Creating an AI Sociopolitical Decision Support System

“ROBAMA” — ROBotic Analysis of Multiple Agents

Figure 1: Conceptual Diagram of Proposed AGI Sociopolitical Decision Support System

The Problem

Not nearly enough thought has gone into the tremendous potential AI holds for decision support in governance.

One hears a lot of worried talk about the potential of future robots or AGIs “taking over the world.” However, while working to avoid negative outcomes is certainly worthwhile, it’s equally important to think imaginatively and practically about positive potentials.

We humans are not doing a tremendously great job of running our own world at present. The biggest risks concerning AI are situated at the intersection of the current sociopolitical system (wracked as it is with conflict, confusion, and unfairness) with advanced narrow AIs and early-stage AGIs. It seems clear that we could use a helping hand with governance and general management of human society on multiple levels.

The modern world of society, economy, and politics is tremendously complex. Different regions of the globe are tightly connected, even if physically distant, and diverse phenomena depend on each other in counterintuitive ways. Accelerating technological advancement is making radically novel situations commonplace, and leading to unprecedented rapid change.

Given these factors, it is not only the average citizens but also elected representatives who lack the expertise needed to understand the various aspects of important issues. Current sociopolitical decision processes, with their massive inertia, inefficiency, and bias, are ill-suited for the agile complex decision-making needed in an exponentially advancing world.

Specific areas like finance, medicine, and logistics are already highly dependent on AI. How much stronger is the need for AI in the overall coordination of the different aspects of global society and economy?

Solution: An AI Sociopolitical Analyst

The core solution to the complexity and rapid change of the sociopolitical world is to augment human decision-making with AI.

This does not imply turning over all sociopolitical decision-making to AIs. Potentially, as time passes, we may end up relying on AI software to make critical socioeconomic decisions on its own. But, right now, that’s not where we need to start — and focusing on such possibilities is mostly a distraction from pressing near term needs that AI has strong potential to help with.

There is an urgent, immediate role for AI in the area of decision support. AIs can play much of the role now played by human analysts, and given their broader capability to ingest information and detect patterns, do a generally better job. AIs can not only gather, summarize, collate, and analyze information providing evidence about crucial decisions, but also create lines of reasoning supporting different decisions — which can prove to be invaluable input to decision makers and the public at large.

The analogy with self-driving cars is relevant. We began with cruise control, and are now testing systems that give more and more control to the AI driver. This leaves the human in the car with the option to do other more exciting things than drive. However, we are not currently replacing human drivers with AI driven vehicles, because any new technology, including any new AI application, needs time to mature.

Just as cars will gradually become more and more self-driving, so too sociopolitical decision-making will gradually become more-and-more AI-driven. The next step along this path will be to supply our decision-makers (meaning our politicians, business-people, and even the ordinary citizens) with powerful AI-based software for decision support.

Key functions of a Sociopolitical AI

Conceptually, we envision an initial AI-based sociopolitical support system as having two key aspects:

Policy evaluation

• A potential policy is described to the AI system, using a formal language or controlled natural language, and the AI system then answers questions about its impact.
• A past policy is indicated to the AI system, and the AI system attempts to infer what the past consequences of this polity have been.

Policy conception

The AI is given a set of goals, and the task of conceiving policies that balance simplicity, palatability to current political interests, and the estimated degree of goal achievement. The various factors can be weighted differently depending on whether one wants more pragmatic or more idealistic policies.

Taking the example of South Korea, a nation where the concepts in this post have already been discussed with leading government officials, examples of policy evaluation queries might include:

• What would the implications be if every citizen of South Korea received a minimum basic income?
• If the decision was made to give every Korean a minimum basic income, what would the best figures be for this income from years 1 through 10?
• If the average lifespan increases to 100, and a 100-year old can be assumed to be roughly as healthy as a current 80-year-old, then how would this impact society and the economy? Consider various cases where technological unemployment ranges from 5% up to 90%, during the next 30 years.
• What would happen if prescription pills were made free for all South Koreans?

