“In my talk, I would like to go back to the basics of what is, or what makes up, an artificial intelligence; how AI is often mistaken for general artificial intelligence—and nowadays narrowed down to neural networks and machine learning; how understandings of “intelligence” can differ between engineering and arts/humanities; how and where artificial intelligences pre-existed digital computing and still do exist outside of it; machine learning-based AI as a remixer on steroids—that is, on top of that, structurally conservative; typical long-term frustrations with AI in general and this type of AI in particular, from the viewpoint of the arts.”
“Can humans think?”—this question posed by the former Fluxus artist Tomas Schmit is reminiscent of the child wondering about the emperor’s new clothes: While there are endless debates about whether or not machines can think, (almost) no one asks whether humans can think, too.1Schmit , Tomas Schmit, Können Menschen denken? = Are Hhumans Ccapable of Tthought?_, Cologne: Walter König, 2007. Exhibition catalog published on the occasion of an exhibition with the same title in the Museum Ludwig cologne and in the Sammlung Falckenberg – Kulturstiftung Phoenix Art.Museum Ludwig ; PhoenixArt, Sammlung Falckenberg ; Walther König, 2007 The fundamental question might be: What are we actually talking about when we discuss intelligence and cognition? A fundamental problem might be the lack of a generally agreed-upon definition, philosophical concept, or scientific theory of “intelligence.” How can one define “artificial intelligence” when even human—or animal—intelligence and “intelligence” as such do not have a clear and generally accepted definition?
A simple example: A 2 euro pocket calculator exceeds my own intelligent abilities—and those of most people—when it calculates, for example, “11293/37” within milliseconds and with floating point precision. So it is an artificial intelligence, not only by my perhaps idiosyncratic definition, but also according to computer science textbooks. We don’t even have to take an electronic calculating machine as an example: Even a two-thousand-year-old wooden abacus qualifies as such an artificial intelligence.
The same is true for other systems that extend or outsource human cognitive labor. Writing and signs are forms of artificial memory that also qualify as artificial intelligences if one counts memorizing as a cognitive activity. This example could in fact be extended to scrap notes and chalk marks as artificial cognitive devices that can memorize things better than human brains.
Perhaps the first thing to clarify is that in almost all cases where scientists and engineers refer to artificial intelligence technology today, they do not mean general artificial intelligence but specialized artificial intelligence. This is the major difference to most science fiction movies and to popular culture, where AI is almost always imagined as general AI. In all practical applications today, including (the AI language model and generator) GPT-3, neural networks, and machine learning, AI is only specialized, not general artificial intelligence. For example, GPT-3—as far as I know—cannot even do what the 2 euro calculator can do in my initial example.
Today, AI is primarily understood to be machine learning based on neural networks. This limited understanding is the result of both the technological advances and the commercialization of this technology over the last decade. However, neural networks and machine learning are only one AI technology and one AI approach among many. Until a few years ago, they were not even the dominant form of computer-based AI. People of my generation may remember AI based on programmed or hard-coded rules—so-called symbolic AI—for example, in the online translator Babel Fish from the 1990s.
Machine learning/neural networks, on the other hand, do not really work with sets of logical rules but are based on heuristics. Good examples of this are modern online translators like DeepL and today’s Google Translate (which only switched to neural network machine learning in 2016), as opposed to Babel Fish.
Apart from these different approaches and the historical paradigm shifts in AI research, one cannot draw very clear lines between computation, computer science, cybernetics, informatics, and AI: They are all closely related, if not competing labels for the same overall discipline (similar to how, for example, “cultural studies,” “aesthetics,” and “media studies” are related disciplines in the humanities).
The term “artificial intelligence” was coined in 1955 by the computer scientist John McCarthy in a research grant application. So it’s fair to say that the term was an elevator pitch, almost along the lines of “fake it till you make it.” “Artificial intelligence” was proclaimed in a time when computer technology was still very limited and couldn’t do much more than today’s 2 euro pocket calculators (but about as much as today’s 20 euro programmable pocket calculators).
