Perspective

By Siwei Lyu 

Over the course of 2025, deepfakes improved dramatically. AI-generated faces, voices and full-body performances that mimic real people increased in quality far beyond what even many experts thought would be the case just a few years ago. They were also increasingly used to deceive people.

For many everyday scenarios—especially low-resolution video calls and media shared on social media platforms—their realism is now high enough to reliably fool nonexperts, including not only ordinary people but even some institutions.

And this surge is not limited to quality. The volume of deepfakes has grown explosively: Cybersecurity firm DeepStrike estimates an increase from roughly 500,000 online deepfakes in 2023 to about 8 million in 2025, with annual growth nearing 900%.

I’m a computer scientist who researches deepfakes and other synthetic media. From my vantage point, I see that the situation is likely to get worse in 2026 as deepfakes become synthetic performers capable of reacting to people in real time.

Several technical shifts underlie this dramatic escalation. First, video realism made a significant leap thanks to video generation models designed specifically to maintain temporal consistency. These models produce videos that have coherent motion, consistent identities of the people portrayed, and content that makes sense from one frame to the next, producing stable, coherent faces without the flicker, warping or structural distortions around the eyes and jawline that once served as reliable forensic evidence of deepfakes.

Second, voice cloning has crossed what I would call the “indistinguishable threshold.” A few seconds of audio now suffice to generate a convincing clone—with natural intonation, rhythm, emphasis, emotion, pauses and breathing noise. This capability is already fueling large-scale fraud. Some major retailers report receiving over 1,000 AI-generated scam calls per day.

Third, consumer tools have pushed the technical barrier almost to zero. Upgrades from OpenAI’s Sora 2 and Google’s Veo 3 and a wave of startups mean that anyone can describe an idea, let a large language model such as OpenAI’s ChatGPT draft a script, and generate polished audio-visual media in minutes.

This combination of surging quantity and personas that are nearly indistinguishable from real humans creates serious challenges for detecting deepfakes. There has already been real-world harm—from misinformation to targeted harassment and financial scams.

Looking forward, the trajectory for next year is clear: Identity modeling is converging into unified systems that capture not just how a person looks, but how they move, sound and speak across contexts. The result goes beyond “this resembles person X” to “this behaves like person X over time.” I expect entire video-call participants to be synthesized in real time; interactive AI-driven actors whose faces, voices and mannerisms adapt instantly to a prompt; and scammers deploying responsive avatars rather than fixed videos.

As these capabilities mature, the meaningful line of defense will shift away from human judgment. Instead, it will depend on infra- structure-level protections. These include secure provenance such as media signed cryptographically, and AI content tools that use the Coalition for Content Provenance and Authenticity specifications. It will also depend on multimodal forensic tools such as my lab’s Deepfake-o-Meter, currently the only open-source free platform for ordinary users to test content.

Simply looking harder at pixels will no longer be adequate. 

Adapted from an article published in The Conversation.

ABOUT THE AUTHOR: Siwei Lyu, a SUNY Distinguished Professor, is SUNY Empire Innovation Professor in the Department of Computer Science and Engineering, director of the Institute for Artificial Intelligence and Data Science, director of the UB Media Forensic Lab and co-director of the Center for Information Integrity.