Marc Andreessen, the last of whom included “TESCREAList” in his Twitter profile for several weeks in 2023 (Gebru & Torres 2024)
We have traversed a lot of conceptual territory in this course, and this week we move to another term: critique, and a related term of criticism.
All of the articles and videos offer insight on different AI and machine learning topics. I'll be providing very brief coverage of two lines of criticism: what we can think of external (Goodlad and Stone; Gebru and Torres) and internal (Sutton, LeCun) criticism.
But the area I really want to focus on is the question raised by the Stahl article: what is immanent critique?
I raise this question not only because it arises in the context of Hegelian scholarship - and because he is the key thinker to inaugurate a concern about this question. It is because one of the difficulties today concerns how we address deep epistemic divides between proponents and critics of AI. Let me take as an example the remark Gebru and Torres make in the context of their article on TESCREALism:
> Marc Andreessen, the last of whom included “TESCREAList” in his Twitter profile for several weeks in 2023
Now this term is of course a critical one, and we can assume Andreessen is both aware of this and proudly applies the term to himself *ironically*. So here we get a key difficulty in the context of critique and criticism: today it is very easy simply to embrace the criticism and move on. If criticism is simply a game of name-calling and point-scoring, we might say it is ultimately ineffectual - something like a fashional pose we adopt, a sign of learning and distinction. How do we make criticism *stick*? This is the problem Stahl - and Hegel before him – seek to address via the idea of an immanent critique. In the first two readings, we see criticism applied to AI from the standpoint of scholars working in the field of what has been termed "critical AI studies". These are what Stahl might term "external" criticism: levelled at the general field of AI from outsiders (scholars typically working in the humanities and social science traditions).
The article by Goodlad and Stone is an introduction to a recent special issue launched by a new journal, in fact entitled "Critical AI". It develops a wide-ranging and lengthy but useful array of critical AI topics. These summarize a number of topics many of you will be familiar with, and that we have dealt with in previous weeks.
The article by Gebru and Torres develops criticism in another direction: what I would call an effort to characterize the "spirit" of Silicon Valley, which today motivates the development of Artificial Intelligence. Their description of TESCREAL seeks to join together otherwise disparate and even contradictory elements. On the one hand, the pursuit of scientific, mathematical and technological progress. On the other, the fantasy – at least for today – of an overcoming of mortality and the limits of life, which fall under the banner of transhumanism.
They further develop their analysis with a comparison with eugenics: a movement that coincides with the history of statistics, which underpins machine learning. This they connect specifically with the recent focus on AGI. They note one of the recurring oddities of the generative AI moment: that those researching or investing to develop AI are often those also expressing concerns about its existential risk. What is downplayed in the "existential" crisis of AGI is the perpetuation of everyday risks: algorithmic racism and sexism, environmental harms and the concentration of political power.
In the follow-up two videos, featuring Yann LeCun and Richard Sutton, we see a different type of criticism. Both are highly respected scholars in the field of Artificial Intelligence, responsible for some of the foundational developments in neural networks and reinforcement learning. Both are however critical about the overreliance upon the Transformer architecture we discussed in Week 4. In different ways, each thinks language models can never arrive at general - much less super - intelligence, because they are inherently limited.
Both offer criticisms that overlap, but also have differences, certainly in how each sees these limitations being overcome. For LeCun, this limitation relates to the relatively *static* nature of the language model. For both, the limits of language models relate instead to the lack of *experience*.So far we have examined two examples of two types of criticism. The first we can describe as *external*: developed by well-informed scholars who sit, like most of us, outside the immediate field of machine learning. The second we can describe as *internal*: developed by experts in machine learning itself, who are nonetheless critical of the way the field has evolved.
But whether we agree with these different critiques or not, we might still be troubled by whether *opposing* standpoints would find them convincing....Now we turn to the fifth article by Stahl, "what is immanent critique?" What *is* immanent critique? And what is its relevance for us, to think about machine and human learning?
