Media, Deepfakes, and Synthetic Content

In 2025, deepfake files on public platforms reached an estimated 8 million — up from 500,000 in 2023, a 1,500% increase in two years. Deepfake-enabled voice cloning fraud alone caused over $3 billion in losses in the US between January and September 2025. And deepfakes now account for 40% of all biometric fraud attempts globally.

The scale has changed. But the more important shift is qualitative: deepfakes have crossed the threshold of reliable human detection.

A 2025 iProov study found that human accuracy at identifying high-quality deepfake videos is 24.5% — barely better than random chance. Only 0.1% of participants correctly identified all synthetic media across the full test. These weren't naive subjects. They were adults actively trying to spot fakes.

The takeaway isn't that people are gullible. It's that the task of visual and audio verification has become genuinely impossible to do reliably without technological assistance.

Why Your Instincts Fail

Three perceptual signals that used to work — and don't anymore:

Audio quality — early synthetic voices sounded slightly mechanical. Modern voice clones, trained on as little as 60 seconds of real audio, are rated as indistinguishable from the original by most listeners.

Visual artifacts — blurred edges, unnatural blinking, lighting inconsistencies — these were the classic deepfake tells. Current generation tools produce videos that pass casual visual inspection with over 90% success against detection tools.

Behavioral fluency — "they seem too stiff" or "the lip sync looks off." These cues still catch low-quality fakes. They don't catch high-quality ones.

The implication is uncomfortable but important: you cannot reliably detect a high-quality deepfake by watching or listening to it. Your perception is not the right tool for this job.

The Provenance-First Approach

If perception fails, what works? Shifting from content evaluation to provenance evaluation.

Instead of "does this look real?", ask "where did this come from, and can I verify that origin through a channel that wasn't itself provided in the suspicious content?"

Specifically:

For video or audio of a person making a significant statement — find the original source. Did the organization whose spokesperson is apparently speaking publish this? Check their official website, verified social accounts, and press contacts independently — not via links in the content itself.

For urgent instructions allegedly from a senior figure — verify through a separate, pre-established channel. Not a reply to the email. Not a callback to the number provided. A call to a number you already have on file. This is now standard protocol in financial institutions, where deepfake CEO fraud has caused verified losses in the millions.

For news events — check whether multiple independent, established outlets are reporting the same thing. A significant real event will have coverage across sources with different known positions. A fabricated event typically breaks on new or fringe sources first.

Content Provenance Technologies

The technical response to synthetic media is developing rapidly. The Coalition for Content Provenance and Authenticity (C2PA) — whose members include Adobe, Microsoft, Google, and the BBC — has developed a standard for embedding cryptographic provenance data directly into media files at the point of creation.

When adopted, this allows a viewer to verify that a video was captured by a specific camera on a specific date and hasn't been altered since. The EU AI Act's transparency requirements, coming into force in August 2026, will mandate disclosure of AI-generated content in certain contexts.

These tools don't yet cover most content in circulation. But they represent the direction of reliable verification — not better human perception, but traceable provenance.