Perhaps an analogy will help here: “artificial intelligence” as a term or concept is about as clear as the hypothetical term “artificial creativity,” under which one could lump almost any creative technology, from a wooden perspective ruler to a movie camera to a simple drawing computer program to GPT-3: they all qualify as “artificial creative agents” in one way or another. If one now imagines that the term would be narrowed down to technologies based on machine learning, such as GPT-3 or Adobe Photoshop’s “contextual fill,” the same terminological problems would arise as with “artificial intelligence” today.
But even if we reduce our debate to machine learning as artificial intelligence, what are its possibilities and limitations?
Perhaps it is no longer necessary to explain how machine learning works in principle. (If not: A software program is given training sets, for example pictures of cats and dogs. It is told by user feedback which images show cats and which show dogs, and then tries to figure out for itself—by trial and error via feedback—which pixel characteristics constitute a cat and which constitute a dog. When the system is able to distinguish cats and dogs well enough, it is deployed [for example, for camera detection at an airport], usually without anyone knowing what distinguishing criteria it uses internally. Contrary to popular misconceptions, the system does not develop or use an algorithm to distinguish cats from dogs but “only” a heuristic for pattern recognition. Algorithms exist only at the meta-level that tells the system how to “learn” and work.)
What can go wrong with AI machine learning? For example, if the cat images in the training set were all taken with one camera and the dog images were taken with another, and the cat images have a more yellowish white balance, the system might simply conclude that everything on a yellowish image is a cat.
This problem can become more pronounced when such a system is used for analysis—or, in contemporary tech jargon, “analytics.” Let’s imagine a warning system for forest fires based on machine learning. (This example, by the way, is not mine.) Analyzing surveillance camera images of the forest, the system might find that children walking in the forest and eating popsicles coincide with forest fires—and conclude that children eating popsicles are causing the forest fires. However, the system does not reach the conclusion that the true cause of both children eating popsicles and forest fires is hot weather.
Another example (which I have used before in the book Pattern Discrimination) is a machine learning-based “smart building” system that keeps track of the chairs in a building and dynamically assigns them to different rooms, for example in a school, in order to save materials and make the building more environmentally sustainable.2Wendy Hui Kyong Chun, Hito Steyerl, Florian Cramer, and Clemens Apprich, Pattern Discrimination (Minnesota: University of Minnesota, 2018). If machine learning had existed a hundred years ago and the system had been trained on the chairs that existed at the time, it certainly would not have recognized a Bauhaus tubular steel chair as a chair. It might even have set off an alarm for an unknown, potentially dangerous object. (I think of this as a very realistic near-future scenario.) The worst consequence of this could be that new, experimental chair designs are no longer being developed, either because they trigger alarms or because it is simply too time-consuming and expensive to update the “smart building” training sets to recognize these new chairs as chairs.
To return to the example with the children in the forest: Machine learning/neural networks are based on the detection of correlations, not on the reconstruction of causality.
This may also be where the bias and discrimination of these systems originate. The forest fire example, among others, can be extended to predictive policing algorithms that stigmatize immigrant neighborhoods by becoming self-fulfilling prophecies: When a neighborhood is more heavily policed, more street crimes are reported and prosecuted, and crime statistics show that more policing is needed, while in a wealthier neighborhood where policing is correspondingly reduced, fewer street crimes are detected and crime statistics show improvement. Thus, a youth growing up in a rich neighborhood is much less likely to obtain a criminal record than a youth in a poor neighborhood, even if both commit the same petty crimes.
The neural network-based machine learning paradigm can be simply explained with the word remix. You feed the learning algorithm—or a language model like GPT-3—all the Shakespeare poems, and then it starts writing Shakespeare poem remixes. These remixes are “better” from a technical point of view, the more they retain the characteristics of Shakespeare’s original poems and act as a Shakespearean ghost, avatar, or bot.
