Definitions

Bridging

bridging goal (intuition): An increase in mutual understanding and trust across divides, creating space for productive conflict, deliberation, or cooperation.
bridging (formalism): An improvement in relation metrics that corresponds to the bridging goal.; We say that an attention-allocator (see below) is bridging to the extent that it causally supports the bridging goal through its allocations.

Note. We distinguish between the intuition and formalism of bridging. Our aim in doing this is to decrease the likelihood that the formalism (and its optimization) will overshadow the rich human experiences it is intended to make legible. Where formalism conflicts with intuition, intuition should generally be favored (in order to avoid Goodharting).

Note. We define bridging as an outcome, rather than a process, but processes or systems can be described as bridging to the extent that they bring about that outcome.

Attention Events

slot: A finite resource. (Metaphorically, a container which can be allocated to hold an object.)
object: Anything that can consume that resource. (Metaphorically, anything that can occupy a slot.)

Note. We define object and slot circularly, but in a given context it should be clear which is which. For example, an object might be an item of content on a social media platform, and the corresponding slots might be discrete positions within a recommender feed. To give another example, slots might be continuous intervals of first-person experience, in which case the corresponding objects would be anything that can be attended to.

atomic allocation: The filling of one slot with one object. It can be modeled formally as (slot, object, properties).; Examples: Choosing the third item to show a user in a feed, introducing an idea in a discussion.
(general) allocation: A set of atomic allocations.; A potential allocation is one which has not (yet) happened (and may not happen), and a realized allocation is one which has already happened.; Examples: Choosing the first 20 items to show a user in a feed, a discussion.

Allocation Systems

allocation system / allocator: A system that takes as input a set of potential allocations and outputs a set of realized allocations. It consists of an allocation process and, optionally, a learning process.
allocation process: A process that takes as input a set of potential allocations, and outputs a set of realized allocations. It is the core component of an allocation system.
learning process: A process that updates the state and prediction models used in the allocation process with data that has been collected, retrieved, or elicited (including e.g., data collected through an allocation process). It is an optional component of an allocation system.
value model: A method for aggregating the multiple predicted impacts of a potential allocation into an overall measure of value. It is used in an allocation process.

Attention

attention: The selective processing of information by a system, such that the state of that system might materially change.; Attention is defined with respect to a system boundary, so one can talk about human individual attention, human group attention, or algorithmic attention (here we focus primarily on individual human attention).
attention allocation system / attention allocator: An allocation system that is involved in the allocation of attention.; Examples. A recommender system; a human facilitator of a deliberation; your own brain.

Note. We define “attention” such that both humans and algorithmic systems can be all be said to have attention. The reason for including algorithmic forms of attention is that we can conceive of algorithmic processes — imagine an “automated small business” or “automated newspaper” — for which the allocation of algorithmic attention can have morally-relevant impacts on humans.

Optimization Stack

optimization stack: The multiparty, multilevel optimization that occurs in an attention-allocator. At minimum, it consists of optimization for stakeholder objectives during system design, accuracy optimization during the learning process, value optimization during the allocation process, and the strategic behavior of users who optimize for personal objectives when interacting with the system.
system design: The process of making decisions about the design of an attention-allocator that are made outside the normal course of operating the system.
accuracy optimization: Optimization for accuracy during the learning process.
value optimization: Optimizing for value during the allocation process.

Data & Modeling

relation model: A formal representation of the relationships between people in a population.; Can be decomposed into the triple (people, items, relations): the people who interact with the system, the alternative items to which they can attend (which may also be people), and the one-to-one relations between people and items, intended to capture goodwill, agreement, affinity, reactions, or similar.; Common examples include graph-based and space-based models.

Signals

signal: Numbers used for ranking. (More precisely, numbers that are aggregated into an overall measure of value and used as the basis for value optimization during an allocation process.); Signals can be either observed or predicted, depending on whether the information they represent is already known.; Signals can also be either causal or heuristic, depending on whether allocating attention to the corresponding object is known to causally contribute to bridging, or is merely thought to do so.
motif: A pattern of interaction, observable on a platform, that is thought to be associated with certain outcomes (e.g., bridging).

Metrics

metrics: Statistics used for system evaluation (e.g., evaluation of attention allocators).
relation metrics: Statistics summarizing the state of a relation model (or a subset of the relation model) at a given point in time.; Relation metrics should have a clear normative interpretation. Within a given context, we should be able to say that an increase in relation metrics is good (i.e., corresponds to the bridging goal), or that certain configurations of multiple relation metrics are better than others.
bridging metrics: Statistics summarizing a change in relation metrics.; Positive bridging metrics should mean that relations are improving, and negative bridging metrics should mean that relations are deteriorating.

System Properties

division bias / bias towards division: A property of attention allocators if their bridging metrics are negative.