A Concept Paper for an Experimental Research Program
Executive Summary
Human beings have accumulated enormous knowledge about human behavior, yet our ability to explain, anticipate, and prevent large-scale political conflict remains limited. Existing artificial intelligence systems can retrieve, synthesize, and reason over much of this accumulated knowledge. A more ambitious possibility deserves systematic investigation: Can AI systems generate genuinely novel hypotheses about human behavior through abductive reasoning, test those hypotheses against historical evidence, and improve them through adversarial interaction with other AI systems?
This proposal calls for an interdisciplinary experimental program to investigate that question.
The project would not begin by asking an AI to reproduce established theories of psychology, sociology, political science, or history. Instead, AI systems would be given carefully selected historical evidence and asked to identify surprising relationships, propose explanatory hypotheses, derive their implications, and subject those hypotheses to systematic attempts at falsification. Independent AI systems would then critique one another, generate competing explanations, and test those explanations against additional historical cases.
If this process produces hypotheses that are demonstrably novel, empirically supported, and more explanatory or predictive than existing models, the project would have demonstrated a potentially important new form of AI-assisted scientific discovery.
A second phase would place competing behavioral models into adversarial simulations of political crises. The objective would not be to predict a particular political event or to help one political faction defeat another. Instead, the simulations would investigate how societies move toward escalation, stabilization, de-escalation, and peaceful resolution, and which interventions appear to increase democratic resilience and reduce the probability of political violence.
The immediate motivation is the possibility of severe political instability in the United States, including during the period surrounding the November 2026 elections. However, the proposed research should not assume that civil conflict is inevitable or even probable. Its purpose should be precisely the opposite: to determine, through evidence and adversarial analysis, which dangerous scenarios are plausible, which are not, and what conditions favor peaceful outcomes.
1. The Central Research Question
The foundational question is:
Can cooperative and adversarial AI systems discover empirically testable knowledge about human behavior that is not merely retrieved or recombined from explicitly known human theories?
This question is inspired in part by Charles Sanders Peirce's conception of abduction: the generation of an explanatory hypothesis in response to an observation that requires explanation.
The proposed research would distinguish three activities that are frequently blended together in present-day AI systems:
Abduction: What might explain this observation?
Deduction: If that explanation were true, what else should we observe?
Empirical testing: Does the additional evidence support or contradict those consequences?
The AI would therefore not be judged primarily on whether it can produce persuasive explanations. It would be judged on whether its explanations survive attempts to prove them wrong.
That distinction is fundamental.
A system capable of producing thousands of plausible theories is not necessarily a scientific discovery system. A system capable of producing hypotheses that repeatedly survive independent empirical testing would be something considerably more consequential.
2. The Novelty Problem
A central difficulty is determining whether an AI has actually generated a new idea.
An LLM has absorbed an enormous amount of human intellectual production. It may therefore produce an apparently original hypothesis that is actually a recombination of existing ideas encountered during training.
The project should therefore explicitly investigate several levels of novelty:
- Rediscovery: The AI independently produces an explanation already well established in human scholarship.
- Recombination: The AI combines known concepts in a new formulation.
- Novel hypothesis: The AI proposes a hypothesis not readily identifiable in the existing literature.
- Novel and empirically useful hypothesis: The AI produces a previously unrecognized hypothesis that explains historical observations and generates successful predictions or discriminating tests.
The fourth category would constitute the most important result.
The objective is not to prove that AI possesses some mystical form of machine intuition. The objective is to determine experimentally whether AI systems can participate meaningfully in the generation of new explanatory knowledge.
3. Proposed Experimental Architecture
The detailed architecture should be developed by a multidisciplinary team rather than prescribed in advance. Nevertheless, the experiment can be conceived as a series of interacting functions.
A. Historical Evidence Engine
The system receives carefully selected historical records concerning human behavior, political movements, institutional crises, social conflicts, cooperation, violence, reconciliation, and political transitions.
The data should be divided into separate discovery and testing sets so that an AI cannot generate a hypothesis from a historical case and then "confirm" it using the same evidence.
B. Abductive AI
One or more AI systems examine observations and generate candidate explanations.
They should be explicitly encouraged to consider explanations that differ from established theories rather than merely summarize existing scholarship.
Each proposed hypothesis should be required to state:
- what it claims;
- what observations motivated it;
- what mechanism it proposes;
- what evidence would support it;
- what evidence would contradict it;
- and what additional observations should follow if it is correct.
