The scientific community needs to treat Artificial Intelligence as a machine for speculating possible scenarios—because that is exactly what it is. It shouldn't be treated as an optical medium or instrument that retrieves real images or true results.
Technical opacity in AI processes is an issue that must always be brought into the discussion. This is because it often hides the fact that we are no longer observing reality, but rather interrogating a database that already contains the exact answer we want to hear.
I'm not the one saying this—author Fabio Offert argues this in his paper "Latent Deep Space." The text discusses the scientific community's use of Artificial Intelligence to generate or recover images. I found it so fascinating that I decided to bring it here. It takes a little bit of time to explain, so bear with me or save this for later!
Hi, I'm Laura. I hold a Master's degree in Communication and I'm an Internet Systems student.
When a generative AI model—the author specifically names Generative Adversarial Networks, or GANs—is used to reconstruct an astronomical or medical image, it isn't "revealing" a hidden truth buried within the data. It is actually performing an exercise in "controlled hallucination," where physical reality is forced to conform to a space of possibilities or visual rules that have already been sealed by the training dataset.
To illustrate this, the paper highlights astronomy. Historically, recovering a sharp image from noisy data captured by a telescope was a problem solved through physics and geometry. Today, instead of calculating the actual trajectory of light, the neural network fills in the information gaps based on galaxy patterns it has already "seen" before.
If the network encounters a real physical anomaly—something that has never been documented—the generative model tends to smooth out this "strangeness" so it looks like a standard, textbook galaxy. We are trading the uncertainty of observation for the certainty of the algorithm, and as a consequence, we might be losing our very ability to detect the radically new.
Here is where we must locate the theoretical rupture. The classical scientific image, as discussed by authors like Peter Galison, was always a "technical image" with a causal link to the physical object. There was a baseline of mechanical objectivity. An AI image, however, does not "represent" the world through a physical index; it "calculates" the world through probabilistic synthesis.
It lives in a gray area between the mimicry of nature and pure mathematical abstraction. While traditional photography is a record of something that was actually there, a GAN image is a projection of what should be there according to statistics. It usurps the authority of photography to validate what is, at its core, just a visualized mathematical model.
The catch is that "Latent Space" is not the real world; it is a mathematical compression of the world. When a scientist navigates this space, they aren't exploring the physical universe—they are exploring the internal logic of a silicon architecture that prioritizes verisimilitude over actual truth.
The opacity of these processes within AI isn't just technical—in the sense that we don't know what every single neuron is doing—it is, above all, epistemological. Because it ends up assuming a completely different function than it was meant to. It alters how we produce knowledge.
If the tool we use to see the world already "knows" what the world is supposed to look like, science runs the risk of becoming a tautology: a system that merely confirms its own preconceptions. Transparency in data science cannot be achieved simply by making the code open-source; we must recognize that AI is not a lens, but a speculative engine.
Therefore, we need to shift how AI functions in science. It should not be treated as an optical medium that transparently reveals the "truth." Its true utility lies in its capacity to generate hypotheses and simulate scenarios within statistical bounds. We must use it to understand what our data could mean, rather than letting it dictate what reality IS.