Deepfakes and Synthetic Media: The Threat You Can't See | State of Surveillance

TL;DR: Deepfake volume is growing 900% annually: from 500,000 in 2023 to an estimated 8 million in 2025. A 2025 study found only 0.1% of participants could correctly identify all fake and real media. Voice clones now take just seconds of audio to create and are indistinguishable from real voices. Deepfake-enabled fraud caused over $200 million in losses in Q1 2025 alone. Europol estimates 90% of online content may be synthetic by 2026. The TAKE IT DOWN Act became the first federal law criminalizing nonconsensual AI-generated intimate images. We've reached the "synthetic reality threshold": humans can no longer distinguish real from fake without AI assistance. Here's what you need to know about the threat and how to protect yourself.

The Scale of the Problem

Deepfake proliferation has accelerated beyond what most people realize ^[1].

Volume growth:

500,000 deepfakes shared on social media in 2023
Estimated 8 million deepfakes by 2025
900% annual growth rate
Europol projects 90% of online content may be synthetic by 2026

Human detection has failed:

A 2025 iProov study: only 0.1% correctly identified all fake and real media
Human detection of high-quality deepfake video is only 24.5% accurate
For audio deepfakes, people claimed 73% accuracy but were frequently fooled
For deepfake images, controlled studies show 62% human accuracy

What this means:

We have crossed the "synthetic reality threshold": a point beyond which humans can no longer reliably distinguish authentic from fabricated media without technological assistance. In practical terms, deepfakes are now indistinguishable from real recordings for ordinary people in everyday scenarios.

Types of Synthetic Media Threats

Deepfakes are just one category of synthetic media. The threats are evolving rapidly ^[2].

Deepfake Video

Face swaps: one person's face mapped onto another's body
Face reenactment: making someone appear to say things they never said
Lip sync manipulation: altering what someone appears to say
Full body synthesis: generating entirely fabricated people

Voice Cloning

A few seconds of audio now suffice to create a convincing clone
Clones include natural intonation, rhythm, emotion, pauses, and breathing
Voice phishing jumped 442% in late 2024
Voice cloning has crossed the "indistinguishable threshold"

AI-Generated Images

Photorealistic images of people who don't exist
Altered photos with faces, bodies, or contexts changed
Nonconsensual intimate imagery generated from clothed photos
Fake evidence, documents, and screenshots

Text Generation

AI-generated articles, comments, reviews
Fake social media posts attributed to real people
Synthetic email and messaging content
Fabricated documentation and communications

The Fraud Epidemic

Deepfakes have become a major tool for financial fraud ^[3].

Financial impact:

Over $200 million in losses from deepfake fraud in Q1 2025
Average business loss: nearly $500,000 per incident
Large enterprises: losses up to $680,000
Fraud attempts using deepfakes increased 2,137% over three years

Common fraud scenarios:

CEO fraud: deepfake video or audio of executives authorizing wire transfers
Biometric bypass: deepfakes now account for 40% of biometric fraud attempts
Voice phishing: cloned voices of family members requesting emergency money
Cryptocurrency scams: deepfake-related crypto incidents up 654% from 2023 to 2024

Real examples:

A Hong Kong company lost $25 million after employees were fooled by a deepfake video call appearing to show their CFO
Family members have wired thousands after receiving calls from cloned voices of relatives claiming emergencies
Crypto promotions using deepfake celebrity endorsements have stolen millions

Nonconsensual Intimate Imagery

The most personal form of deepfake abuse continues to grow ^[4].

The problem:

AI tools can now generate realistic intimate imagery from any photograph
No actual nude photos needed: clothed photos are sufficient
Victims are predominantly women and minors
Used for harassment, extortion, bullying, and revenge

School-based abuse:

Students using deepfake tech to create explicit content of classmates and teachers
Existing cyberbullying laws struggle to address the new threat
Schools facing unprecedented challenges in response

Legal response:

TAKE IT DOWN Act: enacted May 19, 2025, first federal law criminalizing distribution of nonconsensual intimate images including AI-generated content
All 50 states and DC now have laws targeting nonconsensual intimate imagery
Some states have updated language to specifically include deepfakes

The technology to create this content is freely available. The legal framework is playing catch-up. Prevention and detection remain inadequate.

How to Spot Deepfakes: Visual Clues

While AI detection tools are increasingly necessary, some visual tells still exist: though they're becoming less reliable ^[5].

Face and Skin

Skin texture: does it appear too smooth or inconsistent with age?
Facial boundaries: look for blurring or artifacts at the edge of the face
Facial hair: unnatural mustaches, beards, or sideburns
Skin tone consistency: color mismatches between face and neck

Eyes and Blinking

Blinking patterns: AI often produces inconsistent or absent blinking
Eye reflections: light reflections should be consistent in both eyes
Gaze direction: unnatural or fixed gaze
Pupil shape: sometimes irregular in deepfakes

Mouth and Speech

Lip sync: misalignment between audio and lip movements
Teeth: sometimes blurry, distorted, or unusually shaped
Inside of mouth: dark or unclear in many deepfakes

Lighting and Physics

Shadows: inconsistent or missing shadows
Lighting direction: face lit differently than surroundings
Head movement: unnatural or limited range of motion
Background: warping or distortion during movement

Important caveat: These tells are becoming less reliable as deepfake technology improves. High-quality deepfakes can fool expert observers. Don't rely solely on visual inspection.

