Most people are aware of deepfakes, yet their ability to distinguish them from reality is barely better than a coin flip.

KEY FINDINGS

The detection deficit

Despite the rapid advancement of generative AI, respondents in the US report the lowest familiarity with deepfakes compared with other analyzed markets. Only 63% of US adults are familiar with the term "deepfake," lagging behind both the UK (74%) and Brazil (67%). Furthermore, while 80% of Brazilian respondents report encountering deepfakes online, the US mirrors the UK, with only around 60% reporting such encounters.

Are you familiar with the term ‘deepfake’?

In most countries, younger users are more aware of deepfakes.

However, the US is an exception – younger American users do not show higher awareness of deepfakes compared to older demographic.

“There's a paradox at play: the US is the global epicenter of AI development, yet American consumers are the least familiar with one of its most dangerous byproducts. In markets like Brazil, where digital fraud and social engineering are a persistent part of daily life, people have been conditioned to question what they see online. In the US, there's historically been higher baseline trust in digital content, with the conversation centered more on data privacy than content authenticity. The problem is that low awareness doesn't reduce risk – it amplifies it. If you don't know what a deepfake is, you're far less likely to pause and verify when you encounter one.

This research shows that as AI-generated content becomes indistinguishable from reality – and we're already there – the human eye alone is no longer a reliable line of defense. Awareness remains a critical first step: it gives people the instinct to question rather than trust by default. But businesses and policymakers in the US need to close this awareness gap urgently, while simultaneously investing in automated verification technologies that can catch what humans simply can't.”

When respondents were tested on their ability to distinguish between real and AI-generated visuals, performance across all markets, including the US, was extremely low. On a scale from -1 to 1, where:

• 1 represents perfect accuracy
• 0 represents random guessing (a coin flip)
• -1 represents consistently incorrect answers

American respondents achieved an average detection score of 0.07, matching the UK and slightly behind Brazil's 0.08. A score of 0.07 indicates that, while participants performed slightly better than chance, the difference is minimal. In practical terms, these scores show that the average person is almost guessing when attempting to identify deepfakes, performing only a tiny fraction better than a coin flip.

The US respondents' distribution of results further highlights this challenge

A significant portion of the US public overestimates its ability to spot manipulated media. Around half of the users in the US (and Brazil) are confident in their ability to identify deepfakes, which is notably higher than the 44% confidence rate observed in the UK. While higher confidence generally correlates with higher accuracy in the US and UK, a dangerous "high-risk" segment exists.

Across the markets, approximately 7% of users are classified as high-risk: 

• Demonstrate low detection accuracy
• Express high confidence in their abilities
• Rarely or never verify suspicious content

The high-risk segment is generally less common among the oldest age group in most markets, but again, the US is the exception, meaning older Americans are just as likely to fall into this category. Respondents with a university education are less likely to fall into the "high-risk" category of users who are overly confident but inaccurate.

Approximately 7% of users across markets are classified as high-risk 

Our research reveals what may be the most dangerous dynamic in the deepfake era: overconfidence.

"Around half of US respondents believe they can reliably spot manipulated media and while a slight majority do perform better than random chance, detection scores across both confident and less confident groups remain close to guessing. This confidence-competence gap creates a false sense of security that fraudsters and bad actors are primed to exploit. When people believe they can't be fooled, they stop looking for the signs – and that's precisely when they're most vulnerable, whether to a synthetic identity used in financial fraud or a fabricated video designed to manipulate trust.

Most concerning is the roughly 7% of users across markets who fall into what we classify as the 'high-risk' segment – people who perform poorly at detection, are highly confident they'd catch a fake, and rarely bother to verify suspicious content. This group represents the perfect target for AI-driven fraud and manipulation.

For businesses, the implication is clear: any organization that still relies on manual review processes or customer self-attestation is inheriting this vulnerability directly. Human judgment is becoming an increasingly unreliable safeguard, and verification needs to be built into systems by default – automated, technology-led, and not dependent on the end user's self-assessed ability to tell real from fake."

KEY FINDINGS

User behaviors and concerns

Creating AI visuals is already a widespread practice in the US. Nearly 49% of US respondents report having created AI-generated images or videos at least occasionally (18% multiple times, 31% rarely). This makes Americans significantly more likely to dabble in AI generation than British respondents (38% overall), though they trail behind highly active Brazilian users (59%).

Those with experience creating AI visuals performed notably better at spotting fakes: 61% of AI creators scored above chance, compared to 51% of non-creators. However, this 10 percentage point difference likely reflects a broader profile – age, digital fluency, and general familiarity with AI-generated content – rather than the creation experience alone.

