Images are how we perceive the world. They are important.
In this post I will talk about the concept of operational images, how the use of these images has risen and shifted, and why that is important in light of recent events such as the Anthropic/Department of War dispute, the war in Ukraine, or the rise of slop on social media. I hypothesize that these images allow systems to optimize for a single metric. By creating an optimization loop in a system, humans are far from excluded from it and are in fact often part of the metric as the systems improve. However, they lose legibility and understanding of a system that acts on their reality.
Historical context
As early as WWII, cameras were attached to bombs. At that time the images could not be recorded as they were filmed, the footage we have is someone filming a TV screen. Cameras attached to weapons however gained much more widespread use and diffusion to the public during the Gulf War in 1991
The military used cameras to monitor and control bombing attacks from a safe distance. Cameras were also used to monitor assembly lines away from the noise and heat. If these images contained beauty, it was not intentional.
Harun Farocki defines operational images as images that serve a functional rather than esthetic purpose. I would explicitly add[1] they do not describe the world but act on it by being used in a loop with machines or humans to optimize for a particular metric. The classic example of operational images are videos of an aerial view with a crosshair in its center. There is a building that is highlighted. Then a flash, then the building is gone.
It is important to note that although they can be used as such later, these images are not propaganda: they serve to accomplish a military goal.
It may seem silly to give a definition to images that fit this definition today, as they are very common and therefore don't feel special. The rest of this post will try to explain why the distinction matters, starting with the fact that despite operational images being common they are very recent, especially at the scale they exist today.
Although operational images may have existed before (I think some kinds of maps fit their definition), they became much more common and seen by many more people during the Gulf War period, as cameras allowed for the first time the creation of many of these images, much faster (or even in real time). Maps didn't have to be drawn by hand, they could be autonomously created. Humans didn't have to plot paths in the map, an algorithm did. Humans didn't even have to travel that path, a missile did.
The idea of "operational image" seems relevant to present events.
These historical events seem relevant to events happening recently:
AI art slop: The speed at which we can produce images has only risen since 1991, and this led to new operational images.
Some accounts on social media create shocking viral content with AI. They seem to have no other motive than make something viral, their images addictive but meaningless to humans. Since they are optimizing for a single quantifiable metric with these images, they are some kind of operational images. However there is some kind of new element to it: Unlike war which has metrics decided by humans, the metric these images optimize for is the satisfaction of an algorithm which itself was not directly designed by humans: engagement and watch time emerged as metrics pushed by capitalism, then the mapping from content to predicted engagement score was learnt by an opaque system ("the algorithm") rather than designed by hand. We are seeing images made by machines for machines. Humans are out of the loop.
I am reminded of Goodhart's law: When a measure becomes a target, it ceases to be a good measure.
"At the same time, capitalism still depends on meaning systems to maintain a degree of order. Accordingly, Deleuze and Guattari (1983: 34) claim that ‘capitalism institutes or restores all sorts of residual and artificial’ meanings and beliefs in order ‘to rechannel persons who have been defined in terms of abstract quantities’. Producing functional subjectivities requires resuscitating all kinds of archaic and antiquated ideologies: states and their ideals of citizenship, nationalist ideologies, family values, religious beliefs and folkloric traditionalism.2 Given its denaturalizing and (re)combinatory aims, the ideological landscape of capitalism can be seen as ‘a motley painting of all that has been believed’ (Deleuze and Guattari, 1983); capitalism aims to freely combine existing signs and traditions to produce the affective and aesthetic contexts it deems necessary for its perpetuation. Arguably, the operational, decontextualizing logic of capitalism finds its most faithful aesthetic mirror in AI slop, which likewise describes a process of aesthetic appropriation and computationally optimized reassembly and circulation." - Jernej Markelj and Jakko Kemper, Doing the numbers: the operational aesthetics of AI slop
I like this quote because it helped me understand why, despite current image models being quite capable, we do not see any good art made with them: Their nature attracts creator that are not interested in making art. I would guess the same phenomenon is what also repulses artists from using them.
As an aside Claude told me this was the most falsifiable claim of my argument. Is it ? If you happen to really like some AI art, please share it with me. I am obviously biased as an artist but I genuinely haven't seen any AI art that I liked for its artistic qualities. It would make sense there is some out there, but compared to the quantity of it being produced, I think the point still stands.
That later part, AI reviewing data produced by machines in a decision/action loop, seems to fit perfectly this kind of operational image shift from military to social that I try to point at in this post.
On the autonomous weapons part: the AI safety community goes through a lot of effort to emphasize the AIs of the future may be fundamentally different from the AIs we have now. In that sense, AI has controlled weapons autonomously for a long time and the argument for saying LLMs would be more dangerous seems weaker to me after writing that post: Not only have narrow AI systems controlled weapons for a long time, but they have been part of a non human loop too.
The Ukraine and Russia war, which saw the advent of drones.
Soldiers talk about a "Kill zone" surrounding the frontlines. In this 20-30km deep zone, very cheap drones can create huge damage to humans. Therefore only machines can penetrate.
Their performance is evaluated partly by the drone footage, these images are therefore true operational images, part of a loop that creates more of them. (soldiers buy drones that make images that award them with points they use to buy drones etc.). Once again like with the AI art slop, humans are somehow excluded from the loop.
Conclusion
I think operational images are useful as they have existed for decades, therefore we can observe when their medium and relationship to humans changes. I observed here three cases:
It has recently been made possible with AI to easily create photorealistic images that are not real. Photographs and videos used to be the ultimate proof something happened, and contributed to our own loop of updating our model of the world from information and acting on it. That loop has now been compromised by images targeting engagement over everything else.
Operational images emerged in the army, coupled with relatively narrow AI systems. We now see that the army is looking to incorporate LLM technology into its operational loops for autonomous weapons and domestic surveillance. Combined with the first case, this is one more evidence of operational images starting to affect humans by including them in the targets of optimization loops while making the oversight of these loops by humans harder.
Despite autonomous weapon systems having improved a lot since WWII, the Russia-Ukraine war has seen the advent of drones, killing machines with human pilots. And yet even when a human hand is controlling the machine, its decisions are part of an optimization loop provided by the footage of this drone being used as a performance metric.
Introduction
Images are how we perceive the world. They are important.
In this post I will talk about the concept of operational images, how the use of these images has risen and shifted, and why that is important in light of recent events such as the Anthropic/Department of War dispute, the war in Ukraine, or the rise of slop on social media. I hypothesize that these images allow systems to optimize for a single metric. By creating an optimization loop in a system, humans are far from excluded from it and are in fact often part of the metric as the systems improve. However, they lose legibility and understanding of a system that acts on their reality.
Historical context
The idea of "operational image" seems relevant to present events.
Conclusion
I think operational images are useful as they have existed for decades, therefore we can observe when their medium and relationship to humans changes. I observed here three cases:
Based on the viewing of the later documentary War at A Distance (2003) by Farocki, where in my opinion he leans more into that framing.