Only 0.1% of people can reliably tell a deepfake from real media - one in a thousand, from an iProov study of 2,000 US and UK consumers who were even told in advance to look for fakes (iProov, 2025). Machines do better, but unevenly: the top audio detector in a neutral 2026 benchmark scored 98.1% accuracy, while the weakest commercial tool managed 71.3% and open-source models landed between 48% and 63% (Resemble AI, 2026). Meanwhile deepfake fraud attempts jumped more than 1,300% in a single year across 1.2 billion analyzed calls (Pindrop, 2025), and Deloitte projects US generative-AI fraud losses hitting $40 billion by 2027 (Deloitte, 2024). Detection has become a measurable arms race with published accuracy scores, benchmark leaderboards, and a market growing near 48% a year. This analysis consolidates data from iProov, Resemble AI, Pindrop, Gartner, Deloitte, and 12 other primary sources into the detection numbers that matter for 2026.
TL;DR
- Average human deepfake detection accuracy across 56 studies and 86,155 participants: 55.5% (Diel et al., 2024).
- Consumers who correctly identified every real and fake clip shown: 0.1% (iProov, 2025).
- Top audio detector accuracy on a neutral 2026 benchmark: 98.1% (Resemble AI, 2026).
- Open-source deepfake detectors accuracy range on that benchmark: 48-63% (Resemble AI, 2026).
- Detector performance lost under real-world image degradation: up to 50% (RedFace dataset, arXiv 2025).
- Organizations hit by a deepfake attack in the past 12 months: 62% (Gartner, 2025).
- Year-over-year surge in deepfake fraud attempts: 1,300% (Pindrop, 2025).
- Projected US generative-AI fraud losses by 2027: $40 billion (Deloitte, 2024).
- Deepfake detection market size by 2034: $5.6 billion, 47.6% CAGR (Market.us, 2025).
- AI image generators using adequate watermarking: 38% (Rijsbosch et al., 2025).
- EU AI Act Article 50 deepfake-labeling rules apply from: 2 August 2026 (EU AI Act).
1. Humans Are Barely Better Than a Coin Flip
The single most important detection statistic is not about software - it is about people, and it is bleak. Pooling 56 controlled studies, researchers found average human accuracy sits at 55.5%, statistically indistinguishable from chance once confidence intervals are accounted for (Diel et al., 2024). People are marginally best at spotting fake audio and worst at fake images, but no modality clears a reassuring bar. The practical upshot: unaided human judgment is not a control you can rely on, which is exactly why automated detection and content provenance are becoming mandatory.
| Metric | Value | Source |
|---|---|---|
| Average human detection accuracy (all modalities) | 55.5% | Diel et al., 2024 |
| Human accuracy on audio deepfakes | 62.1% | Diel et al., 2024 |
| Human accuracy on video deepfakes | 57.3% | Diel et al., 2024 |
| Human accuracy on image deepfakes | 53.2% | Diel et al., 2024 |
| Consumers who flagged every real and fake clip correctly | 0.1% | iProov, 2025 |
| Drop in accuracy on synthetic video vs image | 36% | iProov, 2025 |
| Adults 65+ who had never heard of deepfakes | 39% | iProov, 2025 |
| People overconfident in their detection ability | 60%+ | iProov, 2025 |
Overconfidence compounds the risk: more than 60% of iProov respondents rated themselves confident detectors regardless of whether they were right, and 48% did not know how to report a suspected deepfake (iProov, 2025).
