AI Voices Are Getting Better and People are Trusting the Real Ones Less

A new large-scale study of human deepfake perception – 1,768 participants, 35,532 judgments, 138 different AI voice systems – found that people are about as good at spotting fake voices as they were four years ago (71.2% accuracy in 2026 vs. 72.9% in 2021). What collapsed is their ability to recognize genuine audio: accuracy on real voices dropped from 72.7% to 64.1%. Researchers Nicolas Müller and Wei Herng Choong (Fraunhofer AISEC, 2026) call this a “skepticism shift.” The threat isn’t that people are being fooled more. It’s that they no longer believe what’s real.

How accurate are humans at spotting AI-generated/deepfake voices in 2026?

Müller and Choong’s 2026 study is the largest human deepfake-listening experiment published to date, directly replicating a 2021 baseline study four years later to measure how perception has changed as text-to-speech technology has advanced.

Source: Müller, N.M. & Choong, W.H. (2026), “Eroding Trust in Real Speech: A Large-Scale Study of Human Audio Deepfake Perception,” arXiv:2605.26136.

People are trusting genuine voices less

Two numbers barely moved. One fell off a cliff.

Even the study’s own baseline ML detector, run as a reference point alongside the human participants, doesn’t close that gap by much on the categories that matter most. Here’s how human accuracy, the study’s baseline detector, and ValidSoft’s AI Voice Intelligence Platform compare across the four scenarios that count:

That’s the gap worth paying attention to: on commercial-grade fakes, the ones built with the same widely available APIs anyone can sign up for, human accuracy sits at 61.3% while ValidSoft’s platform holds at 99.7%, a 38-point spread. Even against the study’s own baseline detector, that’s a meaningfully larger gap than the study’s average 94.5% ML performance would suggest.

Why do people distrust real audio now, even when it’s genuine?

The researchers’ explanation is straightforward: four years of exposure to increasingly convincing synthetic voices has made people generally more suspicious of audio, not just more skilled at catching fakes. Participants aren’t failing to notice synthesis artifacts. They’re second-guessing recordings that never had any.

That distinction matters. A detection failure is a technical problem. A trust failure is a behavioral one, and it doesn’t reverse just because detection tools improve.

Which AI voice generators are hardest to detect?

Not all synthetic speech is equally convincing, and the gap is wide:

• Hardest to detect: commercial APIs (61.3% human accuracy) and autoregressive language-model systems like VALL-E and Bark (65.9%)
• Easiest to detect: traditional seq2seq systems (75.4%) and flow-matching models (76.8%)

The uncomfortable part is which category sits where. Commercial APIs and AR-LM systems are also the most widely accessible synthesis technologies available today. The systems most likely to be used at scale are precisely the ones people are worst at catching by ear.

Is deepfake detection enough to stop voice fraud on its own<strong>?</strong>

Is deepfake detection enough to stop voice fraud on its own?

No, but it means detection alone was never going to be the whole answer, and this study is good evidence why.

The paper’s own ML detector held steady at 94.5% accuracy, comfortably outperforming humans across every category. That’s reassuring for the “is this audio real?” question. It says nothing about the question that actually determines fraud losses: is the real, correctly-identified human who they claim to be, and did they actually authorize what’s about to happen?

A skepticism shift makes that gap more dangerous, not less. If contact centre agents, compliance teams or customers are increasingly primed to doubt genuine audio, a system that only answers “real or fake?” risks adding friction to legitimate interactions without addressing the wider trust problem.

And that trust problem is expanding. Increasingly, the party receiving, interpreting or acting upon a human instruction may itself be an AI agent. In those environments, establishing that the voice is real and that the human is the right human is still not enough. The enterprise must also know whether the AI agent involved is the right agent: authenticated, authorised for the task in question, and operating within the mandate it has been given.

Only then comes the final question: was the resulting action actually the right outcome?

That matters because authorised push-payment fraud, coerced transactions, compromised or impersonated AI agents, and agentic systems operating outside their mandate do not necessarily require a fake voice at all.

Real Human? Right Human? Right Agent? Right Outcome?

This is where the skepticism shift lands squarely in ValidSoft’s territory. The problem is no longer simply determining whether audio is real. Trust has to be established across the entire interaction.