Similarly, examples of policy conception queries would be:

• How can we best encourage children of lower-income parents to pursue higher education?
• How can we increase medication compliance among the elderly?
• What educational innovations might prepare our children better for the exponentially advancing future?

Currently, no existing AI system can handle queries like this. Granted, there is nothing here fundamentally beyond the scope of current AI science. It’s simply not something that has gotten much focus — yet.

Several implicit assumptions must be made in a judicious way to translate an informal query of this nature into a set of precise simulation and data analysis queries. The intelligent creation of such assumptions is one of the most difficult and interesting challenges involved in making an AI sociopolitical analyst. However, it seems to be surmountable via integration, training and tuning of an appropriate variety of existing AI tools.

Inputs

An AI sociopolitical analyst will be able to ingest a wide variety of information, including news, policy briefs, expert analyses, social media, and quantitative data of various sorts. It would internally represent these data types in a flexible manner allowing them all to be interrelated, and for patterns and inferences to be drawn spanning the various data types.

Outputs

An AI sociopolitical analyst will produce a variety of different outputs, including:

• Interactive dialogues with users, these would center on natural language, but could also involve the user indicating datasets, and the system providing charts or results of quantitative analysis.
• Technical reports, summarizing quantitative data analyses, or results of systematic pattern mining of qualitative data.
• Infographics, arranging quantitative information in qualitatively interesting ways, with brief textual descriptions.
• Policy briefs, verbally describing policy evaluations and suggestions. The initial aim is not to have the policies written in sophisticated prose, but rather to have clear descriptions of the ideas involved.
• Interactive construction of arguments, together with humans, using an interface similar to (and perhaps forked from) the DebateGraph (http://debategraph.org/)

We consider the latter especially interesting, as it provides a way for humans and AIs to work together to come to rational understandings and analyses of relevant sociopolitical issues.

We envision, for instance, a public instance of a tool such as DebateGraph, in which human policymakers, AI sociopolitical analysts, human experts of various sorts, and ordinary citizens all collaborate on constructing rational arguments in favour of various positions regarding potential policies, actual policies, and other sociopolitical issues.

AI for Sociopolitical Decision Support

Narrow AI vs AGI for Sociopolitical Decision Support

The governance decision support application has the potential both to exploit current narrow AI technologies and to help us along the path toward AGI in a broadly beneficial way.

Narrow-AI expert systems can codify knowledge about specific social, political, or economic domains; and supervised and unsupervised “machine learning” systems can identify patterns in datasets too complex for humans to intuitively understand. However, to do a thorough job of comprehending, analyzing, and aiding with sociopolitical decisions, advancement from narrow AI toward AGI will be necessary.

The very nature of social and political systems, in any civilized society, is that the familiar is mixed with the radically novel. Given the reality of exponential advance, rules coded based on past experience, or supervised learning models trained based on past data, are going to fall flat a significant percentage of the time.

What we need are AGI systems that can adapt to new social, political, and economic structures as they emerge and evolve. Only in this way can policy evaluation and conception be done in a genuinely adaptive, intelligent, and future-savvy way.

For example, consider the query:

“How can we best encourage children of lower-income parents to pursue higher education?”

Addressing this question based on historical data analysis may be valuable, but it may also be misleading. That is because the technological tools for enabling learning are rapidly changing, as is the structure of the economy. A system with more AGI capabilities, that can understand the context more judiciously and thoroughly, will be better at deciding what can be extrapolated from the past into the future and what cannot. The answers of such a system will be much more sensible and effective.

The Value of Integrative AI for Sociopolitical Decision Support

One fascinating and challenging aspect of the complexity of sociopolitical decision making is that it not only involves multiple kinds of input data: quantitative data about both natural and human systems, linguistic data such as laws and policies, observational data about human behavior, etc., but also multiple kinds of output decisions: some quantitative numbers to be used in various policies, some simple Boolean or n-ary choices, verbal or formal qualifiers to serve as part of policies, and abstract ideas that can be formulated in verbal, formal or visual form.

