Scenario Mapping Advanced AI Risk: Request for Participation with Data Collection

Bottom Line: I am a graduate student researching advanced artificial intelligence (AI) futures and typologies that could result from variations in technological development. I am developing a modeling technique that captures many of the more complex non-quantifiable dynamics and will fill an important gap in the literature.

I would greatly appreciate your perspectives on potential AI paths through a brief survey. All questions are multiple-choice and will remain completely anonymous.

Edit: The Survey from the previous post was updated to add more thorough definitions and descriptions and modify some of the language and framing.

Survey overview

The survey presents several AI dimensions (key factors—e.g., rate of change) followed by conditions (scenario path—e.g., fast) and asks participants to:

  1. Rank the likelihood (plausibility) of the condition to occur (from highly likely to highly unlikely).

  2. Rank the potential impact of the condition on security and stability (from greatly increase to decrease or no effect)

The survey typically takes less than 10 minutes to complete. This project is not for prediction or forecast, but to assess a broad range of scenario conditions (including the speculative and subjective) to create broad categories for the model.

Likelihood values are derived from the MITRE risk scale.

Key questions that the dimension and condition values seek to measure include:

  • What AI paradigm could lead to high-level capabilities?

  • Which technologies could accelerate progress?

  • What are the most concerning mid to long-term risks?

  • How do differences in development, technological transitions, and distribution impact social-political stability? And how do the differences impact power asymmetries and control?

  • What governance structures map most effectively to typologies of advanced AI?

There is a solid body of research that aims to forecast AI development (e.g., here, and here), and this does not attempt to add to it. Several of the scenario elements in the survey are highly speculative and qualitative, as the primary goal is to explore all options, including very messy non-quantifiable dimensions, as best possible, to develop unique unexplored scenario combinations that may provide some insight.

These conditions are relatively imprecise and difficult to measure but an AI researcher’s view on whether deep learning could lead to AGI I believe is much more trustworthy than a random guess.

Another Scenario Survey?

This survey varies from similar projects in that it:

  1. Seeks to identify unique combinations, and potentially, new scenario options.

  2. Requests participants to rank individual elements of possible scenarios, not fully developed scenarios themselves.

  3. Uses the participant’s rankings as inputs to the morphological model to cross-compare, combine, and cluster scenario elements (conditions, rankings) to exhaust combinations.

This work primarily contributes to the literature through the futures modeling technique and is the first time the method has been applied. There is a need for more comprehensive scenario modeling in the AI risk space in general, notwithstanding some notable exceptions (e.g., Gruetzemacher, Avin) However, scenario methods overall tend to struggle with capturing high degrees of complexity and uncertainty, with outcomes generally confined to a limited number of options.

Framework

I am modeling AI through an evolutionary complex systems framework to unravel some of the less-studied structural risks and dangers from competitive agent processes (e.g., decision and value erosion).

This research presents a classification framework, starting with three broad high-level components, followed by 13 nested dimensions that influence AI risk, 43 individual conditions of possible futures, and 129 (tentative) indicators that could potentially monitor technological and social-political developments. (a more comprehensive description of purpose and methods can be found here: https://​​tinyurl.com/​​AIfutures)

The purpose of this research is to: 1) classify components of AI development that impact risk outcomes; 2) present a systematic structuring technique for AI modeling that highlights the spectrum of risks, while exploring unanticipated pathways; 3) map AI development scenarios to system typologies from variations in technological progress, and 4) construct a hypergraph of influence paths between the system dimensions and conditions to suggest how actions in one could impact others.

I plan to develop potential indicators from each condition that could be used to monitor the long-term direction of development and provide strategies for more effective governance.

Methodology

The methodology uses general morphological analysis (GMA) as a foundation (see: Johansen and, Ritchey), where a standard two-dimensional problem is transposed into a multidimensional problem space in which variables are arrayed against each other (e.g., matrix), yielding millions of potential outcomes. The framework composed of 13 dimensions and 43 conditions, will make up the body of the matrix (and the survey values will define the overall boundaries).

The GMA process aims to identify and structure all possible aspects and solutions for irreducible, complex problem spaces which in most cases involve human behavior, qualitative social issues, and political dynamics. The process is iterative through repeated sequences of analysis, refinement, and synthesis. I will conduct follow-on discussions/​interviews with subject-matter experts to refine the results.

Problem Formulation and Multidimensional Matrix

The first step requires an exact as possible formulation of the problem. Next, the problem must be broken down into a parameter set that frames the overall issue—in this case, 13 dimensions, with three to four possible outcomes (conditions) for each, totaling 43. The third step involves the construction of a multidimensional matrix that contains all solutions related to the problem (e.g., Figure-1, GMA Matrix, see Johansen). The matrix contains within itself the entire problem space of the given issue. The fourth step reduces any inconsistency between each pairwise condition.

Figure-1. Example—GMA Multidimensional Matrix. The dimensions are arrayed along the X and Y axis for cross comparative and consistency assessment. The example above is from Johansen’s work on military strategy

Next, a Cross-Consistency Assessment (CCA) winnows down the prospective futures to a smaller set of configurations, weeding out inconsistent pairs while discovering unexplored relationships. From the 13 unique dimensions, the 43 conditions that are assessed in the survey are arrayed along the horizontal (x) and vertical (y) axis to create the GMA matrix; each cell in the matrix represents a combination of two conditions (excluding duplicative or inconsistent values).

The cells are then evaluated to rank the consistency of each value pair, possible combinations (fast takeoff, moderate distribution), and probable (fast takeoff, innovation). The total 43 conditions yield millions of possible scenarios—too many to evaluate independently, but comprehensive and exhaustive (including outliers), which is the value of GMA.

Scenario Mapping

Each condition is then ranked on its assessed impact on stability and security (increase-decrease) and plausibility (likely-unlikely) through the above survey. We created a tool to calculate the impact and plausibility values provided by participants in the survey into a separate matrix for each value class (likelihood and impact).

The tool then averages all impact and likelihood rankings for each condition (all impact values ranked in the survey for condition 1, for example), and combines the ranking for each condition value pair in each cell where the two conditions meet to average the distributions (alternative probability calculations could be used such as conditional probabilities).

After all the separate values for each condition are calculated and combined, I will use a clustering algorithm to cluster like values into potential scenario classes. The clusters are user-defined, allowing thin distinct clusters for detailed analysis, or broad for more in-depth narratives. This method provides a systematic structuring of all independent, mutually exclusive, and interdependent variables of a problem to ensure that as many elements as possible are represented.

Next Steps

Ultimately, the GMA method is primarily used to decompose a multidimensional problem for alternative futures analysis. However, the pairwise relationships between each condition across the cells are also edges of a network graph, with each condition a node in the network. The graph can be a valuable tool to model the complexity, the strength of relationships, and possible directionality.

For example, the results from the CCA could be used to build a network graph to display the interdependent relationships between each variable and condition (from consultations with domain experts and based on the values from the ranking) to analyze the degree of dependency and relationships.

After the survey values are collected, all impact and likelihood values will be averaged for each condition in the GMA matrix. The next steps include:

  • Conduct a Cross-Consistency Assessment (CCA) between dimensions to examine unexplored relationships and ensure consistency of the value pairs.

  • Cluster the values into user-defined sets for analysis and scenario development.

  • Develop indicators for each condition that could potentially be used to monitor developments (TBD).

  • Develop a hypergraph model using the CCA output to evaluate the influence or possible directionality between conditions (TBD).

Note: the project is still in development and there are other methodological options and data sources I may integrate, if viable. This will be iterative, so I hope to conduct interviews with domain experts following the survey for further refinement.