2. What Detection Tools Actually Score on Benchmarks
Software closes much of the gap - on clean inputs. In a neutral May 2026 audio benchmark administered through Podonos across roughly 25 text-to-speech systems and six audio formats, commercial detectors clustered high while open-source models trailed badly. Resemble AI topped the field at 98.1% accuracy (F1 0.981), but the spread is the story: the weakest commercial entrant scored 71.3%, and open-source detectors barely beat human performance at 48-63% (Resemble AI, 2026). Pindrop, the sole large-organization winner of the 2024 FTC Voice Cloning Challenge, reports 99% accuracy on known engines from two-second audio chunks.
| Metric | Value | Source |
|---|---|---|
| Resemble AI accuracy (2026 audio benchmark) | 98.1% | Resemble AI, 2026 |
| Aurigin AI accuracy | 96.8% | Resemble AI, 2026 |
| Hive accuracy | 83.5% | Resemble AI, 2026 |
| Reality Defender accuracy | 71.3% | Resemble AI, 2026 |
| Open-source models accuracy range | 48-63% | Resemble AI, 2026 |
| Pindrop accuracy on known engines (2s speech) | 99% | Pindrop, 2024 |
| Pindrop detection on unseen engines | 90%+ | Pindrop, 2024 |
| Best ASVspoof 5 equal error rate (no augmentation) | 5.56% | ASVspoof 5 Challenge, 2024 |
Note: the 2026 audio benchmark was run by a vendor (Resemble AI) using privately held test labels; the underlying Podonos test set and metrics are disclosed. Pindrop’s 99%/<1%-false-positive figure originates in a 2023 study and is the most recent published headline accuracy (Pindrop, 2024).
3. The Robustness Gap: Lab Scores vs Real-World Inputs
Benchmark accuracy is a ceiling, not a floor. Detectors that shine on pristine files degrade fast against the compression, noise, and codec artifacts of real phone calls and social feeds. On the RedFace real-world dataset, most detection methods lost up to 50% of their performance under image degradation (arXiv, 2025), and every system in the 2026 Podonos audio benchmark struggled once audio was compressed to phone-call quality. Speed is a second constraint: real-time defense needs a real-time factor below 1.0, which not every accurate tool achieves. For the voice channel specifically, our voice biometrics statistics for 2026 cover how liveness and anti-spoofing layers are being stacked to compensate.
| Metric | Value | Source |
|---|---|---|
| Detector accuracy loss under image degradation (RedFace) | up to 50% | RedFace dataset, arXiv 2025 |
| Reality Defender false-positive rate on genuine voices | 53.7% | Resemble AI, 2026 |
| Reality Defender real-time factor (over live threshold) | 1.52 | Resemble AI, 2026 |
| Streaming-capable detectors real-time factor | 0.33-0.40 | Resemble AI, 2026 |
| Face-biometric injection attacks increase (2023) | 200% | Gartner, 2024 |
| Human accuracy with feedback training or AI support | 65.1% | Diel et al., 2024 |
Outlier: a detector trained on the older ASVspoof 2019 attack set can post a sub-2% error rate in the lab yet fail to generalize to 2026-era voice-cloning engines - the benchmark ages faster than the threat.
4. Fraud Losses Are Driving Detection Demand
Detection spending tracks fraud pain, and the pain is escalating on the voice channel above all. Pindrop logged a 1,300% surge in deepfake fraud attempts across 1.2 billion analyzed calls, from roughly one a month to seven a day, with synthetic-voice attacks on insurers up 475% (Pindrop, 2025). Deloitte’s Center for Financial Services models US generative-AI fraud losses tripling from $12.3 billion in 2023 to $40 billion by 2027. This section focuses on the threat pressure behind detection; for overall incident volume and victim counts, see our companion deepfake statistics for 2026, and for the voice-specific angle, our voice cloning fraud statistics and vishing statistics.
| Metric | Value | Source |
|---|---|---|
| Organizations hit by a deepfake attack (past 12 months) | 62% | Gartner, 2025 |
| Year-over-year deepfake fraud attempt surge | 1,300% | Pindrop, 2025 |
| Synthetic-voice attacks on insurers | 475% | Pindrop, 2025 |
| Projected US generative-AI fraud losses by 2027 | $40 billion | Deloitte, 2024 |
| Rise in sophisticated fraud year-over-year | 180% | Sumsub, 2025-2026 |
| Deepfakes as share of top first-party fraud schemes | 11% | Sumsub, 2025-2026 |
| Loss from a single deepfake CFO video call (Arup) | $25 million | CNN, 2024 |
| Enterprises to distrust standalone ID verification by 2026 | 30% | Gartner, 2024 |
Context: Gartner’s 62% figure comes from a September 2025 survey of 302 cybersecurity leaders across North America, EMEA, and Asia/Pacific; 32% of the same group also reported attacks on their AI applications.
