Artificial intelligence has not created a new psychology of fraud. What it is changing is the ability to make old scams more personal, polished and convincing. From tailored phishing messages to cloned voices and synthetic video, AI is weakening some of the signals consumers once relied on to decide whether a message could be trusted.
For years, many online scams carried small warnings inside them. The language felt unnatural. The email contained obvious errors. The caller did not sound quite right. A supposedly urgent message knew surprisingly little about the person receiving it.
Artificial intelligence is gradually removing some of those imperfections.
Generative AI can produce natural-sounding messages, translate them across languages, adapt communication to different audiences and create convincing audio or video. None of these capabilities automatically makes a scam successful, and fraudsters still depend on familiar psychological tactics such as urgency, authority, fear and trust. But AI can make those tactics easier to personalise and potentially cheaper to deploy at scale.
The financial backdrop is already significant. The FBI’s 2025 Internet Crime Report, released in 2026, recorded more than one million complaints and nearly $21 billion in reported internet crime losses in the United States. For the first time, the bureau included a dedicated artificial intelligence category, recording 22,364 AI-related complaints associated with almost $893 million in losses.
The figures do not mean AI caused the wider rise in cybercrime. The FBI’s category covers different ways AI may have appeared within fraudulent activity. But its decision to separate the technology in its annual reporting indicates how difficult AI-enabled deception has become to treat as a fringe issue.
Europol has reached a similar conclusion. Its 2026 Internet Organised Crime Threat Assessment describes AI less as an entirely new form of criminality and more as an accelerator for existing threats. Criminal networks can use it to produce content faster, overcome language barriers and adapt social-engineering narratives to different targets.
Catherine De Bolle, Europol’s executive director, said criminals were exploiting AI “to enhance the speed, efficiency, and scope of their illicit activities.”
For consumers and brands, that shift creates a basic problem. Digital trust has traditionally depended partly on recognition. A familiar name, voice, logo or writing style made communication feel credible. Increasingly, those signals can be reproduced.
Personalisation is making the ordinary scam harder to dismiss
Deepfakes attract attention because they are visually dramatic. The more immediate change in fraud, however, may be much less visible.
It is personalisation.
Traditional phishing works largely through volume. Attackers send large numbers of messages knowing that only a small proportion of recipients may respond. More targeted phishing tends to require information about the recipient, making the message appear connected to that person’s circumstances.
Historically, collecting that information and writing personalised messages required time. AI can compress parts of that process.
A large-scale study presented at the USENIX Security Symposium in August 2026 tested personalised phishing among more than 7,700 participants as part of a university awareness programme. Generic phishing messages achieved a click rate of about 3.9%. Automatically personalised messages created using large language models averaged roughly 10%.
The researchers estimated that automated personalisation could be produced for approximately three cents per email.
The comparison requires caution. Human-written personalised messages performed considerably better in the same study, producing a 24.2% click rate. The research therefore does not show that AI is automatically more persuasive than a skilled human fraudster.
Its importance lies somewhere else.
AI may reduce the amount of human effort required to create individually relevant communication. A fraudster no longer has to choose only between sending generic messages to thousands of people and manually researching every target. Automation can increasingly sit between those two approaches.
That changes the economics of persuasion.
A message that refers to a person’s employer, location, recent activity or interests can feel less random. The scam itself may still be familiar, but the communication surrounding it becomes more believable because it appears to contain context.
Once somebody responds, conversational AI introduces another possibility. Instead of relying entirely on a fixed script, a system can maintain language, answer questions and adapt the conversation.
This matters because fraud rarely succeeds through one perfectly written sentence. Persuasion often develops through interaction. A target hesitates. The supposed sender reassures them. A question is answered. Urgency is reinforced. Trust accumulates over several exchanges.
AI does not need to invent those techniques. It can help sustain them.
Research outside cybersecurity provides some evidence of how powerful personalised AI communication can become. A 2025 study published in Nature Human Behaviour involved 900 participants debating either another person or GPT-4. When the AI was given basic sociodemographic information about its counterpart, it became substantially more persuasive.
In cases where the human and AI did not perform equally, personalised GPT-4 was more persuasive 64.4% of the time, corresponding to an 81.2% increase in the odds of higher post-debate agreement compared with the human baseline.
This was a controlled debate study, not an experiment involving scams, so the result cannot be translated into a fraud success rate. What it demonstrates is the underlying capability to adjust arguments according to information about the audience.
Francesco Salvi, the study’s lead author, described the effect by saying AI systems can make persuasive points while knowing “exactly how to push your buttons.”
For marketers, that finding has an uncomfortable parallel. Personalisation is one of the central promises of modern marketing technology. The same ability to make communication feel relevant can also be misused when the sender’s identity or intention is false.
When a familiar voice stops being proof
AI’s persuasive power becomes more difficult to manage when personalised language is combined with impersonation.
Fraud involving false identities is hardly new. Criminals have long posed as banks, officials, employers, technology companies and relatives. What AI changes is the number of cues that can support the impersonation.
A message can resemble official brand communication. Synthetic images can imitate familiar visual styles. Audio can reproduce a recognisable voice. Video can create the impression that an authority figure is physically present.
The recipient is no longer judging only the story. They may be judging several pieces of apparently supporting evidence at once.
Voice cloning illustrates the problem.