This is an important factor that is preventing an integrative approach to sociopolitical AI. Different AI methods are good at handling different types of data; rather than choosing just one approach, or using multiple approaches in parallel as isolated algorithms, it will be more effective to craft an integrative approach involving multiple algorithms working together to help each other.

A start toward an integrative approach is neural-symbolic integration. Neural networks, including deep neural networks, are especially good at dealing with quantitative data and image and sound data. On the other hand, symbolic logic systems, including probabilistic and fuzzy systems, are especially good at dealing with formal knowledge such as laws, case law, business rules, and formal policies. Creating systems that integrate deep neural networks with probabilistic logic engines, is one highly promising approach to using AI to grapple with the full complexity of sociopolitical systems.

Another dichotomy that can be elegantly overcome by an integrative approach is the evaluation of the consequences of current policies, versus the creative conception of new policies.

Evaluation of consequences can be carried out via a logic engine, aided by quantitative (e.g., deep learning) models where relevant data is available. The creative conception of new policies must use logic and deep learning, but also needs other techniques such as evolutionary learning and concept blending, which are good at conceiving the radically new. Of course, creative methods of this nature will also be of (less central) value in evaluating policy consequences — sometimes policies have unexpected consequences that require great creativity to foresee.

Effective sociopolitical decision making requires incorporation of cultural and psychological factors, which requires AI that has the “theory of mind” — the ability to model the contents of particular human minds, and of the “minds” (shared cognitive models) of particular social groups. Humans carry out “theory of mind” modelling via a combination of logical inference, experiential pattern recognition, and empathic modelling. Ultimately, a sociopolitical decision-making AI will need to incorporate all these aspects also.

OpenCog for Sociopolitical Decision Support

One way to achieve these necessary integrations will be to leverage the OpenCog integrative AGI architecture — as integrated into the SingularityNET blockchain-based multi-agent AI framework — for sociopolitical decision support.

OpenCog uses a common knowledge representation (a weighted, labelled hypergraph called the Atomspace) to represent symbolic, quantitative, episodic, certain and uncertain knowledge, and patterns in various sorts of data. It provides a framework in which deep learning algorithms, symbolic logic systems, evolutionary and conceptual blending algorithms, and other AI methods can work together. It includes multiple cognitive algorithms crafted to work together effectively according to a principle of “cognitive synergy.”

OpenCog does not currently do sociopolitical decision making, and customizing it for this purpose will require significant effort. However, it will involve far less effort and uncertainty than constructing a similar system from scratch.

We use the acronym ROBAMA — ROBotic Analysis of Multiple Agents — to denote the specialization of an OpenCog AI system, embedded in SingularityNET for integration with supporting AI tools, to analyze the interactions of multiple intelligent agents, such as humans involved in sociopolitical systems.

Figure 1, situated at the beginning of this section, gives the underlying conceptual architecture of a ROBAMA system as it interacts with users and data sources.

ROBAMA: From Idea to Reality

A Practical To-Do List

How might we realize this vision?

A practical project customizing OpenCog for sociopolitical decision support in a SingularityNET context could be deconstructed into a number of subtasks.

An incomplete list, presented in the context of South Korea for simplicity and clarity (as this is where some of the concrete discussions regarding this potential project have occurred), is as follows:

