Can Humans Detect Deepfakes? What the Research Shows

Can People Reliably Spot Deepfakes?

People cannot reliably spot deepfakes, and the gap has widened as generators improved. The clearest evidence comes from controlled studies where participants know they are looking for fakes, which is easier than real life, where nobody warns you.

Sophie Nightingale and Hany Farid published one of the most cited studies in PNAS in February 2022. Across 128 faces, 315 participants classified real and AI-generated faces with an average accuracy of 48.2%, slightly below chance. A second group of 219 people received training on rendering artefacts and feedback after each answer, and reached 59.0%, with no improvement over the course of the trials.

Video is no easier. In iProov's February 2025 study of 2,000 UK and US consumers, only 0.1% correctly identified every real and fake image and video, and participants were 36% less likely to identify a fake video than a fake image. iProov sells biometric verification, so its framing favours that remedy, but the numbers match the academic record.

Can You Fall for Deepfakes Even If You Know About Them?

Yes. Knowing deepfakes exist does not protect you, because the studies above were run on people who had been told to look for them. Real scams add pressure that the lab does not: a familiar voice, an urgent request, a deadline, a reason not to call back.

Confidence is part of the problem. More than 60% of iProov's participants said they were confident in their detection skills, whether or not their answers were right. Nightingale and Farid found that AI-generated faces were rated slightly more trustworthy than real ones, 4.82 against 4.48 on a seven-point scale, possibly because generated faces look more average.

That is why criminal use is effective. The FBI warned in December 2024 that criminals use synthetic video of executives and law enforcement in live calls, cloned voices of relatives in emergency scams, and cloned audio to get into bank accounts. Each of these works by reaching a person before any check runs.

Can You Tell a Cloned Voice by Ear?

Not reliably. Kimberly Mai and colleagues tested 529 listeners on English and Mandarin speech and published the results in PLOS ONE in August 2023. Listeners identified deepfakes 73% of the time, with no difference between the languages, and classified genuine clips correctly only 67.78% of the time in the one-clip-at-a-time setting. Hearing five deepfake examples beforehand improved accuracy by 3.84 percentage points on average.

Phone audio makes the task harder, because calls compress speech and remove much of the detail that might give a clone away. The authors' title states their conclusion plainly: humans cannot reliably detect speech deepfakes. They also found that when two clips were played side by side, one real and one fake, accuracy rose to 85.59%, a setting they note is unrealistic because a genuine reference recording is rarely available during a scam call. For contact centres and families alike, that means the decision has to rest on something other than how the voice sounds.

Why Are People So Bad at Detecting Deepfakes?

People look for visible flaws, and modern generators have removed most of them. Advice that circulated a few years ago, such as watching for unnatural blinking, warped hands or mismatched earrings, describes older tools. Current models produce faces with plausible skin, lighting and motion.

Viewing conditions add to the problem. Most media is watched on a phone, compressed by a platform and seen for a few seconds. In the FaceForensics++ user study of 204 participants, accuracy dropped from 68.69% on raw video to 58.73% on low-quality video, and some subtle manipulation methods were detected below chance. Automated detectors read statistical traces that people cannot see at all, although they also lose accuracy on compressed and unfamiliar media.

What Works Better Than Trying to Spot a Deepfake?

Verification by procedure works better than inspection by eye or ear. The aim is to make the decision independent of whether the face or voice is convincing:

Trained forensic analysts remain valuable for contested cases. The Deepfake-Eval-2024 study found that commercial detectors outperformed open-source models on media circulating in 2024 but did not yet match deepfake forensic analysts, which supports pairing tools with expert review rather than replacing either.

• Call back on a number from your own records before acting on a request for money, credentials or sensitive information.
• Agree a code word with family members, as the FBI recommends, and verification phrases or second approvers for urgent requests at work.
• Treat urgency, secrecy and a request to switch channels as warning signs in themselves.
• Check where a clip came from, whether earlier versions exist, and whether official channels confirm it.
• Use deepfake detection tools on media that will drive a decision, and keep a human in the loop for uncertain results.

How Neural Defend Supports Human Reviewers

Neural Defend's tools are built to give reviewers evidence they cannot see unaided. NeuroDetect shows the original file beside explainability heatmaps and attention views down to regions, frames and audio segments, with an LLM-assisted summary linked to that evidence and a downloadable forensic report; the analyst makes the final call. NeuroVerify scores face images, video and audio across six signals in one API call, returning eligible image results in under 200 ms at p95, so checks happen before a person has to judge. NeuroCall, for live voice channels, is in Beta.

On the public VendorBench-100 benchmark of 100 adversarial still images, Neural Defend ranked first of 36 systems with an MCC of 0.876 against 0.646 for the next vendor. One VendorBench author also co-wrote Neural Defend's ARCAS 1B paper, and the benchmark covers still images only.