C. Adversarial AI
A separate system is assigned the task of attacking the hypothesis.
It searches for:
- historical counterexamples;
- alternative explanations;
- hidden assumptions;
- selection effects;
- confounding variables;
- logical inconsistencies;
- and situations in which the hypothesis makes incorrect predictions.
The adversarial system should have no incentive to preserve the originating hypothesis.
D. Independent Hypothesis Generator
A third system, preferably operating independently, generates alternative explanations for the same evidence.
This prevents the project from becoming an iterative exercise in polishing the first plausible theory.
E. Empirical Evaluator
The competing hypotheses are tested against historical cases that were not used during their generation.
The evaluation should be as quantitative as the subject permits.
The system should record not merely successful predictions but failures.
A hypothesis that predicts nine things correctly and misses one should not be presented as having predicted ten.
F. Human Review
Historians, psychologists, political scientists, philosophers of science, statisticians, and AI researchers would independently evaluate the strongest hypotheses.
Importantly, human experts should also be permitted to conclude that an apparently novel AI discovery is simply a rediscovery of an existing theory.
4. Cooperative Training Through Adversarial Interaction
The term "cooperative" in this project should not mean that all AI systems are trained to agree.
It should mean that different systems cooperate in the pursuit of knowledge by disagreeing systematically.
One possible cycle would be:
Observation → Hypothesis → Criticism → Alternative Hypothesis → Deduction → Historical Test → Failure Analysis → Revised Hypothesis
This process could be repeated across thousands of historical cases.
The result would resemble an artificial research community in which different AI systems perform different epistemic roles.
One system proposes.
Another attacks.
Another seeks counterexamples.
Another searches for mathematical structure.
Another compares competing causal models.
Another evaluates whether the evidence actually distinguishes among them.
A final system attempts to determine whether the apparent discovery is genuinely new and empirically meaningful.
The purpose would be to make intellectual disagreement a computational resource.
5. From Understanding Human Behavior to Simulation
Only after behavioral models have demonstrated some empirical credibility should they be introduced into large-scale simulations.
The second research phase would construct adversarial simulations of political and social crises.
Rather than assuming that a particular outcome is inevitable, the simulation would explore multiple pathways:
stability → stress → polarization → escalation
but also:
stress → institutional adaptation → negotiation → de-escalation
and many intermediate possibilities.
Competing AI models would represent different assumptions about human motivation, group behavior, institutional legitimacy, information, status, fear, material incentives, and collective action.
The models would then compete in controlled environments.
The purpose would be to determine which models produce simulations that more closely reproduce historical experience and whether they reveal previously underappreciated mechanisms of escalation or stabilization.
6. The November 2026 Application
The immediate motivation for the project is concern about the possibility of serious political instability in the United States during the period surrounding the November 2026 elections.
The research should not assume that civil war, widespread political violence, or democratic breakdown will occur.
Instead, November should serve as a real-world stress test for a broader question:
What can be learned about preventing political violence and preserving peaceful democratic competition under conditions of severe political stress?
The simulations could examine alternative interventions involving:
- civic and community organization;
- public communication;
- institutional coordination;
- protection of vulnerable populations;
- election administration resilience;
- information integrity;
- conflict de-escalation;
- peaceful demonstrations;
- mechanisms for resolving disputes;
- and preservation of legitimate constitutional processes.
The objective would be to identify strategies that maximize the probability of peaceful political resolution and minimize escalation.
The simulations should also be required to identify what would cause their own recommendations to fail.
This is crucial. A system that tells decision-makers only what they want to hear is dangerous precisely when the stakes are highest.
7. A Principle of Democratic Neutrality
Because the application concerns politically contentious circumstances, the project should establish a strong neutrality principle.
The AI should not be tasked with determining which political faction is morally or politically entitled to prevail.
Instead, the optimization target should be things such as:
- preservation of lawful democratic institutions;
- peaceful political competition;
- reduction of political violence;
- protection of civilians;
- truthful information;
- lawful exercise of political rights;
- institutional continuity;
- and peaceful transfer or retention of political authority through legitimate constitutional mechanisms.
This distinction is important both ethically and scientifically.
The project should be designed to reduce the probability of catastrophic political outcomes, not to become an AI strategist for one side of a political conflict.
8. The Most Important Safeguard: The AI Must Be Allowed to Say "No"
A central design principle should be institutionalized from the beginning:
The system must be rewarded for discovering that its assumptions are wrong.
