Deepfakes are impossible to detect
The claim that deepfakes are completely impossible to detect is an overstatement: automated detection tools exist and continue to improve, but the gap between deepfake generation and detection remains significant, and humans alone are poorly equipped to identify modern deepfakes reliably.
What we know
Deepfakes, AI-generated or AI-manipulated video, audio, and images, are a rapidly evolving technology. Contrary to the claim that they are entirely undetectable, multiple detection approaches exist and are deployed by companies, governments, and researchers. However, the detection problem is genuine and serious: as generation technology advances, detection tools often lag behind, making this a continuously shifting technical contest rather than a settled question in either direction.
A landmark peer-reviewed study published in iScience in 2021 found that humans could correctly identify deepfakes only 57.6% of the time, barely above chance, and that raising awareness or providing financial incentives did not improve accuracy. This establishes that human visual inspection alone is unreliable as a defense. A 2025 benchmark study, Deepfake-Eval-2024, found that even the best open-source automated detectors experienced 45 to 50% drops in accuracy when evaluated on real-world deepfakes collected in 2024, compared to their performance on older academic datasets, a gap that illustrates how detection tools trained on one generation of fake content often fail to generalize to newer generation techniques.
Commercial detection platforms and forensic analysts perform substantially better than either untrained humans or open-source tools. The Deepfake-Eval-2024 benchmark found that the best commercial video detectors achieved approximately 78% accuracy and the best audio detection systems approximately 89% accuracy. Human forensic analysts, using specialized training and dedicated software tools rather than unaided visual judgment, achieve approximately 90% accuracy, higher than current fully automated systems but still meaningfully short of perfect. DARPA's Media Forensics program and separate internal initiatives at Meta, Google, and Microsoft continue to fund and advance detection research, reflecting an institutional recognition that this remains an active and unresolved technical problem.
The practical situation on the ground is nuanced rather than binary. Deepfakes that circulate widely on social media are often detectable with dedicated forensic tools, and many AI-generated images and videos still carry detectable artifacts, including inconsistent lighting, unnatural blinking patterns, or subtle audio-visual synchronization errors that specialized software can flag even when human viewers miss them. At the same time, the most sophisticated deepfakes created specifically for targeted fraud, often produced using expensive, non-public tools and considerable production effort, can defeat both human perception and off-the-shelf automated detectors.
The underlying arms race between generation and detection technology is ongoing and shows no sign of a permanent resolution in either direction, making both the claim that deepfakes are completely undetectable and the opposite claim that detection has fully solved the problem inaccurate descriptions of the current state. Detection researchers generally describe the realistic policy goal not as achieving perfect detection but as raising the cost and technical sophistication required to produce an undetectable fake, combined with complementary measures such as content provenance standards and platform-level labeling requirements.
Common claims
- Humans can reliably spot deepfakes by careful viewing.False, studies show humans identify deepfakes at only slightly above chance rates.
- AI detection tools can reliably identify all deepfakes.False, detection tools perform well on older-style deepfakes but drop sharply on novel real-world examples.
- Deepfakes are impossible to detect.Overstatement, forensic analysts and commercial tools detect many deepfakes, though not all; no method is 100% reliable.
- Deepfakes pose a serious risk to information integrity.Supported, this reflects scientific and government consensus.
Evidence hierarchy
All sources
- Fooled twice: People cannot detect deepfakes but think they caniScience (Cell Press) · 2021
- Deepfake-Eval-2024: A Multi-Modal In-the-Wild BenchmarkarXiv · 2025
- Artificial intelligence, deepfakes, and the uncertain future of truthBrookings Institution · 2022
- Human performance in detecting deepfakes: A systematic reviewComputers in Human Behavior Reports · 2024
- Deepfake Detection TechnologyUK Government Office for Science · 2026