5. The Detection Market and Where the Money Goes
Every accuracy gap above is a line item in a fast-growing budget. Estimates vary widely by definition - detection-only tools versus the broader deepfake-technology stack - but the direction is unanimous. Market.us sizes the deepfake detection market at $114.3 million in 2024, scaling to $5.6 billion by 2034 at a 47.6% CAGR, one of the steepest growth rates in cybersecurity. North America commanded 42.6% of that spend in 2024, and video-and-image detection made up two-thirds of tooling revenue - a mismatch worth watching, given that the sharpest fraud growth is on the voice channel.
| Metric | Value | Source |
|---|---|---|
| Deepfake detection market (2024) | $114.3 million | Market.us, 2025 |
| Deepfake detection market (2034) | $5.6 billion | Market.us, 2025 |
| Deepfake detection market CAGR (2025-2034) | 47.6% | Market.us, 2025 |
| North America share of detection market (2024) | 42.6% | Market.us, 2025 |
| Video and image detection segment share (2024) | 66.7% | Market.us, 2025 |
| Deepfake AI market by 2031 (MarketsandMarkets) | $7.27 billion | MarketsandMarkets, 2025 |
| Broader deepfake technology market by 2034 | $51.42 billion | Fortune Business Insights, 2025 |
Divergence note: forecasts differ by an order of magnitude depending on scope. Market.us measures detection tooling alone, MarketsandMarkets tracks the full deepfake-AI market ($0.85 billion in 2025), and Fortune Business Insights counts the broadest generation-plus-detection technology market - read each figure against its definition, not against the others.
6. Labeling Laws and Watermarking Adoption
Regulation is shifting the burden upstream, from catching fakes after the fact to marking synthetic media at creation. The EU AI Act Article 50 transparency rules - requiring deepfakes to be labeled even absent intent to deceive - apply from 2 August 2026, with penalties up to EUR 15 million or 3% of global turnover. China’s synthetic-content labeling measures already took effect in September 2025. Yet compliance tooling lags the law badly: a 2025 audit of 50 AI image generators found most ship without the marking the rules will soon demand. Consent-based provenance - where a person deliberately labels media made from their own voice model - is the cooperative side of the same problem detectors solve adversarially.
| Metric | Value | Source |
|---|---|---|
| AI image generators using adequate watermarking | 38% | Rijsbosch et al., 2025 |
| AI image generators applying deepfake labeling | 18% | Rijsbosch et al., 2025 |
| AI image generators audited in the study | 50 | Rijsbosch et al., 2025 |
| EU AI Act Article 50 application date | 2 August 2026 | EU AI Act |
| Max EU AI Act transparency penalty | EUR 15M or 3% turnover | EU AI Act |
| China synthetic-content labeling effective date | 1 September 2025 | China CAC, 2025 |
| Pindrop TTS/voice-cloning systems used to train detection | 120+ | FTC Voice Cloning Challenge, 2024 |
Outlier: the same study warns that audio, video, and text watermarking will be even harder to implement than the image watermarking it measured - so the 38% figure is likely an optimistic ceiling for the modalities most used in voice fraud.
