Julian Hauser

One question we can ask when investigating the nature of self-representation concerns the types of property that must figure in its content. Here, authors have claimed that self-representations must be about spatial, temporal, bodily, or mental properties. However, we can also ask a second question: how do we need to represent a property to self-represent it? I address this latter question. I argue that a distinction between egocentric and allocentric forms of representation—known from spatial cognition—also applies to representations of other kinds of property. I use examples drawn from animal cognition and developmental psychology to show how creatures allocentrically represent their temporal, bodily, and cognitive properties. These representations are minimal self-representations: they represent one’s properties so that an explicit differentiation is made between the system and other objects (or between the system’s actual and merely possible properties), they are directly linked to behavior and sensation, and they are immune to error through misidentification. The upshot is a view on which different creatures may self-represent more or fewer kinds of property. More substantive forms of self-representation (for instance, as exemplified by neurotypical adult human beings) then require integrated minimal self-representations of the right kinds of property.

Joint attention is characterized by openness: when two agents jointly attend to an object, they are immediately and fully aware of each other's attentional states. Existing accounts of openness involve a mental picture in which two agents attend to the same object and where openness is then 'added'. I argue that the experience of openness comes first. Young infants operate under a tacit assumption of openness: they behave as if attentional states were open even when they aren't. The ability to engage in joint attention doesn't arise when infants begin to experience openness, but rather when they can limit these experiences to open interactions. For this, they depend on cognitive processes that detect non-open interactions. Some of these processes develop early and don't require the representation of others' mental states. Other processes develop later and require the infant to differentiate between herself and others as subjects of attentional states.

Phenomenal transparency was once widely believed to be necessary for cognitive extension. Recently, this claim has come under attack, with a new consensus coalescing around the idea that transparency is neither necessary for internal nor extended cognitive processes. We take these recent critiques as an opportunity to refine the concept of transparency relevant for cognitive extension. In particular, we highlight that transparency concerns an agent’s employment of a resource – and that such employment is compatible with an agent consciously apprehending (or attending to) a resource. This means it is possible for an object to be transparent and opaque to an agent, even at a single moment in time. Once we understand transparency in this way, the detractors’ claims lose their bite, and existing arguments for transparency’s necessity for cognitive extension return to apply with full force.

We might worry that our seamless reliance on AI systems makes us prone to adopting the strange errors that these systems commit. One proposed solution is to design AI systems so that they are not phenomenally transparent to their users. This stops cognitive extension and the automatic uptake of errors. Although we acknowledge that some aspects of AI extension are concerning, we can address these concerns without discouraging transparent employment altogether. First, we believe that the potential danger should be put into perspective – many unreliable technologies are unlikely to be used transparently precisely because they are unreliable. Second, an agent who transparently employs a resource may also reflect (opaquely) on its reliability. Finally, agents can rely on a process transparently and be yanked out of their transparent use when it turns unreliable. When an agent is responsive to the reliability of their process in this way, they have epistemically integrated it, and the beliefs they form with it are formed responsibly. This prevents the agent from automatically incorporating problematic beliefs. Responsible (and transparent) use of AI resources – and consequently responsible AI extension – is hence possible. We end the paper with several design and policy recommendations that encourage epistemic integration of AI-involving belief-forming processes.