This article was originally published on The Conversation.
By 2025, deepfakes have improved dramatically. AI-generated faces, voices, and full-body performances that mimic real humans have improved in quality far beyond what even many experts expected just a few years ago. It is also increasingly being used to deceive people.
In many everyday scenarios, especially low-resolution video calls and media shared on social media platforms, the level of realism is high enough to reliably fool non-expert viewers. In fact, synthetic media has become indistinguishable from real recordings to the public and, in some cases, institutions.
And this surge is not limited to quality. The amount of deepfakes is exploding. Cybersecurity firm DeepStrike predicts that online deepfakes will increase from about 500,000 in 2023 to about 8 million by 2025, an annual growth rate of nearly 900%.
I'm a computer scientist who studies deepfakes and other synthetic media. From my perspective, the situation is likely to get even worse in 2026 as deepfakes become synthetic performers that can react to people in real time.
dramatic improvement
Several technological changes underlie this dramatic escalation. First, there have been significant advances in video realism, thanks to video generation models specifically designed to maintain temporal consistency. These models produce videos with consistent movement, consistent identities of the people depicted, and meaningful content from frame to frame. The model separates information related to the representation of a person's identity from information about the behavior, so the same behavior can be mapped to different identities, and the same identity can have multiple types of behavior.
These models produce stable, consistent faces without flickering, distortion, or structural distortions of the eyes or jawline that once served as reliable forensic evidence for deepfakes.
Second, voice cloning exceeds what I call the “indistinguishability threshold.” A few seconds of audio is enough to produce convincing clones with natural intonation, rhythm, emphasis, emotion, pauses, and breathing noises. This feature is already facilitating large-scale fraud. Some major retailers report receiving more than 1,000 AI-generated fraudulent calls per day. Perceptually, we can see that the synthesized speech that was once available has all but disappeared.
Third, consumer tools have pushed technological barriers to near zero. Upgrades from OpenAI's Sora 2 and Google's Veo 3, as well as a wave of startups, have made it possible for anyone to write an idea, have it drafted into a large language model like OpenAI's ChatGPT or Google's Gemini, and produce sophisticated audiovisual media in minutes. AI agents can automate the entire process. The ability to generate consistent, story-driven deepfakes at scale has effectively been democratized.
The combination of this burgeoning volume and personas that are nearly indistinguishable from real humans creates serious challenges for detecting deepfakes, especially in a media environment where people's attention is fragmented and content moves faster than it can be verified. Deepfakes, which spread before people even knew what was going on, are already causing real harm, from misinformation to targeted harassment and financial fraud.
The future is real time
Looking ahead, next year's trajectory is clear. Deepfakes are moving toward real-time synthesis that can produce videos that closely resemble the nuances of human appearance, making it easier to evade detection systems. Frontiers are moving away from static visual realism to models that produce temporal and behavioral consistency, i.e. live or near-live content rather than pre-rendered clips.
Identity modeling is converging into integrated systems that capture not just what people look like, but how they move, sound, and speak across contexts. The result goes beyond “this looks like person X” to “this behaves like person X over time.” I would expect the entire video call participants to be composited in real time. Interactive AI-driven actors whose faces, voices, and mannerisms instantly adapt to your prompts. Some scammers deploy responsive avatars instead of static videos.
As these capabilities mature, the perceptual gap between synthetic and real human media will continue to shrink. A meaningful line of defense moves away from human judgment. Instead, it relies on infrastructure-level protection. These include secure provenance, such as cryptographically signed media, and AI content tools that use the Coalition for Content Provenance and Authenticity specifications. It also relies on multimodal forensic tools such as my lab's Deepfake-o-Meter.
It's no longer enough to just look closely at pixels.
Siwei Lyu, Professor of Computer Science and Engineering. UB Media Forensic Lab Director, university at buffalo
This article is republished from The Conversation under a Creative Commons license. Read the original article.
