One of the most insightful and disturbing works by Fyodor Dostoevsky is *The Double*, where the protagonist faces the appearance of an identical doppelgänger, but radically different in character. This identity play serves as a perfect metaphor to understand the current phenomenon of generative photography: an image that, while resembling traditional photography, is essentially something completely different. For decades, we have coexisted with images that imitate photography without being it. Photography displaced painting as the main medium of realistic representation, but manipulation and fiction soon infiltrated its territory. An early example can be found on the cover of *L'Illustration* from 1891, where a supposed photograph was actually a collage of cutouts and drawings, anticipating the digital montages that circulate on social media today. On that cover, a cut-out woman overlaps an illustrated background, with details like drawn shadows to achieve visual coherence. The low resolution of prints at the time made it difficult to detect the deception, similar to what happens today with low-quality fake images circulating on WhatsApp. Creating these collages required technical skills and resources that were not easily accessible. What has revolutionized artificial intelligence is the speed and accessibility: now, anyone can generate images in seconds, although a deep knowledge is still necessary to achieve convincing results. Director Carlo Padial compared the emergence of TikTok to the fast zombies of modern cinema: the acceleration of audiovisual production is dizzying. Platforms like TikTok and tools like CapCut, with generative AI features, have multiplied the creation of images without real references, generated from simple text instructions. However, the apparent ease of these tools is also their Achilles' heel. The overestimation of AI leads to mediocre results and obvious errors. Therefore, it is essential to analyze how photorealistic images can be created that challenge even the most advanced detection systems. An illustrative case: in 2014, all the issues of the experimental magazine *Lalata* were photographed. For its 25th anniversary, it was decided to update the image by adding the issues published afterward. Repeating the session was unfeasible, so a digital montage was chosen. Accurately describing each issue in a prompt proved useless; the solution was to combine real photographs of the new issues with the original image, using Photoshop and its generative AI to adjust colors and textures. The success of these montages depends not only on image generation but also on the integration of tools like Content-Aware, which has allowed for realistic removal or addition of elements for years. Today, Photoshop incorporates generative AI features that, at times, are less convincing than traditional techniques. The key lies in overlaying methods and leveraging the best of each. Thus, a family image of *Lalata* was achieved that was so credible that even Vericta, a Spanish detector of AI-generated images, could not identify the manipulation. This leads us to the concept of prompting: generating realistic images is not just a matter of asking AI for something, but understanding and simulating the real photographic process, specifying details like optics, light, or composition. A good prompt describes the capture conditions, not just the scene. For example, when recreating a photograph of an abandoned gas station, technical data from the Vivo X300 Pro phone was included, such as aperture, focal length, and shutter speed. AI does not understand reality, but it does recognize patterns associated with these parameters. Imperfection is another crucial factor: the most credible images tend to be the least perfect. Adding blurs, crooked frames, or cut-off elements reduces the likelihood of AI overly beautifying the scene. The more one seeks aesthetic perfection, the easier it is to detect artifacts or anomalies. This connects with the limits of detection systems. As Georgina Viaplana, CEO of Lawwwing and head of Vericta, explains, the system analyzes the image globally and may overlook small modified areas. Granular analysis increases the risk of false positives, making hybrid scenarios—where only part of the image is generated by AI—especially difficult to detect. In practice, these technologies are used to detect fraud in insurance, e-commerce, expense management, or fact-checking, with accuracy rates above 95%. The secret lies not in the easily manipulable metadata, but in the invisible patterns that AI leaves at the pixel level. In short, the real challenge is not the existence of AI-generated images, but their subtle integration into traditional workflows, increasingly blurring the line between the real and the artificial. As in *The Double*, the problem arises when the original and the copy become confused. Today, a discreet intervention is enough to erase that line, and the illusion of accessibility that these tools offer recalls the arrival of the first digital cameras: it seemed that taking good photos was easy, but the reality is much more complex. Source: lavanguardia.com