Thomas Nagel begins What Is It Like to Be a Bat? with a problem that sounds philosophical but increasingly feels practical. We can know all there is to know about the physicality of being a bat, e.g. conceptually understand how echolocation works. But even a complete description of these facts still does not explain what it's *actually* like to be the bat.
TLDR; there is a difference between knowing everything about a mind, through outside observation, and truly occupying its point of view.
The view from nowhere
Nagel describes how different observers can “converge on a single thing, be it an object, a person, a process, an event, or whatever.” He believes that objective inquiry can approach "an endpoint," the “completely objective intrinsic nature” of the thing itself.
The thing that disappears
For most objects, removing point of view gets us closer to the thing. For consciousness, Nagel argues, removing point of view removes the thing we're trying to understand. “We cannot ignore this point of view permanently, since it is the essence of the internal world.”
This matters for collaboration. Observability can tell us what a system did, but it does not automatically tell us what the situation meant from inside. For example - a teammate can say “sounds good” while holding substantial uncertainty. Two people can read the same conversation and leave with different models of what was decided. An agent can produce a useful answer while its summary gives us only a lossy description of how it got there.
Language as an interface
This problem didn't begin with AI. Humans have never had direct access to one another’s internal worlds either. We translate them via these symbols we exchange. Language compresses something private into tokens from which another person constructs an approximation of what we mean. Usually this works astonishingly well. Sometimes it fails catastrophically. Much of collaboration consists of noticing the nuances of tokens
That makes human-agent collaboration feel less alien than it initially appears. We are not suddenly confronting inaccessible minds. We are confronting a much greater diversity of internal representations, capabilities, temporalities, and ways of arriving at conclusions.
Interfaces for partial understanding
The obvious response is to demand greater transparency: show the reasoning, expose every source, quantify uncertainty, record everything. Some of this is useful. At Ando, we care deeply about provenance, context, correction, and making it possible to understand why something happened.
But perfect transparency is probably the wrong endpoint. A human who verbalized every thought would not necessarily become a better teammate, and more information about a mind does not automatically produce greater understanding of it.
Instead, we think interfaces should be designed around partial understanding, e.g. agents should be able to express uncertainty rather than manufacture confidence.
Toward an objective phenomenology
Near the end, Nagel wonders whether we might develop “an objective phenomenology not dependent on empathy or the imagination.” Not a magical ability to become the bat, but a language for describing aspects of another kind of experience without requiring us to reproduce that experience ourselves.
We don’t need humans and agents to possess identical internal representations. We don’t even need to settle the harder question of whether their internal states are phenomenologically comparable. We need ways for unlike minds to establish enough common structure to coordinate.
Perhaps the goal of collaborative software isn’t complete mutual legibility. It’s sufficient mutual intelligibility.
Nagel’s bat is usually read as an argument about consciousness. For us, it also poses a practical design question:
How do you build a shared world for participants who may never fully know what it is like to be one another?
Our answer, at least for now, is not to erase the gap. It is to make it possible to work across it.