Firstly, the purpose of immanent critique is to pose greater difficulties for those who would like to rebut that critique - because it aims to show it stems from the same *ground*, or shared premisses, as that which motivates the rebuttal.
Let's move on by introducing a couple of further terms: the *object* and the *standpoint* of critique. The object is the thing that we wish to judge with our criticism, while the standpoint is the position we adopt, implicitly or explicitly, when we do so. For example, the Goodlad and Stone article is directed toward the *object* of artificial intelligence, and their *standpoint* is that technologies such as AI should be equitable, sustainable and fair with respect to the work of creators and workers. We'll next introduce another term: the opposite of immanent critique, as a way for us to get into what can be quite a difficult idea to grasp, through the form of contrast. It is a term known as *transcendental* critique. What is transcendental, and how does it differ from immanent?
Both ideas come in fact from the world of religion. A transcendental god is one who sits outside the world. Who perhaps gives the world its start, via an act of divine creation, but thereafter remains aloof, watching from afar, perhaps only returning at some future date of Judgment. An immanent god, by contrast, is one who is embedded in our world, who is able to act upon it, or who is infused in every act of creation. We have already seen how Hegel's view is deeply immanent: the Spirit is always in a process of maturing, developing, evolving. This includes the Spirit in the godly sense - he means something like god is itself coming into being through the creation of the world, of us, of history, of technology. God's immanence means, for Hegel, that we are not only part of god's plan; our development is also the development of god, or of spirit. Hegel is in some sense of the term, a transhumanist. So what about critique? Let's stay with Hegel's idea of immanence a moment longer, even if we treat it just as a convenient fiction. If we assume our history - the history of our species, our culture, our technology – is the coming into being of god, then we might also assume there are missteps. These might be moments when god's development is arrested, held up, because we stray from the proper path. How do we know this? We might believe god's ultimate plan is for the realization of Absolute Knowledge; yet we might feel we are instead losing knowledge, becoming less intelligent, less educated - straying from the tendency towards god's own eventual realization.
In more concrete terms: we might believe our culture, science and technology *should* be leading us towards greater equality, mutual recognition, self-consciousness and awareness, political harmony, sustainability and so on. But then we notice in practice the opposite is true: wars continue to be fought, xenophobia – a kind of lack of recogition - is on the rise.
In short, we notice a contradiction between what *ought* to be the case, and what *is* actually the case. And this leads us to criticism.Now what makes this criticism *immanent*? In our story so far, we might note we do not in fact have ten commandments or a set of equivalent rules laid out for us. Instead, it is through our own evolution that we develop a sense of right and wrong, or the standards that allow us to talk about what *ought* to be the case. This sense that our standards of judgment arise historically – via our own trial and error - is fundamental to Hegel's picture, and part of his legacy is precisely this idea of critique that is immanent to our own historical process – and to not some *deus ex machina* who presides over us.
Why does this matter? One of the problems Hegel is trying to address - just like Kant before him – is that of *dogmatism*. Dogmatic beliefs are those that cannot be challenged, because in some way they refer to a transcendental conditions: for example, to a god that *I* believe in, but that *you* do not. Hegel wants to say that dogmatic knowledge, even when it looks like it is true, still rests upon what might be subjective and even irrational beliefs. Such knowledge is itself therefore ultimately irrational.
Rational knowledge, on the contrary, needs to be based upon foundations that understand themselves to be historically contingent - even if we don't know of any better. Then if we see contradictions, these foundations are capable of being revised - this is the sign of a properly rational knowledge, and the basis for immanent critique.
