The quality of AI-generated synthetic media has crossed a threshold. In controlled tests conducted by researchers at MIT's Media Lab, human evaluators correctly identified deepfake videos only 52% of the time — statistically indistinguishable from random guessing. The implications for trust, security, and democracy are profound, and the technology is advancing faster than the institutions designed to govern it.
This is not a future problem. Deepfake audio has already been used to commit wire fraud, with criminals cloning executives' voices to authorise fraudulent bank transfers. Deepfake video has been used in political disinformation campaigns in at least a dozen countries. Synthetic images have been used to create non-consensual intimate imagery at scale. The harms are real, they are happening now, and they are accelerating.
The State of Deepfake Technology
The latest generation of video synthesis models can generate photorealistic video of any person saying anything, in real time, with consumer-grade hardware. The computational barrier that previously limited deepfake creation to well-resourced actors has effectively disappeared. Open-source models available on GitHub can produce convincing results in minutes on a standard gaming PC.
The improvement has been dramatic and rapid. Three years ago, deepfakes were detectable by trained eyes — subtle tells in the movement of hair, inconsistent lighting, the uncanny valley of faces that weren't quite right. Two years ago, they were detectable by automated tools with high reliability. Today, the best models produce output that neither humans nor automated tools can reliably distinguish from genuine footage.
The key technical advance has been the shift from GAN-based (Generative Adversarial Network) approaches to diffusion models. GANs produced characteristic artifacts — repeating patterns, unnatural textures — that detection tools learned to identify. Diffusion models produce fundamentally different artifacts that are harder to detect, and the training process is more stable, making it easier to produce high-quality results consistently.
Detection: A Losing Battle?
Detection tools have struggled to keep pace. The fundamental challenge is that detection models trained on current deepfakes become obsolete as generation technology improves. "It's a classic adversarial dynamic," says Dr. Hany Farid of UC Berkeley, a pioneer in digital forensics. "Every time we build a better detector, the generators adapt. The generators have a structural advantage: they only need to fool the detector once, while the detector needs to catch every fake."
The most promising approaches focus not on visual artifacts — which are increasingly imperceptible — but on physiological signals. Subtle variations in skin colour caused by blood flow (rPPG signals) are difficult to synthesise convincingly, and tools that analyse these signals show detection rates above 90% on current deepfakes. But this advantage is likely temporary: as researchers publish their methods, deepfake generators will be trained to reproduce these signals.
A more durable approach may be forensic analysis of the compression and encoding characteristics of video files. Deepfakes, even when visually perfect, often have distinctive patterns in how they interact with video codecs — patterns that are difficult to eliminate without understanding the specific codec being targeted. This approach is less intuitive than visual analysis but potentially more robust.
The Business of Deepfake Detection
A significant commercial ecosystem has emerged around deepfake detection. Companies including Reality Defender, Sensity AI, and Pindrop are offering enterprise detection services targeting financial institutions, media companies, and government agencies. The market is growing rapidly, but the products are imperfect and the competitive dynamics are challenging — detection companies are in a perpetual race against generation technology that is advancing faster and attracting more investment.
The most sophisticated enterprise deployments combine multiple detection approaches — visual analysis, physiological signals, metadata forensics, and behavioural analysis — and treat the output as a risk score rather than a binary determination. A video that scores high on multiple independent detection methods is treated as suspicious and flagged for human review, rather than being automatically rejected. This probabilistic approach is more honest about the limitations of current technology.
Policy and Platform Responses
Major platforms are implementing content provenance standards. The C2PA (Coalition for Content Provenance and Authenticity) standard, now supported by Adobe, Microsoft, Google, and major camera manufacturers, embeds cryptographic metadata in media files that allows verification of their origin and edit history. When a photo is taken on a C2PA-compliant camera, the device signs the image with a cryptographic key. Any subsequent edits are recorded in the provenance chain. A viewer can verify not just that the image hasn't been tampered with, but exactly what edits were made and by whom.
The C2PA approach is technically sound but faces significant adoption challenges. It requires cooperation from device manufacturers, software developers, and platforms simultaneously. It only works for content created on compliant devices — the vast majority of existing content has no provenance data. And it does nothing to address deepfakes created from scratch, rather than manipulated from genuine footage.
Legislative responses have been fragmented. The EU's AI Act includes provisions requiring disclosure of AI-generated content, but enforcement mechanisms are weak and the definition of "AI-generated" is contested. In the United States, a patchwork of state laws addresses specific harms — non-consensual intimate imagery, election-related deepfakes — but there is no comprehensive federal framework. Several countries have enacted outright bans on certain categories of deepfake content, but enforcement against offshore actors is effectively impossible.
The Democratic Threat
The most concerning long-term implication of deepfake technology is not any specific harm but the erosion of epistemic trust — the shared assumption that video and audio recordings are reliable evidence of what actually happened. Once that assumption is gone, it cannot easily be restored. Every genuine piece of evidence becomes deniable: "That's a deepfake." Every genuine statement becomes questionable: "That's AI-generated."
This "liar's dividend" — the ability to deny genuine evidence by pointing to the existence of deepfakes — may ultimately be more damaging than the deepfakes themselves. It creates an environment where truth is contested not on its merits but on the technical sophistication of the parties involved. Well-resourced actors who can afford deepfake detection services gain an advantage over ordinary citizens who cannot.
What You Can Do
For individuals and organisations, the most practical defence is procedural: establish out-of-band verification protocols for high-stakes communications, treat unexpected video calls requesting sensitive actions with scepticism, and implement multi-factor authentication for financial transactions. The specific threat of "CEO fraud" — where criminals use cloned audio or video of executives to authorise fraudulent transfers — can be largely mitigated by requiring a secondary verification channel (a phone call to a known number, a code word established in advance) for any financial instruction received via video or audio.
For media organisations and individuals publishing content, adopting C2PA-compliant tools for content creation and distribution is worth the investment. It won't prevent deepfakes, but it provides a verifiable provenance chain for genuine content — increasingly valuable as the default assumption shifts toward scepticism.