What is a ‘real’ picture? It’s hard to imagine major corporations weighing in on that one, but that really was the question asked and answered by a Samsung executive in a recent interview. When questioned about last year’s accusations that its smartphones weren’t actually capturing the moon, but overlaying a filter to make it look like they were, Patrick Chomet said: “Everyone was like, ‘Is it fake? Is it not fake?’ There was a debate around what constitutes a real picture. And actually, there is no such thing as a real picture. As soon as you have sensors to capture something…there is no real picture.”
For those who don’t remember, last year a viral post argued that the company’s Space Zoom camera feature basically amounted to making up pictures of the moon. It said that the process that transformed a blurry picture into a high resolution one involved the use of “a specific AI model trained on a set of moon images, in order to recognize the moon and slap on the moon texture on it.” There was a difference, they reckoned, between combining multiple frames to recover detail and creating a new image from scratch.
Samsung has since released a blog in which the company explained that “learning data” (i.e. pictures of the moon) are in fact used to power a “detail enhancement engine”. The upshot is that yes, a generative model is being used to create pictures that may not precisely represent a detailed version of a picture of the moon taken with a Samsung phone.
But that shouldn’t come as a surprise. When we use AI to “enhance” a picture we are asking it to make a best guess as to what it thinks that a higher resolution or more detailed version of a particular image may look like. That is not the same as accurately capturing what a higher resolution of a given image would actually look like. (Of course, there may be instances in which the prediction maps very neatly onto reality—but often we don’t always have the ability to check.)
Take Tomb Raider’s Lara Croft. This was a popular example that circulated on X a few weeks ago that showed a picture of the original game rendering from the first Playstation console ‘enhanced’ using Magnific. The version given the AI treatment looks more or less photorealistic at first glance, but if you look a little closer you can see the backpack’s shoulder strap is no longer fastened under the arm (not ideal for raiding tombs).
I don’t think this error is particularly significant as far as the quality of the image is concerned. And I’m not interested in whether the example proves that the model doesn’t have a world model or anything like that. What I am interested in, though, is what it says about the type of instrument that generative AI is. That it generated a cool-looking but slightly inconsistent image shows us that it’s a tool that we could probably have a lot of fun with, but it should also encourage us to question the extent to which we should see it as a manifestation of precision or imagination. That is to say: generative models don’t enhance, they create.
Now, sceptics would rightly say there is no ‘real version’ of what Lara Croft looks like. It’s a game, buddy, so it doesn’t really matter one way or the other. That may well be true in this case (though even in virtual reality fidelity and permanence are both desirable in lots of circumstances) but it doesn’t really hold at the point at which we are trying to capture the real world. That takes us back to the Samsung defence, which is that there is no such thing as a real picture anyway because the equipment we use distorts the human sensory experience. You could go a layer deeper and say that the human sensory experience doesn’t really represent reality, but I’ll save that one for another time.
Truth to nature
In the case of upscaling pictures of the moon, the question boils down to how we want it to represent the world. Should examples of the natural world represent the ideal or the typical? This tension has sat at the heart of scientific practice since the early 1800s, when the term ‘objectivity’ was first introduced into the scientific lexicon in a way that we understand it today.
Historians of science recognise three types of objectivity. First, there are attempts to describe regularity by overlooking or omitting inconsistency. This is when an observer might sketch a specimen by removing irregular features to create an archetypical version of the artefact of study that represents its essence. We call this type of objectivity ‘truth to nature’. Next, we have so-called ‘mechanical objectivity’ that seeks to capture unique details and describe or reproduce variation. Historically, its introduction was characterised by attempts to automate the observational process to defend against what boosters described as the distorting effect of human perception. And finally, there is ‘structured objectivity’ that amounts to the use of expertise to represent the world, often through the use of data and statistics.
Daston and Galison, on whose excellent Objectivity much of this post is based, argue that epistemic virtues do not replace each other “like a succession of kings”. What they mean by that is that that truth to nature, mechanical objectivity, and structured objectivity each continue to inform scientific practice in their own way long after the introduction of a new way of seeing.