Why is this important? In the final analysis, transcendental critique - the kind of criticism that depends upon standards brought from outside the object – only holds if the speaker and the listener share some *a priori* values. For example liberalism: if we are both liberals, your critique of how AI (for example) offends liberal sensibility is likely to resonate. But what if I am opposed to liberalism? We are stuck: you appear to me as dogmatic; I appear to you as sceptical. What breaks this impasse? If you can show me how we are together by a historical process to understand certain values, such as equality, as precious - not as god-given, but rather acquired via painful historical trial-and-error. This doesn't require us both to be atheist; only that we see how history delivers us a set of ideals that seem to improve upon earlier versions. Then you might lead me to a conviction that AI (or something else) contradicts this historical and ethical achievement. Pointing out such a contradiction leads to a criticism that is *immanent*, because it does not depend us having some shared metaphysical values - only that we both acknowledge this historical process (a separate problem).As the Stahl paper argues, this process of immanent critique involves a complex dialectic exchange: we judge an object against a normative, ethical standard, like the printing press, and then we also judge our own standards too - recalibrating both object and stanard together. This process, as Stahl reminds us, is what Hegel means by the term "experience", and is the genuine form of learning that takes us forward toward self-consciousness and, ultimately, Absolute Knowledge. How do we bring this all home in the context of AI? First, we might want to think in terms of our own processes of response to AI as an object: how do we judge it against what we take to be our normative expectations of technology? Or in less formal language: what are our hopes for technology, and in what ways does AI fail to live up to them?
But second, how are our hopes themselves being modified as we start to assess AI? Can we begin to see that our standards themselves need to evolve - not just that we need new rubrics, for example, but that the foundations for those rubrics also need examination?
And on a wider front, leading us towards the differing visions of history and futurity we will discuss next week: is the advent of AI itself the kind of technology that might be essential to the very task of our critical self-examination? Is it the technology that can help transform the admitted obscurity of reflexive immanent critique? Can it lead us toward, as Mary and Bill might say, reflexive pedagogy? Or, as Gebru and Torres claim, is the AI project instead a massive fantasy projected by a disguised eugenics movement - a contemporary form, as Stahl discussed in relation to Marx, of "social domination" (rather than "conceptual self-determination")?Let's examine now the case of the printing press, as an example of an immanent critique of technology. We can look back upon the arrival of this invention and note its pros and cons: its democratization of information, on the one hand, and its propensity for the spread of misinformation, on the other. But then we might step back to say: where does our *desire* for democracy come from? In other words, what explains the basis of our judgment? We might say: we learn it, from parents, schools, community, all of which instils in us an appreciation for democracy, for an equal right to vote, and so on. But we might probe further, and note our education is in part depedent upon the low cost of the technologies for storing and disseminating information itself. These include the printing press, or derivatives from it. So a proper "immanent" critique acknowledges our debt to the very technology that we are busy critiquing.
Immanent critique involves then a process of trying to bring together the objects and our standards for assessing them in a single account. If we didn't have the object of the printing press, we might not have the benefit of several hundred years of the democratization of education to allow us to judge it. And this must form *part of our judgment*!
However that does not mean that critique needs to become mute. Instead it can observe the emergence of contradictions: when, for example, the technology of print is used to subvert the ideals of democracy and equality that it also gave rise to. In that case, we need to deliberate: either the object is not living up to the standards we expect of it; or our standards may have been too high; or some combination of the two. We need to be *reflexive* in our critique, and consider how both the object and the critique itself might need to be recalibrated.
Now this becomes more difficult in terms of technologies of the present, because we do not yet experience its benefits nor its deficits fully. So we have to *anticipate* many of its effects, while also acknowledging the role of similar technological innovations, like the printing press and the Internet, in the past. In breakout groups, we want to try to develop a rubric for a reflexive, immanent critique!
We'll start by imagining we are evaluating a new AI technology for education. We want to develop a rubric that seeks to judge the following:
a. the pros and cons of the technology itself - will it improve education? What are criteria for evaluting improvement?
b. whether our existing standards are sufficient for evaluation - or do we need new ways to evaluate? How do we evaluate our evaluations?
c. Can AI be a reflexive technology - one that helps us develop a critique that raises the standards of both human and machine? What would be criteria for assessing this?