A 2025 peer-reviewed study published in PLOS One found that cloned voices were classified as human in 58% of trials, compared with 62% for genuine human voices in the relevant experiment. The researchers found no statistically significant difference between those conditions, suggesting participants had limited ability to reliably distinguish the cloned voices from the real ones.
Generic AI-generated voices were easier to identify, showing that synthetic speech is not uniformly convincing. Audio quality, the type of model and the listening environment can all affect detection.
Still, the study raises a practical question for fraud prevention. If a voice sounds like someone familiar, how much weight should that familiarity carry?
Increasingly, the answer may be less than before.
The FBI highlighted the problem again in July 2026 when it warned that fraudsters were impersonating the Internet Crime Complaint Center itself. According to the bureau, criminals had used AI-generated videos and impersonation techniques in schemes targeting people who had already experienced fraud.
That example shows why synthetic media should not be viewed only as a deepfake problem.
Its role can be psychological.
A potential victim might first receive a credible message, then see a professional-looking digital property, then encounter synthetic audio or video apparently confirming the identity behind the communication. Each additional signal makes the same false story appear more coherent.
For businesses, this creates an unusual consequence. Years of investment have gone into making brand assets consistent and recognisable. Logos, service scripts, spokespersons, notification styles and customer communication have all been designed to create familiarity.
Fraudsters can attempt to borrow that familiarity.
The challenge for marketers is therefore no longer simply protecting trademarks from misuse. Brand impersonation can affect customer trust directly. If consumers repeatedly encounter fake support agents, false brand messages or synthetic representatives, even legitimate communications can become harder to believe.
India’s fraud numbers show what is at stake
The issue has particular relevance in India because digital payments and customer communication operate at enormous scale.
Official Indian statistics do not yet provide a comparable national figure isolating AI-enabled scams. It would therefore be misleading to attribute the country’s broader cyberfraud losses directly to artificial intelligence.
The underlying exposure, however, is considerable.
In a February 2026 Rajya Sabha response, the Ministry of Home Affairs said the National Cyber Crime Reporting Portal received 24,02,579 financial-fraud complaints in 2025, involving approximately ₹22,495 crore in reported losses.
In 2021, the corresponding figures were 2,62,846 complaints and ₹551 crore.
The ministry cautions that cybercrime portal data are dynamic. More importantly, these numbers represent cyber financial fraud broadly, not AI-specific crime. They show the scale of the environment into which increasingly capable synthetic-media and conversational tools are arriving.
India’s Digital Threat Report 2025-26, prepared for the BFSI and payments ecosystem by CERT-In, CSIRT-Fin and SISA, describes a related challenge as “AI asymmetry.”
The idea is straightforward. Technology can lower the resources and specialist knowledge needed to carry out activities that previously required larger teams, more time or deeper expertise.
For financial fraud, this matters because criminals may not always need to technically defeat a payment system. Persuading the genuine account holder to authorise an action can achieve a similar result.
That shifts some of the cybersecurity problem towards human decision-making.
“As India’s financial ecosystem becomes more interconnected, real-time and technology-driven, cyber resilience must be treated as a shared responsibility,” CERT-In Director General Sanjay Bahl said when the report was released.
For banks, retailers, telecom providers and digital platforms, that shared responsibility increasingly includes customer communication. Fraud can begin wherever a trusted identity can be imitated, including social media, messaging apps, email, search results or fake customer-support channels.
Trust may need to move from recognition to verification
The familiar advice for identifying fraud has often focused on mistakes.
Look for poor spelling. Check whether a message sounds strange. Be suspicious of badly produced websites or unnatural voices.
Those signals remain useful when they appear. They are becoming less reliable as a complete defence.
An AI-generated message can be grammatically correct. A cloned voice can sound familiar. A synthetic video can appear polished. The absence of obvious mistakes no longer carries the reassurance it once did.
That is why fraud prevention is increasingly moving towards verification rather than visual or linguistic judgement.
The FBI advises consumers not to respond immediately to unexpected pressure and to verify unusual communications through independently trusted channels. The principle is particularly important for requests involving money, credentials or sensitive information.
The same approach has implications for brands.
Companies can make legitimate communication easier to verify by maintaining clear customer-service channels, explaining the types of information they will never request and reducing inconsistency between different parts of the customer experience. Internally, sensitive financial or account decisions may require stronger confirmation processes rather than relying solely on a familiar voice, message or video call.
This is not because everything online should suddenly be distrusted.
It is because the definition of credible communication is changing.
AI has not invented urgency, impersonation or emotional manipulation. Nor does the research show that machines have become universally better fraudsters than humans. In several cases, skilled human persuasion remains stronger.
The change is one of scale and accessibility.
Personalised messages can be produced more cheaply. Conversations can be maintained more consistently. Synthetic voices can imitate identity cues that people have historically trusted. Several pieces of false evidence can be combined around one believable narrative.
For marketers, that means digital trust can no longer be treated as only a security department’s problem. Brand recognition, customer communication and fraud prevention are beginning to overlap.
A familiar logo may be copied. A professional email may be generated. A known voice may be cloned.
What AI cannot automatically reproduce is independent verification.
That may become the more valuable signal in the next phase of digital communication. As scammers become better at making something appear real, consumers and companies will increasingly have to ask a different question.
Not simply, “Does this look genuine?”
But, “Can I prove that it is?”
Disclaimer: All data points and statistics are attributed to published research studies and verified market research. All quotes are either sourced directly or attributed to public statements.
Related Articles