Summary: Deepfake Detection by the Numbers
| Metric | Value | Source |
|---|---|---|
| Average human detection accuracy (56 studies) | 55.5% | Diel et al., 2024 |
| Human accuracy on audio deepfakes | 62.1% | Diel et al., 2024 |
| Consumers who flagged all real and fake clips | 0.1% | iProov, 2025 |
| People overconfident in detection ability | 60%+ | iProov, 2025 |
| Top audio detector accuracy (2026 benchmark) | 98.1% | Resemble AI, 2026 |
| Open-source detector accuracy range | 48-63% | Resemble AI, 2026 |
| Pindrop accuracy on known engines (2s speech) | 99% | Pindrop, 2024 |
| Detector accuracy loss on degraded real-world data | up to 50% | RedFace, arXiv 2025 |
| Reality Defender false-positive rate on real voices | 53.7% | Resemble AI, 2026 |
| Organizations hit by a deepfake attack (12 months) | 62% | Gartner, 2025 |
| Year-over-year deepfake fraud attempt surge | 1,300% | Pindrop, 2025 |
| Rise in sophisticated fraud year-over-year | 180% | Sumsub, 2025-2026 |
| Projected US generative-AI fraud losses by 2027 | $40 billion | Deloitte, 2024 |
| Loss from the Arup deepfake CFO scam | $25 million | CNN, 2024 |
| Enterprises to distrust standalone ID verification by 2026 | 30% | Gartner, 2024 |
| Deepfake detection market by 2034 | $5.6 billion | Market.us, 2025 |
| Deepfake detection market CAGR | 47.6% | Market.us, 2025 |
| AI image generators using adequate watermarking | 38% | Rijsbosch et al., 2025 |
| EU AI Act Article 50 application date | 2 August 2026 | EU AI Act |
Methodology and Sources
Data was gathered by aggregating primary reports, peer-reviewed and preprint academic studies, named public benchmarks, government and regulatory sources, and vendor disclosures with published methodology, cross-referencing market and fraud figures across multiple sources and flagging divergences inline. Every statistic here was verified against the cited source during research; where the best available figure predates 2025, its year is stated.
- iProov - Deepfake Detection consumer study, 2025 (press release)
- Diel, A. et al. - Human performance in detecting deepfakes: a systematic review and meta-analysis of 56 papers, 2024 (ScienceDirect)
- Resemble AI / Podonos - Audio Deepfake Detection Benchmark, 2026 (report)
- Pindrop - 2025 Voice Intelligence and Security Report (report) and Pindrop Pulse detection (overview)
- ASVspoof 5 Challenge - audio anti-spoofing evaluation, 2024
- RedFace real-world forgery dataset - arXiv, 2025 (paper)
- Gartner - cybersecurity leader survey, September 2025 (via Infosecurity Magazine); and 30%-by-2026 identity-verification prediction, 2024 (press release)
- Deloitte Center for Financial Services - generative-AI fraud forecast, 2024 (insight)
- Sumsub - Identity Fraud Report 2025-2026 (newsroom)
- Arup deepfake CFO scam - Hong Kong Police via CNN, 2024 (report)
- Market.us - Deepfake Detection Market report, 2025 (report)
- MarketsandMarkets - Deepfake AI Market report, 2025 (press release)
- Fortune Business Insights - Deepfake Technology Market report, 2025 (report)
- Rijsbosch, B. et al. - Missing the Mark: watermarking adoption under the EU AI Act, 2025 (arXiv)
- EU AI Act - Article 50 transparency obligations (text)
- FTC - Voice Cloning Challenge winners, 2024 (press release)
- China Cyberspace Administration - synthetic-content labeling measures, 2025
Data watch: several cited sources refresh on annual cycles - Pindrop’s Voice Intelligence and Security Report typically lands mid-year, Sumsub’s Identity Fraud Report publishes around year-end, and Gartner reissues its security-leader survey and identity-verification forecasts annually; the Resemble AI / Podonos audio benchmark and NIST-adjacent evaluations update as new voice-cloning engines appear, and EU AI Act transparency obligations become enforceable 2 August 2026, which will prompt fresh watermarking-compliance audits.
Last updated: July 9, 2026. We review and update this page quarterly as new data is published.