Synthetic Media's Price Collapse & The End Of Visual Proof

Welcome to Memorandum Deep Dives. In this series, we go beyond the headlines to examine the decisions shaping our digital future. 🗞️

This week, we examine how the economics of synthetic media have shifted, starting with a thought experiment economists run when a technology becomes cheap enough to change behavior at scale. They stop asking what the technology does, and start asking what becomes possible once the price falls below the threshold that previously kept most people from trying.

Synthetic media, once expensive and limited to research labs, became accessible to anyone with a consumer-grade gaming PC after Stable Diffusion was released to the public in 2022. By 2026, a deepfake model trained on a specific person can be built in about 15 minutes on consumer hardware, and a wholesale market for synthetic images, cloned voices, and forged identities now operates at prices measured in single dollars.

Each of the visible harms, the intimate-image abuse, the eight-figure corporate fraud, the election-cycle audio clip, gets reported as its own story. They are not separate stories. What ties them together and what each one is doing to the underlying assumptions is where this week's deep dive begins.

Detection is dead, long live verification

There is a useful thought experiment that economists run when a technology becomes cheap enough to change behavior at scale. They stop asking what the technology does, and start asking what becomes possible once the price falls below the threshold that previously kept most people from trying.

In generative AI, a similar story is at play. Synthetic media, or media generated by AI, was once expensive and limited to research labs. All that changed when Stable Diffusion was released to the public in 2022.

As access to this technology became easier, anyone with a consumer-grade gaming PC could produce photorealistic images of people who do not exist, or of real people in situations that never occurred. What changed was not the technology but who could afford to use it.

From research labs to gaming PCs

As of 2026, the cost of generating synthetic media has fallen sharply, and the technical barriers that once constrained the technology have largely disappeared.

One of the clearest examples is a technique known as low-rank adaptation, or LoRA, which allows an existing image-generation model to be trained on a specific person. According to research from the Oxford Internet Institute, creating such a model requires roughly 20 reference images, about 24 gigabytes of GPU memory, and around 15 minutes on consumer hardware. Once trained, the model can generate an effectively unlimited number of synthetic images depicting that individual.

The result was an onslaught of synthetic media on online channels. Between 2017 and 2022, researchers documented just 22 notable deepfake incidents. However, with reduced costs, 179 cases were reported in the first quarter of 2025, according to Resemble AI.

The impact of this increase in synthetic media goes way beyond AI slop taking over the internet. It now undermines a basic assumption: that photographs, video, and recordings can serve as proof of who someone is.

Three harms, one root cause

Historically, the scarcity of the technology imposed a natural limit on misuse since producing convincing synthetic media required expertise, computing power, and effort. Once those constraints weakened, entirely new categories of activity became economically viable. The resulting harms appear different on the surface, but they are all downstream of the same shift in production costs.

On the surface, the harms are broadly categorized along three tracks: non-consensual intimate imagery, financial fraud through synthetic identities and cloned voices, and political manipulation. Though they are often discussed as separate problems, they are better understood as expressions of the same shift.

The first track, non-consensual intimate imagery, is also the best-sourced corner of the phenomenon. Oxford Internet Institute researchers have identified nearly 35k publicly downloadable deepfake model variants, downloaded almost 15M times since late 2022, with 96% targeting women. The ease with which synthetic media can be generated has also led to a shift in who can be targeted. Early deepfakes largely focused on celebrities because source material was abundant. As generation costs fell, the victim pool expanded to ordinary social media users and, increasingly, minors.

The second track is best understood through a single incident. In January 2024, a finance employee in Arup’s Hong Kong office joined a video conference call in which every other participant, including the company’s chief financial officer and several colleagues, was an AI-generated. He had suspected a phishing email, but set the concern aside once he saw familiar faces. He then wired $25.6M to accounts controlled by the fraudsters. The significance of the incident was that the technology was now cheap enough for attackers to use to impersonate individuals for fraud.

The third track, or the political impact, is where the popular narrative and the evidence diverge most sharply. Recorded Future documented 82 pieces of AI-generated deepfake content targeting public figures across 38 countries between July 2023 and July 2024.

Examples range from a fabricated video targeting Moldova’s Maia Sandu to a fake election audio clip in Slovakia and a cloned voice of President Biden used in robocalls aimed at New Hampshire voters.

In such cases, the impact is often misunderstood, as it is assumed the aim is to spread misinformation. But that assumption overlooks how such incidents also erode public trust in what people see and hear.

A review by the Knight First Amendment Institute found that many AI-generated, election-related posts were not meant to deceive. The most influential deepfakes were often technically crude, drawing their power from timing and existing beliefs rather than realism. AI’s more consequential contribution is making authenticity itself contestable, turning disputes over what is real into part of the political message.

This is precisely the point where the three tracks begin to converge. The intimate-image victim, the defrauded employee, and the targeted politician appear to face different threats. In each case, however, the underlying capability is the same: the ability to manufacture convincing evidence at a cost low enough to support mass production.

The extent of this shift can be better understood by examining the emergence of markets for synthetic media. According to Group-IB, deepfake images, cloned voices, and synthetic identities can now be purchased for as little as a few dollars. The numbers suggest that synthetic deception is no longer the domain of hobbyists or isolated actors. It increasingly resembles an industry, complete with supply chains, specialized services, and economies of scale.

And the downstream effects are already appearing in verification systems. Gartner reports that 62% of cybersecurity leaders experienced a deepfake attack during the previous year and predicts that many organizations will soon stop treating standalone biometric verification as reliable. Human judgment is proving equally vulnerable. In iProov testing, only 0.1% of participants correctly identified every real and synthetic image or video they were shown.

An industry built on proving you’re real

The response to the marketization of synthetic media has fueled rapid growth in authentication, provenance, watermarking, and verification technologies. The combined deepfake detection and generation sector was estimated at $857M in 2025 and is projected to exceed $7B by the early 2030s.

Within the industry, the most ambitious efforts focus on proving authenticity rather than detecting deception. Initiatives such as the Coalition for Content Provenance and Authenticity (C2PA) embed cryptographically signed metadata into files, recording who created them, what tools were used, and whether AI was involved. But the modern internet complicates that approach, as content is routinely compressed, edited, and reposted in ways that can strip away metadata.

A similar challenge is emerging in identity verification, prompting institutions to increasingly combine biometrics, device signals, behavioral analysis, and liveness checks rather than relying on any single indicator.

The pattern is familiar: email created spam filters, social media created moderation systems, and synthetic media is creating an authentication economy. The unresolved question is whether verification technologies can evolve faster than generation technologies, or whether society will simply adapt to a world where authenticity must be continuously proven rather than assumed.

When the real becomes harder to believe than the fake

Despite efforts to counter the mass production and dissemination of synthetic media, a deeper fear is now taking hold. The conventional fear was that people would believe fake content, but the greater risk is that authentic content will become easier to dismiss. Legal scholars Robert Chesney and Danielle Citron called it the ‘liar’s dividend.’

Once sophisticated deepfakes become commonplace, denying genuine evidence becomes far easier than proving it authentic. That weakens a core assumption underpinning courts, journalism, and public accountability: that documentation can generally be trusted.

So, while the cost of producing synthetic media continues to collapse, the conversation has shifted away from what problems it solves and toward what people will actually do with the technology. And when cybercriminals have access to such technology, it threatens the very fabric of authentication that human society relies upon. The real cost of synthetic media, then, is much more than the dissemination of misinformation. It is the collapse of the assumption that what you see and hear is real.