The age of AI has implications for all three types of scientific objectivity, but I’m going to spend the rest of this piece focusing on the relationship between truth to nature and mechanical objectivity. To illustrate the difference between the two, we begin with Swedish biologist Carl Linnæus. Best known for efforts to formalise binomial nomenclature and trips to Lapland, Linnæus aggressively selected which plants to discard and which to include throughout his career. These choices were especially pronounced within Hortus Cliffortianus, a visual encyclopaedia detailing the rich assortment of plants found in the garden of Amsterdam banker (and director of the Dutch East India Company) George Clifford. In it, Linnæus carefully guided the illustrator Georg Dionysius Ehret to create sketches of plants that he believed best represented their true form.
The rub, of course, was that creating an idealised version of an entire species of plant necessarily involves removing any elements that may be described as non-typical. The result is an image that may be true to nature in some sense, but is not true to how a particular species is likely to be encountered in the wild. There are plenty of other examples that explain how truth to nature manifested itself in the 19th century—from anatomist William Hunter’s efforts to retouch human corpses after death to naturalist George Edwards’ decision to pose birds in “as many styles” as he could invent in displays—but in each instance the motivation was to curate examples that scientists thought best represented the natural world.
But truth to nature wouldn’t last forever. By the early 20th century, scientists began to wonder whether the representation of so-called working objects (specific examples that stood in for a broader range of natural objects) were contrived manifestations of wishful thinking. German physiologist Otto Funke, for example, described the process by which he painstakingly copied images of blood seen through a microscope, which sat in contrast to what he saw as the idealised crystallographic diagrams that came before. Another example that draws into focus the differences between the two schools came in 1892 when meteorologist Gustav Hellmann captured close-up images of snowflakes in all of their irregular beauty. The work of Hellmann stood in stark contrast to the 1820 effort of Arctic explorer William Scoresby that aimed to describe snowflakes as perfectly symmetrical.

But mechanical objectivity isn’t foolproof. A famous episode that demonstrates its shortcomings came in 1905 when American businessman, mathematician, and astronomer Percival Lowell attempted to draw what he thought were canals on the martian surface. Lowell created the sketches by looking through his telescope in regularly defined intervals for around seven months. Sketches accomplished, he turned his attention to the introduction of new equipment to capture photographs of the canals. But the photos were poor quality. When readers couldn’t see the canals, Lowell questioned whether he ought to alter the images to highlight the presence of the canals that his sketches made clear—but relented after deciding that how the images were captured was as important as what the images captured.
Telescope or paintbrush?
Upscaling models are artistic tools, not scientific instruments. That is the core of the problem at the heart of the Samsung episode, which drew into focus the tension between two very different types of seeing. For those hoping for an accurate representation of the moon, on first blush Space Zoom seems to represent a move away from mechanical objectivity towards approaches that centre truth to nature. All those blurry pictures of the moon are gone in favour of pristine shots that represent the essence of Earth’s satellite.
But that’s not quite right. Clearly, professional astronomers would not use generative models to upscale a picture of the moon any more than a biologist would use a generative AI tool to enhance pictures of the human body. That’s because generative models don’t really enhance or upscale, they create something entirely new.
AI generated pictures of the moon are not demonstrations of truth to nature because they separate seeing from knowing. When Linnæus asked Ehret to omit or alter some features of plants in his drawings, he was doing so on the basis that he believed his expertise allowed him to curate idealised images that represented their own form of objectivity.
Image generation models clearly cannot exercise this type of judgement, which means that—idealised or not—AI enhanced pictures of the moon don’t represent the natural world. As for whether we’re dealing with a version of mechanical objectivity, plastering a new image on top of an existing one bears no resemblance to the careful work of scientists to sketch snowflakes or blood as seen from a microscope. Mechanical it may be, but objective it is not.
That leaves us in an interesting place. Generative AI tools are incapable of dealing in both mechanical objectivity and truth to nature, an issue that I think goes some way to explaining the backlash to their use for capturing the night sky. Amateur astronomers wanted a telescope, but they got a paintbrush instead.