When a colleague shares a suspiciously polished photograph and asks "is this AI?", the instinct is to reach for the newest detection tool. But the question has roots that stretch back further than Midjourney, further than DALL-E, and further than the GAN portraits that startled the art world in 2018. A clearer sense of that history helps us understand what synthetic images actually are, and why the trust problem they pose to journalism is structural rather than incidental.

The earliest machine-made images

The story of machine-generated imagery does not begin with neural networks. Harold Cohen, a British artist and computer scientist, began developing a program called AARON in the late 1960s and continued refining it through the 1970s and beyond. AARON used rule-based logic to produce original drawings and paintings autonomously. Cohen never claimed the machine was conscious or creative in a human sense, but the output was undeniably original visual content produced without direct human authorship of each mark. AARON is now widely cited by historians of digital art as the first sustained example of a computer system functioning as an image-maker rather than merely a rendering tool.

Through the 1980s and 1990s, fractal generation, ray-tracing, and procedural texture synthesis expanded what computers could produce visually. These were largely deterministic systems: given the same inputs, they returned the same outputs. They were generative in the broad sense, but they lacked the capacity to surprise their operators in the way later probabilistic models would.

The GAN breakthrough and its consequences

The shift that most directly shapes our current situation came in 2014, when Ian Goodfellow and colleagues introduced generative adversarial networks, or GANs, in a paper that has since become one of the most cited in machine learning. The architecture pits two neural networks against each other: a generator that produces images and a discriminator that tries to identify them as fake. As the two networks train together, the generator becomes progressively better at producing images that fool the discriminator, and by extension, human observers.

The cultural moment that brought GANs to mainstream attention came in 2018, when a GAN-produced portrait titled "Edmond de Belamy" sold at Christie's auction house. The work, created by the Paris-based collective Obvious using a GAN trained on historical portraits, fetched $432,500, far exceeding its pre-sale estimate. The sale prompted widespread debate about authorship, authenticity, and the market value of machine creativity. For newsrooms, it was also an early signal that synthetic imagery could carry institutional weight and deceive audiences accustomed to treating visual art as evidence of human intention.

GANs produced the first wave of photorealistic synthetic faces that journalism had to grapple with seriously. The website "This Person Does Not Exist," launched in 2019 and built on Nvidia's StyleGAN architecture, demonstrated to a general audience that a convincing human portrait could be generated on demand, at no cost, in seconds. Profile pictures, byline photos, and source images all became newly suspect. Our field guide to synthetic media for newsrooms examines the full range of content types this shift affected.

Diffusion models and the present landscape

The dominant architecture today is not the GAN but the diffusion model. Systems such as Stable Diffusion, Midjourney, and DALL-E 3 work by learning to reverse a process of adding noise to images. They are trained on vast datasets of image-text pairs, which gives them the ability to respond to natural-language prompts with striking fidelity. The outputs are harder to detect than GAN images, partly because they do not carry the characteristic artefacts that earlier detection tools were trained to identify.

For verification desks, this evolution has real operational weight. The technical fingerprint of a synthetic image changes with each new model generation, which means detection approaches that rely on artefact-spotting are perpetually playing catch-up. The more durable response is provenance: knowing where an image came from and what touched it before it reached a newsroom. That is the logic behind content credentials and the C2PA standard, which attaches cryptographically signed metadata to media files at the point of capture or generation.

Why the history matters for trust

Understanding this arc changes how we frame the problem internally. Synthetic imagery is not a sudden rupture caused by one company releasing one product. It is the latest development in a decades-long progression of tools that separate visual output from direct human observation of a scene. Each step in that progression, from AARON to GANs to diffusion models, expanded both the creative possibilities and the verification burden.

The news industry's existing frameworks for image verification, source confirmation, caption accuracy, and contextual integrity were built in an era when photographic images had a presumptive evidential relationship to reality. That presumption is now contested at scale. Policies built around it need revisiting, and the revisiting should be informed by a realistic account of how we arrived here.

Two concrete areas deserve immediate attention. First, the difference between content credentials and watermarking is not merely technical: it determines what claims a newsroom can actually make about an image's origin. Second, tracking which platforms have adopted content credentials tells us where provenance signals can currently be trusted and where they cannot. The longer history of content authentication shows that the industry has navigated provenance problems before, though never at this speed or scale.

The first AI-generated image is less a single object than a moving category. What matters for journalism is not pinning down an origin date but recognising that the category now encompasses content indistinguishable from photography, produced at negligible cost, and circulating without reliable provenance signals in most of the channels our audiences use every day.

Sources

  • Harold Cohen and the AARON program: Computer History Museum documentation and Cohen's own writings, widely cited in digital art histories.
  • Goodfellow, I. et al., "Generative Adversarial Nets," Advances in Neural Information Processing Systems, 2014.
  • Christie's auction result for "Edmond de Belamy," October 2018, Christie's official sale records.
  • Nvidia StyleGAN and the "This Person Does Not Exist" demonstration, 2019, covered extensively in technology press including MIT Technology Review.
  • C2PA (Coalition for Content Provenance and Authenticity) technical specification, c2pa.org.