A new generation of deepfake detectors is moving beyond pattern recognition to check whether what you're seeing could actually happen in the real world. Resemble AI launched DETECT-World, a detection system built on world model architecture, the same cutting-edge technology that major AI labs are racing to develop. Instead of just looking for familiar signs of a fake, the system checks whether lighting, shadows, facial continuity and other physical properties obey the laws of physics.

What Makes Physics-Based Detection Different From Traditional Deepfake Tools?

For years, deepfake detectors worked like pattern-matching systems. They learned what fakes looked like by studying thousands of examples, then flagged new content that matched those patterns. The problem: every time someone builds a new AI generator, the detector becomes less reliable until security teams retrain it on fresh examples. Humans alone miss roughly 7 in 10 deepfakes without help, and pattern-based detectors still struggle to keep pace with new generation models.

DETECT-World adds a physics layer on top of that foundation. In a video of someone speaking, the detector looks for whether lighting changes naturally across the scene, whether shadows match their sources, and whether facial features move consistently. A deepfake that passes every known pattern check can still be caught if it breaks the laws of physics. This approach trains faster against new generators and reaches high-confidence detection on day-zero releases, before security teams have time to retrain traditional detectors.

"Frontier labs are moving fast to build the next generation of generative AI. We're moving just as fast to build the detection that keeps pace with them. Our mission is to be the most reliable line of deepfake defense in a world full of companies whose goal is producing content indistinguishable from reality," said Zohaib Ahmed, CEO of Resemble AI.

Zohaib Ahmed, CEO at Resemble AI

How Accurate Is Physics-Based Deepfake Detection?

DETECT-World has been benchmarked across more than 250 generation models. Audio accuracy reaches 99.5% internally and now covers 54 languages, making it practical for global enterprises handling multilingual communications. The system is designed to catch several attack patterns that traditional detectors miss.

• Executive Impersonation on Live Calls: Real-time face-swaps on Zoom or Teams are discovered by analyzing facial and body continuity across frames, minor distortions in features while speaking or moving, and probable lighting and reflections.
• Synthetic Identity Fraud: Deepfake selfies and synthetic ID photos are caught by inspecting liveness cues like blink patterns and depth inconsistencies, and by detecting injection-attack signatures where a live camera feed has been spoofed or swapped out.
• Claims and Evidence Review: Videos and images submitted for reimbursement or claims processing are inspected for lighting and shadow inconsistencies, unnatural continuity between frames, and signs of digital splicing or manipulation.

The timing matters. Gartner recently named deepfakes one of four critical cybersecurity threats where attackers currently hold the advantage. Generative AI has made deepfake creation more realistic, accessible and increasingly real-time, expanding attackers' ability to impersonate identities in fraud, social engineering and recruitment scams.

Why AI Phishing and Deepfakes Are Becoming a Unified Threat

Deepfake detection doesn't exist in isolation. Attackers are combining synthetic media with AI-powered phishing to make fraud campaigns more convincing. AI phishing uses artificial intelligence to research targets through open-source intelligence (OSINT), generate personalized lures, impersonate trusted people and manage conversations across multiple channels. IBM X-Force researchers demonstrated the speed advantage: a convincing phishing email can be generated in five minutes using five prompts, compared with approximately 16 hours for an experienced social-engineering team to craft a comparable message.

The 2024 Hong Kong incident involving engineering firm Arup illustrates the real-world consequence. An employee joined a video conference populated by fabricated participants and authorized a wire transfer of approximately $25 million. The lesson is direct: visual familiarity is not an authorization control. Finance teams should verify unusual payment instructions through a trusted channel that was not introduced by the suspicious message or call.

How to Build Layered Defenses Against Deepfake and AI Phishing Attacks

• Combine Multiple Detection Methods: Layered verification that combines forensic analysis, provenance checks and human review is more reliable than any single deepfake detection method used alone. Frame-by-frame, lip-sync and audio forensic analysis expose inconsistencies between visual and audio signals that a single detector often misses.
• Implement Out-of-Band Verification: Detection scores indicate a probability rather than proof of authenticity, so high-risk decisions such as payments, credential changes or hiring approvals require independent verification through a channel not introduced by the suspicious request.
• Train Employees on Verification Habits: Continuous, role-based training and multichannel phishing simulations build the verification habits that generic annual awareness training cannot. Employees should recognize the requested action and verification failure rather than rely on spelling errors as the primary warning sign.
• Deploy Phishing-Resistant Controls: Layered controls including phishing-resistant multi-factor authentication (MFA), SPF, DKIM and DMARC email authentication, along with dual-approval payment workflows, reduce the damage a successful deception can cause.
• Measure Human-Risk Metrics: Human-risk metrics such as reporting speed, repeat-failure rate and time to remediation measure resilience more accurately than click-through rate alone.

DETECT-World is now available via real-time streaming and batch API. Deployment options include cloud, virtual private cloud (VPC), on-premises and air-gapped environments, with SIP and SIPREC integration for live call and telephony use cases. The system is part of a broader security stack that includes explainability tools, identity management, fraud-pattern checking and watermarking for content an organization creates.

The broader lesson is that no single control stops deepfake attacks. Detection is one layer. Explainability turns a detection score into evidence an analyst can act on immediately. Identity management enrolls a voice or likeness for protection. And fraud-pattern checking compares callers and content against known attack signatures across audio, image, video and text. Together, these layers protect enterprise decisions from start to finish.