There are four questions that increasingly need to be answered:

• Is it Human? Answered by real-time deepfake and synthetic-speech detection through Voice Verity®.
• Is it the Right Human? Answered by voice biometric aRight uthentication against a known identity through VoiceID™.
• Is it the Right Agent? Where an AI agent participates in the interaction, is it the genuine, authorised agent it claims to be, operating within the identity, permissions and mandate assigned to it?
• Is it the Right Outcome? Has the verified human’s intent been correctly understood and bound to the specific transaction, instruction or action ultimately carried out, creating a cryptographically provable record through VoiceMFA™?

The first question is where this study begins. But its broader significance is that authenticity alone can no longer carry the burden of trust.

Even once audio has been established as real, another question immediately follows: is this known-real human the right human?

And increasingly there is another participant in that chain. If an AI agent receives, interprets, relays or executes the instruction, the enterprise must establish that it is the right agent – not an impersonated, substituted, compromised or unauthorised agent, and not a legitimate agent acting beyond the scope of its mandate.

Only then can the enterprise establish the right outcome.

Authentication is not authorisation. Identity is not consent. Agent identity is not agent authority. Presence is not mandate.

A correctly identified real human on a genuine call can still be coerced, deceived or confused. A correctly identified AI agent can still operate outside its permitted authority. And either party can later dispute what was instructed, understood or authorised.

In a climate where trust in audio itself is eroding, what an enterprise ultimately needs is not simply a more confident determination that a voice was real. It needs a provable trust chain connecting the real human, the right human, where applicable the right agent, and the resulting action.

That means establishing that the transaction was genuinely authorised, that every participant was entitled to perform their role, and that the resulting instruction and outcome are auditable, immutable and capable of being proven to a third party after the event.

Frequently Asked Questions

FAQs

Can humans tell the difference between a real voice and an AI-generated one? On average, humans correctly identify fake audio about 71% of the time, roughly the same as four years ago. But accuracy on genuine, real recordings has dropped from 72.7% to 64.1% over that period, according to a 2026 study of 1,768 participants (Müller & Choong, arXiv:2605.26136). People aren’t getting better or worse at spotting synthesis artifacts; they’re growing more suspicious of audio generally, including audio that’s completely real.

Why can’t people trust recorded audio the way they used to? Four years of exposure to increasingly convincing synthetic speech has made listeners more skeptical of audio by default, not just more skilled at catching fakes. Researchers call this a “skepticism shift”: the erosion of trust in genuine recordings, separate from and in some ways more consequential than the risk of being fooled by a fake one.

Which AI voice tools are hardest for people to detect as fake? Commercial voice-cloning APIs and autoregressive language-model systems (the architecture behind tools like VALL-E and Bark) are the hardest for humans to catch, detected correctly only 61–66% of the time. Older text-to-speech architectures remain easier to spot, at 75–77% accuracy. The systems that are most widely accessible are also the ones people are worst at identifying by ear.

Is AI better than humans at detecting deepfake audio? Yes. A machine learning detector in the same study held 94.5% accuracy across all voice types tested, outperforming human listeners by roughly 23 percentage points, and the gap held steady from the 2021 baseline (95.5%).

What is ValidSoft Deepfake Audio Detection Accuracy? ValidSoft’s AI Voice Intelligence Platform can detect deepfake audio with accuracy rates of up to 99.7%. As market leaders in voice and speech science, ValidSoft measures performance against the realities of enterprise deployment, recognising that real-world audio conditions can vary significantly across channels, devices, networks and operating environments. Our focus is therefore not simply on achieving industry-leading laboratory accuracy, but on delivering consistently high detection performance in the complex environments in which enterprises actually operate.

What’s the difference between voice authentication and deepfake detection? Deepfake detection answers one question: was this audio generated by AI? Voice authentication answers another: does this voice actually belong to the person it claims to be?

But as AI agents increasingly participate in customer interactions, commerce and transaction execution, a further question emerges: is the system acting on that human’s instruction the genuine, authorised AI agent, and is it operating within its permitted mandate?

None of those questions alone establishes whether the resulting transaction or instruction was genuinely authorised.

That final gap is closed by binding verified identity, authority and intent to the specific action ultimately undertaken – creating a cryptographically secured, auditable and non-repudiable trust chain from the real human, through the right human and, where applicable, the right agent, to the right outcome.

How can a business protect against voice-based fraud if people can’t reliably tell real from fake anymore? By not relying on human judgment or detection alone. A layered approach, real-time deepfake detection, biometric identity verification, and a cryptographically bound, auditable record of what was authorized and by whom, closes the gap that a listener’s ear (or growing skepticism) can’t.