The dominant cultural narrative treats AI-generated imagery as a discontinuity, the moment when image evidence stopped being trustworthy. This is wrong in two directions. First, image evidence was never as trustworthy as the narrative assumes; the medium has been manipulable since its first decade. Second, the present moment is genuinely different, but the difference is economic and distributional, not metaphysical. Understanding which historical analogies apply, and which do not, matters because the regulatory and editorial reflexes of the past have a mixed record and the present is borrowing from them.
This page is a brief history of consequential image manipulation, from the wet-plate era to diffusion models. It is selective rather than exhaustive — entire books have been written on individual episodes. What it tries to provide is enough continuity to see what is genuinely new about 2026 and what is just the latest iteration of a long pattern.
The darkroom era: 1840–1990
The first widely documented photographic manipulation predates the medium's first decade. Hippolyte Bayard's "Self Portrait as a Drowned Man," staged in 1840 as a protest against Daguerre's recognition, is in effect a fictional photograph presented as a real one. By the 1860s, composite portraits combining multiple negatives — Henry Peach Robinson's "Fading Away," 1858, was assembled from five separate plates — were exhibited as art, with the manipulation either openly acknowledged or quietly assumed.
Political manipulation followed shortly. The widely reproduced portrait of Abraham Lincoln in heroic pose, circulating from around 1860, is a composite: Lincoln's head on the body of John Calhoun. Stalinist photo retouching, especially the systematic erasure of Nikolai Yezhov from photographs after his 1940 execution, is the textbook 20th-century case. The work was done with airbrushes, scalpels, and bromide; it required skilled labor and access to original negatives. The barrier to entry was real, and it constrained the scale of the practice, but not the seriousness of the cases that did occur.
Newspaper retouching was routine throughout this period — softening complexions, removing blemishes, dropping unwanted bystanders. The 1937 Hindenburg disaster photographs were composited from multiple frames in some published versions. Wartime censors and propagandists in every combatant nation manipulated photographs systematically. The historian David King's 1997 book "The Commissar Vanishes" remains the standard reference on Stalin-era practice and is worth citing because it documents the labor cost: each Yezhov erasure took days.
The Photoshop era: 1990–2014
Photoshop 1.0 shipped in February 1990. The transformation it produced over the following decade was less about new capability — darkroom artists could do most of what Photoshop did, given time — than about removing the labor cost. By the mid-1990s, a college student with a copy of Photoshop and a scanner could produce composites that would have required a professional studio a generation earlier. The newsroom and editorial response was a series of high-profile scandals and the slow construction of explicit ethical codes.
The 1994 Time magazine cover darkening of O.J. Simpson's mugshot, after Newsweek ran the unmodified version, became a teaching case in journalism schools. The 2003 Brian Walski composite from Iraq, in which a Los Angeles Times photographer merged two frames for a more dramatic image, resulted in his firing. The 2006 Adnan Hajj affair, in which a Reuters freelancer cloned smoke in a Beirut photograph, prompted Reuters to terminate its relationship with him and to delete all 920 of his images from its archive.
By the late 2000s, every serious wire service had an explicit policy that prohibited compositing, content removal, or content addition. The cases that drove these policies were not synthetic media in the modern sense; they were retouching scandals. What they established was the institutional muscle that the current provenance debate is now trying to extend with cryptographic tools.
The deep learning prelude: 2014–2022
Ian Goodfellow and collaborators introduced generative adversarial networks in a 2014 paper, "Generative Adversarial Nets," published at NeurIPS. The technique made it possible to train networks that produced novel images statistically similar to a training distribution. Early outputs were low-resolution and obviously artificial. By 2017, NVIDIA's progressive GAN architecture was producing 1024×1024 portraits that could pass a casual inspection.
The term "deepfake" entered public usage in late 2017, originating with a Reddit user who posted face-swapped pornographic videos. Reddit banned the relevant subreddit in February 2018. Through 2018 and 2019, deepfake video became a recurring subject in trade press and policy circles, with much hand-wringing about election impact and very little actual deepfake-driven electoral interference. The threat was theoretical because the technology was hard to use.