Detection Tools and Technology

AI detection tools are becoming essential, but they have significant limitations ^[6].

Available Tools

Reality Defender: Multi-model platform analyzing video, images, audio, and text. Uses probabilistic detection rather than watermarks.

Hive AI: Deepfake Detection API for images and videos, useful for content moderation.

Microsoft Video Authenticator: Analyzes videos for manipulation, provides confidence scores.

InVID/WeVerify: Browser extension for video verification, includes synthetic media detection.

TruthLens: Research framework that detects and explains which facial regions were manipulated.

Critical Limitations

Detection tools cannot be fully trusted:

They fail when confronted with deepfakes from new techniques
A detector trained on one AI generator may fail on another
Performance drops significantly on new "in-the-wild" fakes
Deepfake creation constantly improves, eliminating the artifacts detectors look for

The arms race reality:

Detection is always playing catch-up. Every time detectors learn to spot specific artifacts, new generation techniques eliminate those artifacts. Machine learning experts recommend treating detectors as one tool among many, not a definitive answer.

Protecting Yourself

A layered defense strategy for the synthetic media era ^[7]:

Verify Before Acting

If you receive an urgent video or audio request for money, verify through a separate channel
Call the person back using a number you already have, not one provided in the message
Establish family code words for emergency situations
Question unexpected video calls from executives or authorities

Limit Your Digital Footprint

Fewer public photos means less material for AI training
Limit video content you post publicly
Be cautious about voice recordings shared online
Review and limit what's accessible on social media

Use Multi-Factor Authentication

Don't rely solely on voice or video for identity verification
Use authenticator apps or hardware keys
Require multiple verification methods for sensitive transactions

Build Critical Skepticism

Assume any surprising video could be manipulated
Verify through multiple independent sources
Be especially skeptical of content that confirms your existing beliefs
Remember that seeing is no longer believing

Stay Informed

Follow developments in deepfake technology and detection
Understand what's currently possible
Recognize that capabilities are constantly expanding

For Organizations

Businesses need specific protections against deepfake threats ^[8]:

Financial Controls

Require multi-person authorization for large transfers
Establish out-of-band verification for payment changes
Never act on video/audio instructions alone for wire transfers
Train finance staff on deepfake fraud scenarios

Communication Protocols

Use authenticated communication channels for sensitive matters
Establish verification procedures for unexpected requests
Train employees to recognize social engineering using deepfakes

Technical Defenses

Deploy deepfake detection tools for critical video communications
Implement liveness detection for biometric authentication
Monitor for fraudulent use of executive likenesses
Consider content authentication systems for official communications

Incident Response

Have a plan for responding to deepfake attacks
Know how to quickly verify and counter false videos of executives
Prepare statements and verification procedures in advance

The Detection Industry

The fight against deepfakes is now a multi-billion dollar market ^[9].

Market growth:

Global deepfake detection market: $5.5 billion in 2023
Projected to reach $15.7 billion by 2026: nearly triple
Major investment from tech firms, governments, and platforms

Research approaches:

Deep learning: neural networks trained to spot synthetic artifacts
Transfer learning: adapting detectors to new fake types
LSTM networks: analyzing temporal patterns in video
Model fingerprinting: identifying which AI created the content
Watermarking: embedding invisible marks in authentic content
Gaze tracking: real-time analysis during video calls

Platform responses:

Social platforms implementing detection and labeling
Video calling services adding authentication features
Content authentication initiatives from camera manufacturers

What's Coming

The next wave of synthetic media threats is already emerging ^[10]:

Near-term developments:

Real-time deepfakes in video calls
Higher resolution and more convincing fakes
Easier tools requiring less technical skill
Multimodal synthesis: coordinated fake video, audio, and text

Societal implications:

Video evidence becoming unreliable in legal contexts
"Liar's dividend": real videos dismissed as fakes
Erosion of shared reality and trust in media
New forms of political manipulation and fraud

The authentication imperative:

The future may require provenance tracking for all media: cryptographic proof of origin from the moment of capture. Without this, distinguishing real from synthetic becomes increasingly impossible.

The Bottom Line

We've crossed a threshold. Synthetic media is now good enough to reliably fool humans. Only 0.1% of people can correctly identify all deepfakes. Voice clones take seconds to create. Fraud losses are in the hundreds of millions.

This isn't a future problem: it's happening now. Eight million deepfakes are projected for 2025, with 900% annual growth. By 2026, Europol estimates 90% of online content may be synthetic.

Detection tools help but can't be fully trusted. They fail on new techniques. They're always catching up. Human verification through separate channels remains essential.

The defenses available today:

Verify unexpected requests through independent channels
Use family code words for emergency calls
Limit public photos and videos
Never act on video/audio alone for sensitive transactions
Treat any surprising video as potentially manipulated

The era of "seeing is believing" is over. Adjust accordingly.