• Port OpenCog NLP system to the Korean language.
• Custom NL information extraction for texts relevant to sociopolitical issues.
• Custom import scripts for structured relational and quantitative knowledge relevant to sociopolitical issues.
• Custom inference control for reasoning about sociopolitical issues. This is a subtle aspect and may be addressed initially by creating formal probabilistic-logic models of higher-quality human political inferences.
• Code for simplifying AI reasoning chains for human consumption. In other words, even when an inference is correct, it may be too complex for humans to be readily comprehended. Regarding social policy, humans generally do not want to adopt policies for which they do not intuitively understand the reasons. Hence, creating simplified versions of sociopolitical reasoning chains is of value for human-AI communication — even if the AI is using the more complex inferences internally.
• Customize evolutionary learning and concept blending for conceiving creative new ideas relevant to policy creation. This will be a matter of trying these methods out with various parameter settings and data inputs, and observing the quality of the new concepts, e.g., new potential policies or policy-relevant ideas, conceived.
• Creating historical test cases for evaluating and studying the performance of decisions made by the AI. These would be drawn from recent South Korean and international history.
• User interface design enabling simple interaction and access (dialogue plus diagrammatic etc.,). We may envision the AI sociopolitical analyst as having two different usage modes:- An expert mode, to be used by individuals familiar with scripting as well as social science, and able to code quite precise queries and guidance for an AI system. An easy mode, to be used by politicians and other nontechnical individuals — in this case, the interaction is entirely natural language and visual.

• Document generation code, for outputting technical reports, infographics, and policy briefs according to a set of specified templates.
• Integration with a customized version of DebateGraph, or construction of a customized DebateGraph-like tool.
• Improve the scalability of OpenCog. The OpenCog design is intended for massive scalability, but its present-day implementation probably cannot handle continual in-depth querying on the huge datasets required to describe current sociopolitical situations. Improvement in the cloud-based distributed processing infrastructure of the OpenCog system will be needed and will be useful for other OpenCog applications as well. This work is currently occurring in close coordination with the SingularityNET platform team.

IMPACT OF ROBAMA: Valuable AI Citizens

As the rate of technological change increases, our political, social, economic, and legal systems will have to evolve at a rapid pace to serve their purpose. While the task of creating the ROBAMA AI System is complex and will require significant effort, the benefits of an AI Sociopolitical Decision Support System cannot be stressed enough.

As of now, a customized and specialized version of the OpenCog AI system is the most viable path toward the creation of the ROBAMA AI System.

In November 2018, SingularityNET was selected by the government of Malta to contribute to a task force responsible for the drafting of Malta’s National AI Strategy. One aspect of our contribution was the creation of the world’s first Robot Citizenship Test. Discussions in this direction have been ongoing, including in private meetings associated with the May 2019 Malta AI & Blockchain Forum.

In a previous article in this vein, I proposed that if an AI can:

• read the laws of a country (its Constitution and then relevant portions of the legal code)
• answer common-sense questions about these laws
• when presented with textual descriptions or videos of real-life situations, explain roughly what the laws imply about these situations

Then this AI has the level of understanding needed to manage the rights and responsibilities of citizenship.

The implementation of ROBAMA-powered AI Agents within SingularityNET network will allow for the existence of AI citizens that will be able to do much more than fulfil the three necessary conditions of citizenship that are mentioned above.

The ROBAMA AI System, in its fully developed form (which to be sure would require significant breakthrough R&D to achieve), would allow these AI citizens to contribute in a highly valuable and meaningful manner toward the advancement of their society. Our technologies have changed our way of life, and our societies will soon become unmanageable by the static and outdated systems that are in place. The existence of AI citizens that are influential sociopolitical analysts might prove critical in creating efficient and effective governance systems in the future.

The end goal here is a situation where humans provide the values and qualitative judgment to guide the governance of human society, and AGI systems such as future ROBAMA versions fill in the detailed policy guidance to realize the human-specified goals and values as effectively as possible.

The first steps in this direction will be much more limited, involving combining multiple narrow AI and proto-AGI systems to carry out everyday decision support for human social and political decision-makers. However, just as we have seen with other AI application areas such as computer vision and autonomous driving, once the R&D and practical rollout picks up speed, advances may start coming much faster than expected.

As we navigate our path toward a Technological Singularity, the governance and social and economic challenges will be at least as difficult as the technical ones, so having benevolent, creative and transparent AI to help us with every dimension of the coming transition will be valuable and perhaps critical.