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 expected would be the case just a few years ago. They were also increasingly used to deceive people.<\/p>\n
For many everyday scenarios \u2014 especially low-resolution video calls and media shared on social media platforms \u2014 their realism is now high enough to reliably fool nonexpert viewers. <\/p>\n
In practical terms, synthetic media have become indistinguishable from authentic recordings for ordinary people and, in some cases, even for institutions.<\/p>\n
And this surge is not limited to quality. The volume of deepfakes has grown explosively: Cybersecurity firm\u00a0DeepStrike\u00a0estimates an increase from roughly 500,000 online deepfakes in 2023 to about 8 million in 2025, with annual growth nearing 900%.<\/p>\n
I\u2019m a computer scientist who\u00a0researches deepfakes\u00a0and other synthetic media. From my vantage point, I see that the situation is\u00a0likely to get worse\u00a0in 2026 as deepfakes become synthetic performers capable of reacting to people in real time.<\/p>\n
Dramatic improvements<\/p>\n
Several technical shifts underlie this dramatic escalation. First, video realism made a significant leap thanks to video generation models designed specifically to\u00a0maintain temporal consistency. <\/p>\n
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. <\/p>\n
The models disentangle the information related to representing a person\u2019s identity from the information about motion so that the same motion can be\u00a0mapped to different identities, or the same identity can have multiple types of motions.<\/p>\n
These models produce stable, coherent faces without the flicker, warping or structural distortions around the eyes and jawline that once served as reliable forensic evidence of deepfakes.<\/p>\n
Second, voice cloning has crossed what I would call the \u201cindistinguishable threshold.\u201d <\/p>\n
A few seconds of audio now suffice to generate a\u00a0convincing clone\u00a0\u2013 complete with natural intonation, rhythm, emphasis, emotion, pauses and breathing noise. <\/p>\n
This capability is already fueling large-scale fraud. Some major retailers report receiving\u00a0over 1,000 AI-generated scam calls\u00a0per day. The perceptual tells that once gave away synthetic voices have largely disappeared.<\/p>\n
Third, consumer tools have pushed the technical barrier almost to zero. <\/p>\n
Upgrades from OpenAI\u2019s\u00a0Sora 2\u00a0and Google\u2019s\u00a0Veo 3\u00a0and a wave of startups mean that anyone can describe an idea, let a large language model such as OpenAI\u2019s ChatGPT or Google\u2019s Gemini draft a script, and\u00a0generate polished audio-visual media in minutes. <\/p>\n
AI agents can automate the entire process. The capacity to generate coherent, storyline-driven deepfakes at a large scale has effectively been democratized.<\/p>\n
This combination of surging quantity and personas that are nearly indistinguishable from real humans creates serious\u00a0challenges for detecting deepfakes, especially in a media environment where people\u2019s attention is fragmented and content moves faster than it can be verified. <\/p>\n
There has already been real-world harm \u2013 from\u00a0misinformation\u00a0to\u00a0targeted harassment\u00a0and\u00a0financial scams\u00a0\u2013 enabled by deepfakes that spread before people have a chance to realize what\u2019s happening.<\/p>\n
The future is real time<\/p>\n
Looking forward, the trajectory for next year is clear: Deepfakes are moving toward real-time synthesis that can produce videos that closely resemble the nuances of a human\u2019s appearance, making it easier for them to evade detection systems. <\/p>\n
The frontier is shifting from static visual realism to temporal and behavioral coherence: models that\u00a0generate live or near-live content\u00a0rather than pre-rendered clips.<\/p>\n
Identity modeling is converging into unified systems that capture not just how a person looks, but how they\u00a0move, sound and speak across contexts. <\/p>\n
The result goes beyond \u201cthis resembles person X,\u201d to \u201cthis behaves like person X over time.\u201d <\/p>\n
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.<\/p>\n
As these capabilities mature, the perceptual gap between synthetic and authentic human media will continue to narrow. The meaningful line of defense will shift away from human judgment. Instead, it will depend on infrastructure-level protections. <\/p>\n
These include secure provenance such as media signed cryptographically, and AI content tools that use the\u00a0Coalition for Content Provenance and Authenticity\u00a0specifications. It will also depend on multimodal forensic tools such as my lab\u2019s\u00a0Deepfake-o-Meter.<\/p>\n
Simply looking harder at pixels will no longer be adequate.<\/p>\n
Editor\u2019s Note: This article was written by Siwei Lyu,\u00a0Professor of Computer Science and Engineering, Director, UB Media Forensic Lab, University at Buffalo, and republished from\u00a0The Conversation\u00a0under a Creative Commons license. Read the\u00a0original article.<\/p>\n","protected":false},"excerpt":{"rendered":"