StyleGAN, also from NVIDIA, was released in December 2018 and produced the first widely circulated photorealistic synthetic faces. The site thispersondoesnotexist.com, launched in February 2019, made the capability visible to a general audience. Diffusion models — DALL·E (January 2021), GLIDE (December 2021), Stable Diffusion (August 2022), Midjourney (July 2022) — collapsed the cost further. By the end of 2022, text-to-image synthesis was a consumer product.
The diffusion era: 2023–present
The qualitative break came with Midjourney v5 in March 2023 and intensified with the public release of DALL·E 3 inside ChatGPT in October 2023. Generated portraits and scenes crossed the perceptual threshold for non-expert viewers. The pope-in-a-puffer-jacket image, posted to Reddit on 24 March 2023, became the canonical example of a viral synthetic image that fooled large numbers of viewers and journalists. The image was made with Midjourney v5.
Through 2023 and 2024, several long-standing tells progressively disappeared. Hand and finger generation became reliable. Text rendering, previously a near-certain giveaway, became readable. Eye reflections and earring asymmetry stopped being diagnostic. The visual indicators page tracks what remains diagnostic as of 2026, and the list is shorter every year. Detection guides written in 2023 are largely obsolete by 2026.
| Year | Inflection | Effect |
|---|---|---|
| 1840 | Bayard's staged self-portrait | First documented photographic fakery |
| 1990 | Photoshop 1.0 | Manipulation labor cost collapses |
| 2014 | GAN paper (Goodfellow et al.) | Synthesis becomes statistical, not compositional |
| 2018 | StyleGAN | Photorealistic synthetic faces |
| 2022 | Stable Diffusion public release | Open-weights synthesis at consumer scale |
| 2023 | Midjourney v5; DALL·E 3 | Threshold for non-expert deception crossed |
| 2024 | EU AI Act passed | First major statute requiring marking |
| 2025 | Hardware-backed C2PA on flagship phones | Provenance enters consumer capture path |
What is genuinely new
Three things distinguish the present from earlier cycles. First, marginal cost has reached zero. A 19th-century composite required days of skilled labor; a Photoshop edit required hours; a diffusion generation requires seconds and a few cents of compute. The labor constraint that historically limited fakery has dissolved. Second, distribution speed has collapsed in parallel: the Hajj photograph circulated for days before being caught; the pope-in-the-puffer image was a global story within hours.
Third, the targets have democratized. Stalin retouched photographs at the level of the central state; the diffusion era enables anyone to retouch — or fabricate — anyone, including non-public figures. The intimate-image abuse cases that have driven much of the 2024–2026 legislative wave are without obvious historical parallel in their volume or in the absence of any institutional gatekeeper between producer and victim.
What is not new
The premise that photographs are inherently truthful was never accurate; it was a 20th-century convention sustained by the practical difficulty of fakery and the editorial discipline of news organizations. The current discourse sometimes treats the convention as if it had been a property of the medium, which sets up the wrong baseline for what to defend. The realistic target is not a return to a mythical era of trustworthy images, but the construction of a new convention that takes synthesis as default and trustworthy provenance as the marked case. That is what C2PA, SynthID, and the platform policies under development are trying to build.
The other thing not new is the institutional response pattern. After every prior wave — composite portraits, Soviet retouching, Photoshop scandals — newsrooms developed editorial codes, professional societies issued standards, and a small set of authoritative outlets became the de facto verifiers for the rest of the ecosystem. The same is happening with C2PA, Project Origin, and the wire-service C2PA pilots underway since 2023. The technologies are new; the social architecture is being reassembled from older parts.
Where the field is moving
The historical record suggests two predictions, both modest. First, the current panic will subside as institutional and technical defenses settle into routine; the canonical scandals of 2025 and 2026 will become teaching cases in the way that the Adnan Hajj affair became one. Second, the settlement will not restore the previous trust baseline. Future readers will treat images the way present readers treat anonymous quotes: useful but contingent on the publishing institution's reputation. Provenance technology is the infrastructure that lets that reputation be cryptographically attached to the image rather than left implicit in the masthead.