For example, if researchers believe that a particular political crisis is likely to escalate, the AI should be explicitly tasked with constructing the strongest evidence-based argument that escalation is unlikely.
If two explanations compete, the system should be rewarded for identifying the evidence that would distinguish them.
If the available evidence is insufficient, the correct answer should be:
Insufficient evidence.
This may sound mundane, but it is one of the most important requirements for a system intended to operate in a politically charged environment.
The project should measure not only predictive accuracy but also calibration, uncertainty, falsification ability, and willingness to revise conclusions.
9. What Would Constitute Success?
The project should establish meaningful success criteria before beginning.
Possible milestones include:
Stage 1
AI systems demonstrate reliable generation of hypotheses that are distinguishable from simple retrieval or summarization.
Stage 2
AI-generated hypotheses survive independent historical testing.
Stage 3
Some AI-generated hypotheses demonstrate explanatory or predictive value beyond established baseline models.
Stage 4
Adversarial AI systems reliably identify weaknesses in both human and AI-generated theories.
Stage 5
Behavioral models produce simulations that reproduce known historical patterns better than simpler models.
Stage 6
The simulations identify interventions associated with greater stability and reduced escalation in previously unseen scenarios.
Stage 7
Independent researchers reproduce the results.
The project should be considered unsuccessful if it merely produces impressive-looking narratives.
10. Required Research Community
This cannot responsibly be an AI-engineering project alone.
The proposed research community would ideally include:
- AI researchers and software engineers;
- historians;
- philosophers of science;
- psychologists;
- political scientists;
- economists and game theorists;
- statisticians and causal-inference specialists;
- conflict researchers;
- democratic-institution experts;
- security and safety researchers;
- and independent red-team organizations.
The engineers would build the experimental environment.
The historians would challenge the interpretation of historical evidence.
The philosophers would examine the epistemology of abduction and scientific discovery.
The statisticians would test whether apparent discoveries actually survive quantitative scrutiny.
The political and conflict researchers would challenge assumptions about collective behavior.
The red teams would attempt to break the entire system.
No single discipline should be permitted to become the project's intellectual gatekeeper.
11. Why This Research Is Worth Attempting
The argument for attempting this project does not depend on believing that AI will soon acquire a complete theory of human nature.
The argument is simpler.
Human societies face problems whose complexity may exceed the capacity of any individual or existing institution to reason about them comprehensively.
Meanwhile, AI systems are acquiring increasingly powerful capabilities for abstraction, simulation, mathematical reasoning, pattern discovery, and interaction among multiple agents.
It is therefore reasonable to ask whether these capabilities can be organized into something resembling a scientific discovery process for human behavior.
The experiment may fail.
The AI may merely rediscover existing theories.
It may generate attractive but false explanations.
Historical data may prove too ambiguous.
Simulations may turn out to be unreliable.
Or the project may reveal that some apparently profound human regularities cannot be modeled adequately.
Those outcomes would still constitute valuable scientific knowledge.
But there is also an upside.
If AI systems can generate genuinely novel hypotheses, subject them to hostile criticism, test them against history, and improve them through repeated empirical confrontation, we may have discovered a new instrument for investigating one of the oldest scientific problems:
Why do human beings behave as they do, particularly when they form groups, encounter uncertainty, experience fear, compete for status and resources, and confront institutions that they perceive as legitimate or illegitimate?
That knowledge could have applications far beyond the immediate American political crisis.
Conclusion
The proposed project begins with a modest question and points toward an extraordinary possibility:
Can machines participate in the discovery of new knowledge about human nature rather than merely consuming and reorganizing the knowledge humans have already produced?
The answer should not be assumed.
It should be experimentally demonstrated or disproved.
The most promising route is an adversarial cooperative architecture in which AI systems generate hypotheses, attack hypotheses, construct alternatives, derive predictions, confront historical evidence, and repeatedly revise their models.
If successful, the resulting behavioral models could then be placed into controlled simulations of political crisis to investigate the conditions under which societies escalate toward violence or instead maintain institutional continuity and peaceful political competition.
Given the possibility of severe political instability in the United States and elsewhere, waiting until existing institutions are already overwhelmed would be an unnecessarily expensive experiment.
The appropriate response is not to assume catastrophe.
It is to build better instruments for thinking about catastrophe before catastrophe arrives.
This proposal is therefore an invitation to AI researchers, historians, social scientists, philosophers, statisticians, and democratic-institution experts to determine whether such an instrument can actually be built.
