This talk investigates the following two questions: Q1. Under what conditions do human AI-based beliefs qualify as knowledge? Q2. Do the seemingly crazy errors that AI systems sometimes make pose a threat to human AI-based beliefs qualifying as knowledge? The discussion of Q1 and Q2 is set against the background of a stock of examples of AI errors, including adversarial examples drawn from the large literature on image classifiers and LLMs. Many of these errors strike humans as bizarre or crazy—e.g., LLMs ‘hallucinating’ references or an image classifier correctly classifying an image of a panda but switching the output to ‘gibbon’ after the original image is subjected to a humanly imperceptible manipulation of its pixel structure. The talk brings Q1 and Q2 into connection with mainstream epistemology—more specifically, modal epistemology. The key idea is that, in order for a belief output of a given method to qualify as knowledge in a given world w, the belief must not only be true in w; it must likewise be sufficiently modally robust. The talk discusses the prospects of AI-based knowledge, given modal conditions on knowledge and the wealth of adversarial examples that have surfaced in AI research.
According to scientific realists, the success of a scientific theory provides strong evidence that it is (approximately) true (Putnam, 1975). In response, antirealists argue that the theories we have are successful because they are survivors of a selection process where unsuccessful theories are rejected, so truth is not necessary to explain success (van Fraassen 1980). This paper argues that the training and testing process of artificial intelligence is structurally analogous to the selection process of scientific theories. Convolutional Neural Networks (CNN) achieve human-level performance in image classification through iterative training procedures that adjust weights and biases to minimise errors. Moreover, recent techniques in explainable AI (XAI) can approximate concept-level interpretations of the CNN’s structure. Some of these concepts align with human concepts, while others do not, even when predictive performance is comparable. The CNN is interpreted as encoding a structural representation of the data, analogous to how a scientific theory represents phenomena. To the extent that the AI classifier uses similar concepts to humans, we have support for realist interpretations of successful representation. Conversely, divergence from human concepts lends weight to antirealist interpretations.
A fast-moving debate has emerged over whether LLMs are bullshitters in any significant sense. This talk develops an account of LLM bullshit that, in contrast to the most influential existing accounts, is entirely output-based. I begin with an overview of the best-known treatment of LLM bullshit, due to Hicks, Humphries, and Slater, along with some of the main critical reactions to their views. One response, from Gunkel and Coghlan, argues that Hicks et al.’s process-based account should be replaced by an output-based one. I take this response to be compelling, though it is notable that Gunkel and Coghlan do not attempt to develop a detailed output-based account. To fill this gap, I review Florian Cova’s recent output-based account of bullshit, explain how it can be streamlined, and show how it can be applied to LLMs. The main upshots are: (i) some but not all LLM outputs are bullshit; (ii) LLMs engage in the activity of bullshitting sometimes but not always; and (iii) LLMs are bullshitters in only a rather weak sense.
The de facto situation regarding trustworthy AI is that the principles and supporting guidelines of are largely settled, from a pan-cultural perspective, and that if we build this trustworthy AI—all other things being equal—this will lead to greater AI adoption.
There are some consequences that may be drawn by AI accelerationists from this. First, we don’t need to expend resources engaging with the communities impacted by AI to determine what makes AI trustworthy for them. Instead, it is a matter of building trustworthy AI and getting that AI in front of people to facilitate AI adoption. Second, on balance, this version of trustworthy AI constitutes a societal good: real trustworthy AI mitigates harms while delivering maximal benefits. Third, if we build trustworthy AI according to these assumptions, it is not rational for people to be sceptical of AI.
And, following from that, those who raise fears among the community regarding AI adoption are both doing a disservice to that community and are not acting in a rationally-justified manner. In this paper I critique this common notion of trustworthy AI, discussing AI in the context of a plurality of world views, and critique the claims made above.
Critics rightfully identify that AI models are biased against marginalised groups. These biases deteriorate our shared hermeneutical resources—the narratives, frameworks, and concepts that structure how we understand the world and ourselves—by reflecting and exacerbating existing oppressive narratives. However, this is not the only way that AI models are sources of hermeneutical impoverishment. I propose that AI models warp our hermeneutical resources, not only by reinforcing existing problematic representations of identity groups, but by changing how these groups are represented. That is, AI models are conceptual engineers, capable of revising our social concepts. When certain deep machine learning models perform predictions, they construct social concepts. Crucially, these algorithmic concepts differ from their human-constructed counterparts due to unavoidable trade-offs in model development. In constructing revised algorithmic concepts, AI models act as conceptual engineers. Once introduced, algorithmic concepts can take the place of our own concepts. Through these hermeneutical changes, AI models can also make a difference to our underlying social ontology: in redefining how we think of ourselves, they can redefine who we are. Finally, I offer upshots of attending to AI models as novel sources of epistemic and ontological harm.
