Deepfakes have evolved from crude digital novelties into sophisticated audio-visual forgeries that can convincingly impersonate real people, raising urgent questions about trust and authenticity in an age of generative AI. What began as a niche technical challenge a decade ago has become a mainstream concern as the convergence of multimodal artificial intelligence, massive datasets, and open-source tools has democratized the ability to create convincing fake videos and audio.

What Exactly Is a Deepfake?

A deepfake is any image, video, or audio recording, or combination thereof, that has been digitally manipulated to depict something that never actually happened. In the early days, before generative AI made manipulation easier, the results were often unconvincing. People labored to coordinate video and audio, and glitches frequently gave away the forgery. The process required significant manual work to smooth out imperfections and create seamless results.

The technology drew serious attention from intelligence agencies and law enforcement as early as 2017, when researchers at the University of Washington created a deepfake of President Obama. The concern was clear: bad actors could use deepfakes to convince populations that world leaders said things they never said or did things they never did.

When Did Deepfakes Become Truly Convincing?

The real game-changer arrived in 2021 with the work of Christopher Ume, a visual effects artist who created hyper-realistic deepfake videos of actor Tom Cruise. Ume's breakthrough demonstrated that you didn't need a PhD in artificial intelligence to create convincing deepfakes. He hired a Tom Cruise impersonator to provide base footage, then meticulously refined it to create videos that made viewers doubt their own eyes. Crucially, Ume shared his methods with the online community of practitioners, and his software was open source, which accelerated adoption across the field.

The next major acceleration came in 2023 when OpenAI and other organizations began compiling massive datasets to train large language models, most famously ChatGPT. By 2024 and 2025, multimodal AI, which processes both audio and visual information together, had gone mainstream. What once required weeks of painstaking labor could now be accomplished with minimal effort using newer generative models.

How the Technology Evolved Across Key Milestones

• Early Era (Pre-2021): Deepfakes were clunky, required extensive manual labor to coordinate audio and video, and glitches frequently exposed them as fake. Apps like FaceApp existed but produced obviously synthetic results.
• The Ume Breakthrough (2021): Christopher Ume's Tom Cruise deepfakes proved that highly convincing forgeries were possible without advanced AI degrees. His open-source approach democratized the technique and inspired widespread adoption.
• Generative AI Revolution (2023-2025): Large language models trained on massive datasets enabled multimodal AI systems that could generate convincing deepfakes with minimal human effort, shifting the technology from a niche skill to an accessible tool.

The implications are profound. Unlike earlier eras when deepfakes were curiosities, the current generation of multimodal AI systems can produce audio-visual forgeries that are difficult to distinguish from authentic recordings. This shift from novelty to credible threat has captured the attention of technologists, policymakers, and society at large.

Why Understanding Deepfakes Matters Now

The evolution of deepfake technology reflects a broader pattern in how transformative technologies emerge. They rarely arrive fully formed. Instead, they begin with basic research in university and industry labs, are refined over time, and sometimes fail to materialize. For a technology to become truly disruptive, multiple conditions must align: the underlying technology must be perfected, the necessary hardware must exist to run it quickly and affordably, and societal conditions must create demand for it.

"With Ume's work, we collectively jumped from deep fakes being an amusing novelty to being highly convincing, and with that change, the floodgates opened," noted Alma Katsu, a former tech futurist for the intelligence community.

Alma Katsu, Author and Former Intelligence Community Tech Futurist

Katsu, who wrote her first article on deepfakes over a decade ago and her last comprehensive piece on the topic in 2022, has observed this evolution firsthand. That 2022 article, which documented the technology's rapid advancement, inspired her to write "Incarnate," a novel exploring how deepfakes and multimodal AI could fundamentally alter society. The book's protagonist uses software to create a perfect digital avatar, raising questions about authenticity, identity, and the ethical boundaries of synthetic media.

The broader concern is not merely technical but deeply human. As deepfakes become indistinguishable from reality, society faces questions about trust, verification, and the nature of evidence itself. The technology has the potential to upend life as we know it, which is why understanding its trajectory from clunky experiment to convincing deception matters for everyone.