For poetry and other arts, this can be more of a letdown than a promise. As early as 1967, the Italian novelist Italo Calvino (who was a member of the French writers’ group Oulipo, which was experimenting with formal constraints and algorithms) concluded, with some disappointment, that computer-generated literature was “classicist” or conservative.3Italo Calvino, “Cybernetics and Ghosts,” in The Uses of Literature: Essays (San Diego: Harcourt, 1986), 3–27. (He was not the only literary writer to experience computer-generated literature as initially exciting and ultimately disappointing.) This classicism is not a bug but a feature: neural network-based machine learning as such is structurally and epistemologically conservative, since its paradigm is to extract a norm from the past and apply that norm to the present.
The principle underlying neural network-based machine learning is the same as that of the mathematically much simpler Markov chains discovered by the Russian mathematician Andrei Markov in 1913.4Andreĭ Andreevich Markov, “An example of statistical investigation of the text Eugene Onegin concerning the connection of samples in chains,” in Science in Context 19.4 (2006): 591–600. Markov computed—with pen and paper, without electronic or mechanical aids—remixes of Pushkin’s Eugene Onegin based on the so-called transition probabilities of its words: i.e., the probabilities of certain groups of words in the novel, allowing recombinations of the text that were quite grammatical, although the mathematical text processing involved no understanding of its grammar or meaning.
This evening, I came home late for dinner. \
Because of a traffic jam, she came home late and missed an appointment.
can be recombined as:
This evening, I came home late and missed an appointment.
Markov chains have been frequently used in experimental writing, including by the Oulipo group, the German semiotician and poet Max Bense, and the American poets Jackson Mac Low and Charles O. Hartman. In music composition, Markov chains served as a compositional algorithm in Lejaren Hiller’s Iliac Suite and Karlheinz Essl’s remix of a piece by Johann Sebastian Bach, among others. Markov chains are commonly found in nonsensical Internet chatbots and in everyday applications such as Google’s “GBoard” on-screen keyboard for the Android operating system, which makes real-time suggestions for completing words based on user’s previous writing on the device. (Anyone who has dabbled in algorithmic arts has probably encountered Markov chains at some point.)
I have argued that neural network machine learning is based on similar heuristic principles as Markov chains. In simpler terms, neural network AI could be called Markov chains on steroids. Although neural networks are much more sophisticated than Markov chains in their ability to find correlations, their results still exhibit the same problems, which might be called, “algorithmic bias,” with Italo Calvino, “aesthetic conservatism,” or, as I’d like to call it, “algorithmic posthistoire”: a state of being artificially trapped in the past because of systems that have been trained with that past; systems that have ceased their machine learning the moment they have been moved from training mode to operational mode. (Which, in my opinion, is an enormously significant and underappreciated difference with human and animal learning and action.)
If my characterization of neural network machine learning AI as “Markov chains on steroids” is too techno-esoteric, then a more accessible might be the one of the German hacker and computer security expert Felix von Leitner who called machine learning “a glorified database with weighted records” (eine glorifizierte Datenbank mit Gewichten an ihren Records).5Felix von Leitner, “Zwischen Forschung Und Gelebter Realität […],.” Fefes Blog, 19 January 19,. 2022, https://blog.fefe.de/?ts=9f16df52. Accessed June 22, 2022, https://blog.fefe.de/?ts=9f16df52.]
At the beginning of my talk, I mentioned the frequent confusion of AI with general AI, or the fact that all of today’s AI systems and technologies are “just” special-purpose AI and not the kind of general AI imagined in science fiction and popular culture. But even if we limit ourselves to special AI systems, they tend to reveal their limitations only after prolonged engagement with them. Elsewhere I have suggested that to call this the “kaleidoscope constraint” of programmed systems: Like a kaleidoscope, at first glance they seem rich, surprising, and nearly infinite in their combinatorial creativity, but at some point they become boring and predictable, even if they do not literally repeat their output.