"The reality for businesses is that relying solely on manual review processes or customer self-attestation introduces significant vulnerabilities, especially as threats become more sophisticated. While automation and technology-led verification should be the default to ensure scale, consistency, and speed, the focus should be on strategically integrating human judgment only at critical decision points – where depth, experience, and intuition improve outcomes."

When attempting to identify fakes, American users largely rely on the same visual cues as the rest of the world, though they do so at slightly lower rates. 

The most common indicators cited by US respondents include:

• Unnatural-looking skin (53%)
• Oddities in appearance such as hair or teeth (52%)
• Unnatural movements or facial expressions in video (51%)

While these cues may have been useful in earlier stages of deepfake development, advances in AI have made them increasingly unreliable. Modern deepfakes are capable of replicating these details with high accuracy, reducing the effectiveness of these strategies.

Additionally, the research shows that checking content more frequently does not necessarily improve accuracy, suggesting that users lack a structured or effective approach to verification.

When attempting to identify fakes, American users largely rely on the same visual cues as the rest of the world, though they do so at slightly lower rates. 

The most common indicators cited by US respondents include:

• Unnatural-looking skin (53%)
• Oddities in appearance such as hair or teeth (52%)
• Unnatural movements or facial expressions in video (51%)

While these cues may have been useful in earlier stages of deepfake development, advances in AI have made them increasingly unreliable. Modern deepfakes are capable of replicating these details with high accuracy, reducing the effectiveness of these strategies.

Additionally, the research shows that checking content more frequently does not necessarily improve accuracy, suggesting that users lack a structured or effective approach to verification.

Regardless of respondents’ age or educational background, all groups among Americans are concerned about the real-world impact of deepfakes, though slightly less so than Brazilians, who are highly concerned. In the US, the leading fears are:

However, the US stands out in one important respect: compared to respondents in the UK and Brazil, Americans are more likely to trust social media platforms and digital services to identify and manage AI-generated content. This creates a potential mismatch between perceived and actual protection. While concern is high, reliance on platforms may reduce individual vigilance.

When nearly 8 in 10 Americans say they're concerned about deepfake-driven personal fraud, that's not a hypothetical fear, it reflects a threat that's already materializing.

"We're seeing synthetic identities used to open fraudulent accounts and authorize transactions, and deepfake videos deployed to bypass basic verification checks. The online businesses sit squarely in the crosshairs because they are where identity, money, and trust converge. Every customer onboarding flow, every account recovery process, every high-value transaction is now a potential attack surface for AI-generated fraud.

What makes this particularly urgent for the US market is the combination of high concern and relatively high platform trust. Americans are more willing than their peers in other countries to believe that platforms can handle the problem, which may mean they are less guarded at the individual level, even as the threats intensify. That gap between perceived and actual protection is exactly where fraud thrives. For financial institutions and tech companies, the answer isn't to reassure customers – it's to earn that trust through action. That means deploying AI-powered biometric authentication that can verify a real person in real time, detect synthetic media at the point of interaction, and do so without relying on the customer to spot the fake themselves. The deepfake arms race is an AI problem, and it requires an AI solution."

CONCLUSION

The path forward

Deepfakes are getting better

In the age of generative AI, the most uncomfortable truth this research reveals is not that deepfakes are getting better – it's that human detection was never a sole defense. Across every metric in this study, American consumers demonstrate a consistent pattern: 

The most effective defense is one that keeps humans in the loop, empowered by AI systems that detect what the eye cannot.

Methodology overview

The research was conducted by Kantar in February 2026 using an online access panel.

Total sample: 3,000 respondents

• United States: 1,000
• United Kingdom: 1,000
• Brazil: 1,000

Participants were aged 18 to 64, with quotas applied to ensure nationally representative samples in each country based on age, gender, and region. The survey took approximately 9 minutes to complete and included both demographic questions and practical evaluation tasks.

Respondents were asked to assess 16 visuals:

• 8 real
• 8 AI-generated or manipulated

These included:

• Fully AI-generated images
• AI-generated videos
• Faceswap content

To reduce bias, all visuals were presented in randomized order. Participants completed both individual evaluations and direct comparisons between real and fake content. Detection accuracy was calculated using a scoring index that benchmarks performance against a defined chance-level baseline.

About Veriff

Veriff is a global AI-native identity platform that helps organizations build trust online. Leading companies across financial services, marketplaces, mobility, gig economy, and other digital sectors rely on Veriff’s technology to stay compliant, prevent fraud, protect users, and scale globally.

Veriff’s trust infrastructure supports the full customer journey, from verification to ongoing authentication and fraud prevention, with the least friction for honest people. Built for global scale, Veriff helps businesses expand across borders without the complexity of managing identity verification, compliance, and fraud in multiple markets – creating a single source of truth for trusted identities.