Why Solving Existential Risks Related to AI Might Require Radically New Approaches

Some of the existential risks associated with AI are listed in table 1 below.

Table 1: Existential risks associated with AI.

However, the most important existential risk in terms of its potential for exponential impact might be invisible. That risk is that in any group decision-making process in general there are different problem-solving strategies that each work for different categories of problems. However, most individuals are biased towards a very limited range of strategies (table 2). Without a process enabling groups to cooperate to select the group problem-solving process that is most “fit” in terms of relative collective outcomes for every participant, it is hypothesized that individuals in groups will instead compete to “win” the battle to impose their own view of what is the most optimal problem-solving strategy.

Table 2: Sectors of human decision-making.

This introduces centralization into decision-making that effectively prioritizes the individual outcome of “winning” the battle for power, control, and resources, as opposed to prioritizing impact on collective outcomes like existential risks associated with AI. The reason this centralization is largely invisible is that anecdotal observation suggests that individuals generally don’t see that any problem-solving strategy they are biased towards has a limited range of applicability.

This might be because it’s so tempting to think that when one’s problem-solving strategy is prioritized all problems will be solved. However, the only example we have of problem-solving strategies that can vastly or even exponentially increase collective problem-solving ability so they are reliably capable of solving problems too complex for individual entities, is the collective intelligence that exists at multiple scales in nature.

In the multiple billions of years that life has existed on earth, life has only been observed to have solved its most complex problems like vision and cognition through networks (i.e. through multi-cellularity). If an exponential increase in impact isn’t reliably achievable in any other known way, then the specific properties of this natural collective intelligence that allows the discovery of such networks are profoundly important, particularly where they concern societal problems of existential importance.

However, the entire AI industry shows a heavy slant towards consensus based decision-making, which is only one problem-solving strategy. Theoretically, networks of interventions can cooperate to vastly or even exponentially increase their impact as compared to interventions that compete for impact on their own. Consensus fails in its ability to select the networks of interventions that are potentially capable of this increase in collective impact, including impact on outcomes like addressing the existential risks associated with AI.

If true, this means that all research in this area might be effectively and invisibly confined to lines of inquiry that can’t reliably solve the most important problems in this area. This critically important problems is defined here as "General Collective Optimization" or GCO, in contrast to "General Individual Optimization" or GIO.

A fundamental characteristic of living systems, and consequently their infrastructure, is their inherent inclination towards self-perpetuation. Single-celled organisms, which predominantly employ the GIO strategy, tend to perpetuate this approach, while multi-cellular organisms, which employ the GCO strategy, perpetuate the latter. This implies that every technological innovation and decision made by entities driven by GIO tends to reinforce the absence of the necessary networks and infrastructure essential for GCO to optimize collective outcomes. This leads to the remarkable realization that GIO potentially generates an imperceptible force that sustains GIO-centric systems.

GIO tends to resolve problems in a manner that primarily benefits the fitness or well-being of a single entity, while GCO tends to resolve problems in a manner that primarily the collective fitness and well-being of all participating entities. Whether an entity experiences GIO or GCO depends on one's position within or outside the group that GCO serves. Achieving GCO requires specific infrastructure capabilities, including the systematic identification of collaborative networks of processes that can achieve collective outcomes far superior to what any individual process could accomplish, and infrastructure capable of assessing the state of each participating entity and establishing a metric for collective fitness.

In terms of collective societal outcomes, this implies that without introducing the capacity for GCO, technological and other systems must over time facilitate less and less beneficial collective outcomes where doing so benefits the entity in control of key processes. Because as mentioned, in terms of collective outcomes, GIO implies competing in order to “win” control over market share (and thus win control over power, decision-making, and resources) regardless of the impact on collective well-being. GCO on the other hand is predicted to create an invisible force towards cooperating to sustainably decrease harmful consumption.

The infrastructure hypothesized in this article to be required to distribute and decentralize any decision in general in order to create the capacity to vastly or even exponentially increase the collective intelligence that might be directed at any group problem, while also aligning outcomes with the collective fitness or well-being (that is the infrastructure required to achieve GCO), has been described as a hypothetical General Collective Intelligence or GCI platform with the following properties:

A portable semantic model of concepts and reasoning that is able to represent the knowledge in an AI (the so-called “conceptual space) so that meaning rather than information might be exchanged at vastly greater speed and scale, and so that AI becomes more explainable, and therefore safer.

A private repository of data, processes, and identities in the solve and private control of the individual user that owns them.

On or more intelligent agents that understand the individual user’s perspective, and that can optionally broker interactions with the GCI on behalf of the single user who owns them, thereby removing the limits to the speed and scale at which users can interact with the collective reasoning in order to achieve outcomes that are optimal from their own perspective, rather than from a perspective that is imposed on them.

It’s counter intuitive, but without introducing the capacity to potentially distribute and decentralize ALL decisions, then there is no guarantee that the decision acting as a bottleneck in getting rid of centralization, is a decision that can be made in a way that’s aligned with optimal collective outcomes.

Efforts to validate that this conceptual space can be implemented are underway. Working papers already explore the potential impact of this approach on AI explainability and AI safety. These papers await empirical demonstration of the conceptual space. The challenge remains to create enough awareness of this approach to make organizing this interdisciplinary cooperation so that implementing the conceptual space is reliably achievable.

This is an existentially important issue. Without the ability to reliably optimize collective outcomes, and instead being driven by systems that are complex to the point of being invisible to the vast majority, but that invisibly optimize outcome for come centralized entity, then human civilization isn’t working to solve our most important problems. Instead, civilization is effectively working as a plantation or farm to invisibly grow economic value from human labor, and on behalf of our “owners” to invisibly harvest what might be a greater amount of that value than the plantations of the past.