer invokes social compliance. The several fake colleagues provide the herd. The secrecy isolates the target from anyone who might contradict the story. The deepfake supplies the deception, but deception was always part of fraud. What the technology changes is the cost and scale of producing a convincing performance, not the psychological structure of the con.
Goffman's (1974) frame analysis offers a sociological language for this. He defined a fabrication as "the intentional effort of one or more individuals to manage activity so that a party of one or more others will be induced to have a false belief about what it is that is going on" (Goffman, 1974, p. 83). The target of deepfake fraud is not mistaken about a face. They are mistaken about the situation. At Arup, the fabricated frame was a routine, if confidential, instruction from senior management. The synthetic faces were props that held that frame in place. A defence that only examines the props leaves the frame itself unchallenged.
Why detection alone will not win
Much of the current response to deepfake fraud focuses on detection: training staff to spot visual glitches, or deploying tools to analyse audio and video for signs of manipulation. Post-incident advice on the Arup case included asking people on a video call to turn their heads, or to pick up a pencil, to expose a deepfake (Fortune, 2024).
There are two problems with this. The first is practical. Detection is an arms race, and every improvement in generation erodes the value of the cues staff are trained to look for. Certos itself describes media analysis as forcing finance teams into "a probabilistic guessing game" (Certos, 2026). As a vendor of an alternative approach it has an obvious interest in saying so, but the logic holds.
The second problem is criminological. Advice that places the burden on individuals to spot a fake is another form of responsibilisation. When the fake is undetectable, the person who was deceived will be judged to have failed a test they could not have passed. The Arup employee's suspicion should be read as evidence that awareness worked as far as awareness can. What failed was the absence of a process that did not depend on his judgement of a face on a screen.
Preventing the script, not the scene
Crime script analysis helps us understand money mule activity as a sequence of stages rather than a single act. The same approach clarifies deepfake fraud. The deepfake operates at one scene of the script, the moment of persuasion. But the script cannot be completed without later stages that no synthetic face can fake: an instruction that bypasses normal authorisation, a payment to a new or changed account, and a receiving account controlled by the offender or by a mule.
Those stages are where situational prevention has most purchase. Independent call-back to a number already held on file, dual authorisation for high-value or unusual payments, mandatory holds on new payees and checks that the receiving account belongs to the intended recipient all increase the effort required, and none of them rely on spotting a fake. Certos's survey found that 58 per cent of respondents had increased multi-factor authentication, 51 per cent had stepped up employee training and 37 per cent had added manual review holds for higher-risk payments (Certos, 2026). The fact that holds are the least common of the three suggests that organisations are still investing more in recognising the performance than in interrupting the script.
Secrecy deserves particular attention. A request to keep a transaction confidential is a persuasion technique that works regardless of the medium. An organisational rule that any request for secrecy around a payment automatically triggers independent verification would have applied at Arup whether the instruction came by email, by phone or by a flawless video call.
Activity is not prevention
Closing a mule account is not the same as closing the case. AML controls can exist on paper without working in practice, a raid count cannot tell us whether a scam compound industry has been dismantled or displaced, reimbursing victims measures fairness rather than prevention, and platform governance can be documented without changing what platforms do. The common thread is that visible activity, whether accounts closed, raids conducted, claims paid or policies written, is too easily mistaken for prevention.
Deepfakes invite the same mistake. Investment in detection tools is visible and easy to report. But the crime is persuasion, and persuasion is defeated by designing processes that do not depend on any single person believing what they see. The pixels will keep improving. The principles they exploit have not changed in centuries, and that is precisely why they are where prevention should focus.
References
Association of Certified Fraud Examiners (2026) 'Study: Deepfake fraud surges – and only 7% of organizations are firmly ready', press release for the 2026 Anti-Fraud Technology Benchmarking Report (with SAS), 25 March. Available at: https://www.acfe.com/about-the-acfe/newsroom-for-media/press-releases/press-release-detail?s=2026-anti-fraud-technology-benchmarking-report-pr
Certos (2026) 'AI deepfakes target corporate America's multi-bank blind spot', PR Newswire, 8 October. Available at: https://www.prnewswire.com/news-releases/ai-deepfakes-target-corporate-americas-multi-bank-blind-spot-302898894.html
CNN (2024) 'Arup revealed as victim of $25 million deepfake scam involving Hong Kong employee', 16 May. Available at: https://www.cnn.com/2024/05/16/tech/arup-deepfake-scam-loss-hong-kong-intl-hnk
Fortune (2024) 'A deepfake "CFO" tricked the British design firm behind the Sydney Opera House in $25 million scam', 17 May. Available at: https://fortune.com/europe/2024/05/17/arup-deepfake-fraud-scam-victim-hong-kong-25-million-cfo/
Goffman, E. (1974) Frame Analysis: An Essay on the Organization of Experience. Cambridge, MA: Harvard University Press. Available at: https://archive.org/details/frameanalysisess0000goff_u8p5
Stajano, F. and Wilson, P. (2011) 'Understanding scam victims: Seven principles for systems security', Communications of the ACM, 54(3), pp. 70–75. Available at: https://doi.org/10.1145/1897852.1897872
World Economic Forum (2025) 'Cybercrime: Lessons learned from a $25m deepfake attack'. Available at: https://www.weforum.org/stories/emerging-technologies/deepfake-ai-cybercrime-arup/
Published by The Financial Crime Lab on 16 Nov 2026.
Cite as: The Financial Crime Lab (2026). The Crime Is Persuasion, Not the Pixels: Rethinking Deepfake Fraud.
Put this evidence to work in your organisation.
Book a consultation with Dr Nicola Harding to translate these findings into prevention strategy, product design and frontline practice.