In my opinion, the core of the problem lies in the differences in the understanding of “intelligence” between the engineering sciences on the one hand and the arts, humanities, and social sciences on the other. From an engineering point of view, a system behaves “smart” as soon as it performs some computational or cognitive task (even if it is only the calculation of the sum of one and one). From a humanities point of view, however, this is not a satisfactory concept of intelligence.
We are confronted here, it seems, with the same structural misunderstanding as the one that was characteristic for the so-called “interactive art” of the 1990s and early 2000s. In engineering, anything is called “interactive” that does not run autonomously but involves feedback. In other words, a constantly running fan is non-interactive while a simple light switch is interactive (because it requires user interaction), from an engineering perspective. Of course, this concept of interactivity is not remotely satisfying from an arts, humanities, or social sciences standpoint. This discrepancy of terms also explains a common frustration with “interactive installation art” that, indeed, more often than not boiled down to glorified light switches.
No doubt, machine learning AI systems exceed older, very limited expectations of machine “smartness,” such as the artificial intelligence of a pocket calculator and the interactivity of a light switch. The characterizations I brought in, such as “correlations instead of causality,” “Markov chains on steroids,” “kaleidoscope constraint,” “database with weighted records,” and “algorithmic posthistoire,” illustrate how much machine learning AI has complicated matters while still suffering from the same fundamental issues.
Perhaps the most accessible term for the shortcomings of machine learning AI is “laziness,” which was proposed by the MIT computer scientist Aleksander Madry. In an interview, he characterized the laziness of neural network machine learning as follows:
“Think about being lazy as this kind of smart student who doesn’t really want to study for an exam. Instead, what he does is just study all the past years’ exams and just look for patterns. Instead of trying to actually learn, he just tries to pass the test. And this is exactly the same way in which current AI is lazy.”6Steve Nadis, “The Promise and Pitfalls of Artificial Intelligence Explored at TEDxMIT Event,” MIT CSAIL (blog), January 11, 2022, accessed June 22, 2022, https://www.csail.mit.edu/news/promise-and-pitfalls-artificial-intelligence-explored-tedxmit-event
In my opinion, many of these current issues of AI have to do with the lack of involvement of artists, humanities scholars, and social scientists in the development and application scenarios of this technology. Many of its flaws and limitations seem remarkably obvious from a humanities perspective.
In the case of systems like GPT-3, we are also dealing with commercially developed, proprietary black boxes that delegate artists (and anyone else) working with them into the role of mere users or consumers. Writing poetry with GPT-3, for example, amounts to painting by numbers or using a music keyboard with preset melody and rhythm patterns. This can be an interesting practical use, of course, just as numerous artists have worked with painting templates and musical presets, even in interesting, surprising, and critical ways. But—in my opinion—such uses always presuppose a clear awareness (as in Oulipo) that these systems constitute constraints rather than merely expansions of one’s practice.
Another trap is naive post-humanism. What I mean by this is premature—and ultimately romanticist—jumps to conclusions about living together with a seemingly autonomous non-human fellow being that would ignore the (very human, corporate, interest-driven, political, social, economic, cultural) constructedness of that entity. With any programmed system, no matter how sophisticated, the question always remains: What is the politics underlying its design? Who constructed it with what agenda? Why does it work the way it does? Could it work differently, and how? What could we imagine it to be?
FLORIAN CRAMER, reader in 21st century visual culture/autonomous art and design practices at Willem de Kooning Academy and Piet Zwart Institute, Rotterdam, Netherlands. Publications related to AI include: Pattern Discrimination (University of Minnesota Press and Meson Press, 2018) co-authored with Clemens Apprich, Wendy Hui Kyong Chun, and Hito Steyerl, and “Crapularity Aesthetics” (in Making and Breaking, vol. 1); both publications are Open Access.