“As of 2026, AI-generated videos are realistic enough to fool the majority of viewers without the use of technical detection tools.”

Study 1 found that machines achieved excellent performance in classifying real and deepfake images, with good accuracy in feature classification. Humans, in contrast, experienced challenges in distinguishing between real and deepfake static images. Their classification accuracy was at chance level. Using dynamic video stimuli, Study 2 found that performance of machines was near chance level... humans outperformed machines in the detection of video deepfakes.

Using dynamic video stimuli, Study 2 found that performance of machines was near chance level, with poor feature classification. Further, machines showed greater lie bias and reduced decision confidence relative to humans who outperformed machines in the detection of video deepfakes.

In a large recent study, psychologists and computer scientists at the University of Florida found that AI programs were up to 97% accurate at detecting pictures of deepfake faces. Participants in the study performed no better than chance. However, the algorithms’ performance declined sharply when it came to detecting deepfake videos. In those tests, programs performed at chance levels, while humans correctly identified real and fake videos about two-thirds of the time.

Synthetic video generation is progressing very rapidly. The latest models can produce very realistic high-resolution videos that are virtually indistinguishable from real ones. Although several video forensic detectors have been recently proposed, they often exhibit poor generalization, which limits their applicability in a real-world scenario.

Visual media has long served as a reference for what feels real, but generative video models now produce synthetic footage closely resembling real-world scenes. Despite clear disclosure, the AI-exposure group reported increased doubt about subsequent videos' authenticity, reduced judgment confidence, greater perceptual disruption, and lower social connectedness.

Over the course of 2025, deepfakes improved dramatically. AI-generated faces, voices and full-body performances that mimic real people increased in quality far beyond what even many experts expected... For many everyday scenarios — especially low-resolution video calls and media shared on social media platforms — their realism is now high enough to reliably fool nonexpert viewers. In practical terms, synthetic media have become indistinguishable from authentic recordings for ordinary people.

This paper reports two studies in which non-experts completed a deepfake detection task... raising awareness about the dangers of deepfakes or offering cash incentives for correct detections did not enhance detection accuracy. This suggests that the inability to detect deepfakes reflects a skill deficit.

A new research report from Microsoft warns that no single technology can reliably distinguish AI-generated content from authentic media. The report also expresses concern about AI-based deepfake detectors, which Microsoft's research team described as a useful but inherently unreliable last line of defense. Proprietary detectors built by Microsoft's AI for Good team showed accuracy in the range of 95% under non-adversarial conditions, but the report cautioned that the 'cat-and-mouse' dynamic between AI generators and detectors means no detection tool can be considered fully reliable.

AI systems are far better than people at spotting deepfake images, but when it comes to deepfake videos, humans may still have the edge. That’s the surprising twist from a new study that pits people against machines in the race to detect digital forgeries.

For many everyday scenarios — especially low-resolution video calls and media shared on social media platforms — their realism is now high enough to reliably fool nonexpert viewers. In practical terms, synthetic media have become indistinguishable from authentic recordings for ordinary people and, in some cases, even for institutions.

In 2026, deepfakes will no longer be novel; they will be routine, scalable, and cheap, blurring the line between the real and the fake. This has profound implications for journalism, democracies, economies, courts and personal reputation.

Human detection rates for high-quality video deepfakes are 24.5%. Around 60% of people believe they could successfully spot a deepfake video or image. A 2025 iProov study found that only 0.1% of participants correctly identified all fake and real media shown.

Human detection of high-quality deepfake videos is only 24.5% accurate. For images, human detection averages about 62% accuracy. Only 0.1% of participants across modalities could reliably spot fakes in mixed tests. Participants detect real stimuli about 68.1% of the time, but struggle more with fakes.

Sora 2 is the most realistic and complex model available. The model aims to produce videos that look like they were filmed with a real camera. The AI-generated dialogue highlighted technical features and design, selling the product effectively without context. The ad felt like it was filmed by a real person just now. This is a game-changer for creators needing professional ads without a full production crew.

In controlled studies, human accuracy in identifying high quality deepfake videos plummets to a dismal 24.5%. For images, it's only slightly better at 62%. Surveys consistently show that around 60% of people believe they could successfully spot a deepfake video or image. Humans are extremely inaccurate, correctly identifying high quality deepfake videos only about 24.5% of the time.

The challenge in 2026 is no longer identifying poor-quality deepfakes. It is understanding why highly realistic AI-generated videos are trusted even when nothing explicitly looks wrong. The 2026 International AI Safety Report, cited by The Guardian, notes that AI-generated content has become harder to distinguish from real media compared to just a year earlier.

The gap between AI-generated and traditionally produced video continues to narrow. What started as experimental technology producing inconsistent results has become a reliable production tool. The improvements span multiple dimensions: resolution has jumped from 720p to native 4K, video length has extended from 3-5 seconds to 20+ seconds, and perhaps most importantly, the physics simulation now produces believable real-world interactions.

In a large recent study, psychologists and computer scientists at the University of Florida found that AI systems performed no better than chance while humans correctly identify deep fake videos about 2/3 of the time. Participants were able to detect subtle inconsistencies in movement, facial expressions, and timing that algorithms missed.

The 'uncanny valley' (the uncomfortable feeling when something is almost-but-not-quite human) has been crossed. Current deepfakes are indistinguishable from real people to human perception. This means: Training employees to 'spot deepfakes' is like training them to spot perfect forgeries—impossible; Visual inspection is no longer a viable verification method; Trusting your eyes and ears is now a vulnerability.

By 2026, AI video generators will no longer treat sound as an afterthought. Instead, they will synthesize audio with full contextual awareness, creating seamless alignment between what is seen and what is heard. This evolution turns AI from a generator into an interactive collaborator.

Overall, nearly 83% of consumers in the 2026 State of Video Report say they've watched a video they suspected was AI-generated, with the biggest giveaways cited as robotic gestures (67%), unnatural voices (55%), and lack of emotional tone (51%).

An insider look at deepfakes in 2025. Learn more about how deepfake tech is evolving, what the biggest risks are today, how to detect deepfake.

Multiple peer-reviewed studies from 2025-2026, including those from University of Florida and PubMed-indexed research, consistently show humans at chance level (~50%) or below for detecting deepfake images, and around 65-70% for videos in controlled settings, far below reliable detection. Commercial reports cite even lower rates (24.5%) for high-quality videos, indicating AI videos fool most viewers without tools.

AI content detectors claim 99% accuracy but consistently fail in real-world conditions. As of early 2026, more than a dozen elite universities have disabled AI detection entirely, and OpenAI shut down its own detector after it correctly identified AI text just 26% of the time.

Deepfakes now fool most people. Dr. Ryan Ries explains how to spot AI-generated videos, protect your family, and verify what you see online.

Learn to create lifelike cinematic AI videos in 2026. We need more videos like this but we also need to see the actual time it took, how many failures were generated, and what the real cost is.

In this video, I rank the best AI video generators to use in 2026 after generating over 1,000 videos across every major model. I test each one with the same universal prompt, then push them with use-case-specific scenarios to show where they actually shine, covering realism, motion, audio, speed, and cost.

I walk through how I test the best AI video models in Artlist to see which one gives the most realistic and cinematic AI video. I compare Veo 3.1, Sora 2 Pro, Kling, and Hailuo to see how each AI video model handles camera movement, human motion, prompt accuracy, storytelling, and product shots.

A study found that 1 in 3 people cannot reliably distinguish a deepfake video from a real one. AI-produced video, photo, and audio clips are so convincing that it becomes difficult to identify what is real. It is hard to differentiate AI-manipulated content from human-created content.