Can People Still Tell What’s Real Online? The Deepfake Confidence Gap

Human Deepfake Detection Is Close to Chance in Many Tests

A 2024 systematic review and meta-analysis examined 56 papers involving 86,155 participants and 137 measured effects. Across the studies included in the analysis, total human deepfake-detection accuracy was 55.54%, with a 95% confidence interval of 48.87% to 62.10%. The researchers found that overall deepfake-detection sensitivity was not significantly above chance because the confidence interval crossed the 50% threshold.

Performance also varied by media format. The meta-analysis reported average accuracy of 62.08% for audio, 53.16% for images, 52.00% for text, and 57.31% for video. These figures show that the presence of video does not automatically make synthetic material easy for human observers to identify.

A separate controlled experiment involving 454 participants produced a similar result. Participants classified 20 videos, consisting of 10 authentic videos and 10 deepfakes. Average classification accuracy was 60.7%. Participants were also overconfident about whether individual judgments were correct. Providing participants with visual deepfake-detection strategies did not significantly improve either accuracy or confidence.

Confidence and Accuracy Are Different Measurements

Confidence measures how certain a person is about a judgment; accuracy measures whether that judgment is correct. Experimental deepfake research demonstrates that these measurements do not always correspond.

The 454-participant experiment found overconfidence at the individual-video level even though average classification accuracy was 60.7%. The same participants underestimated their overall performance when asked to estimate how many of the 20 videos they had classified correctly.

Research on fabricated political speeches has also documented substantial differences between media formats. In one preregistered experiment with 501 participants, accuracy was 57% for transcripts, 64% for silent video without subtitles, 69% for silent video with subtitles, 81% for audio, 85% for video with audio and subtitles, and 86% for video with audio but no subtitles.

The same research found that some text-to-speech deepfakes were identified correctly in fewer than 75% of observations. The results demonstrate that detection performance depends partly on the type of synthetic manipulation and the information available to the observer.

Source Verification Provides Information That Visual Inspection Cannot

A website address provides information about where online material is published, but it does not prove that every image, recording, or statement on the site is authentic. Organizations can establish identifiable web properties through registered .com domains, while verification of individual media assets requires additional evidence.

Digital provenance standards are one method for providing that evidence. The Coalition for Content Provenance and Authenticity, or C2PA, develops technical standards that record information about the origin and history of digital media. Its Content Credentials system can record information about an asset’s origin, modifications, editing tools, and use of artificial intelligence. The credentials are cryptographically bound to the asset.

C2PA also states a specific limitation: provenance data cannot independently determine whether the events depicted in an image or video are factually true. The system verifies information about an asset’s history and associated credentials rather than making a factual judgment about the depicted event.

Deepfakes Are Also an Identity-Fraud Tool

Synthetic media has applications beyond misleading public content. Entrust’s 2024 Identity Fraud Report stated that detected deepfake attempts increased by 3,000% between 2022 and 2023. The report also stated that deepfakes represented approximately 30–40% of biometric fraud detected by Onfido during the preceding six months.

These attacks can target identity-verification processes rather than public audiences. In that setting, the objective is to make synthetic biometric material pass a verification system rather than persuade social-media viewers that a video depicts a genuine event.

AI Visibility Now Includes Disclosure Requirements

Online visibility increasingly intersects with the identification of AI-generated material. Businesses publishing digital material therefore operate in an environment where search visibility, source identity, AI-generated content, and content provenance can be separate technical issues. An AI visibility checklist for new businesses can address how business information is presented and discovered, while authenticity mechanisms address the origin and modification history of specific digital assets.

The European Union has introduced legal transparency requirements for certain synthetic content. Article 50 of the EU AI Act requires providers of relevant AI systems to make generated or manipulated outputs detectable in a machine-readable format. Deployers must also disclose image, audio, or video content constituting a deepfake, subject to specified exceptions.

These Article 50 transparency obligations became applicable on 2 August 2026. European Commission guidance states that disclosure of deepfakes must be clear and distinguishable and cannot rely solely on machine-readable markings that require technical tools to detect.

What the Evidence Establishes

Current research does not support the assumption that people can consistently identify synthetic media through unaided observation.

Key measured findings include:

• A 2024 meta-analysis reported overall human deepfake-detection accuracy of 55.54%.
• A controlled video experiment reported average accuracy of 60.7%.
• Detection performance differs across audio, images, text, and video.
• Confidence in an individual judgment can exceed actual detection performance.
• C2PA Content Credentials can document digital provenance but do not determine whether depicted claims are factually true.
• EU transparency rules require disclosure for qualifying deepfakes and machine-readable marking for specified AI-generated or manipulated outputs.

The documented gap is therefore measurable: confidence in judging individual media items can exceed demonstrated detection accuracy, while technical provenance and disclosure systems provide information that visual inspection alone does not supply.