Abstract
This paper proposes a research direction to advance AI which draws inspiration from cognitive theories of human decision making. The premise is that if we gain insights about the causes of some human capabilities that are still lacking in AI (for instance, adaptability, generalizability, common sense, and causal reasoning), we may obtain similar capabilities in an AI system by embedding these causal components. We hope that the high-level description of our vision included in this paper, as well as the several research questions that we propose to consider, can stimulate the AI research community to define, try and evaluate new methodologies, frameworks, and evaluation metrics, in the spirit of achieving a better understanding of both human and machine intelligence.
Introduction
Despite the transformative success of Deep Learning, modern Artificial Intelligence remains largely "narrow." While current models excel at specific tasks such as image classification or natural language translation, they frequently lack the robust generalizability, causal reasoning, and common sense inherent in human cognition. This position paper, authored by a collaborative team from IBM and several academic institutions, outlines a strategic roadmap to bridge this gap by drawing on cognitive science—specifically the dual-process theory of human thought.
The Problem: The Tension Between Connectionism and Symbolism
A central debate in AI research pits end-to-end neural network approaches (connectionism) against symbolic, logic-based AI. Neural networks are adept at learning from vast amounts of raw data but operate as "black boxes" that struggle with transparency and rigorous logic. Conversely, symbolic AI is transparent and excels at structured reasoning but often fails when faced with noisy, real-world data.
The authors argue that neither paradigm is sufficient in isolation. The challenge lies in creating a cohesive architecture that integrates the "intuitive" capabilities of machine learning with the "analytical" rigor of symbolic logic into a single, functional system.
The Conceptual Framework: System 1 and System 2
The proposed framework is modeled after Daniel Kahneman’s "Thinking, Fast and Slow," which describes two distinct modes of human thought:
- System 1 (Thinking Fast): This represents intuitive, unconscious, and rapid decision-making, such as recognizing a familiar face. In the context of AI, this corresponds to Machine Learning (ML) and deep neural networks.
- System 2 (Thinking Slow): This represents effortful, logical, and slow processing used for complex problem-solving, such as calculating a mathematical proof. In AI, this corresponds to Symbolic Logic, Planning, and Optimization.
The authors advocate for a Neuro-Symbolic approach where these two systems work in tandem. A critical component of this architecture is Governance—a meta-layer responsible for determining when the AI should rely on its fast, intuitive model and when it must "slow down" to employ formal logic for the sake of accuracy, safety, or ethical alignment.
Key Insights for Hybrid Intelligence
The paper identifies several pillars necessary for this integration:
1. Symbiosis and Heuristics
System 1 is not merely a precursor to System 2; it serves as a vital heuristic engine. In complex environments with massive search spaces (e.g., chess or scientific discovery), System 1 intuitively narrows the field of possibilities, allowing System 2’s logical engine to focus its computational resources on the most promising paths.
2. Introspection and Switching
For a hybrid system to function, it requires "introspection." The AI must possess the meta-cognitive ability to recognize its own limitations—identifying when a problem is too complex, high-stakes, or novel for the intuitive System 1 to handle alone, thereby triggering a System 2 intervention.
3. Ethics as a System 2 Override
The authors suggest that ethical behavior can be formalized as a System 2 process. While a System 1 model might generate a biased or "fast" response based on patterns in its training data, a System 2 layer can evaluate that response against explicit ethical constraints and logical rules, overriding the output if it violates predefined safety or fairness standards.
4. The Sequentiality Bottleneck
In humans, System 2 is sequential; we generally focus on one complex task at a time. The authors pose a provocative research question: Is this sequential nature a biological limitation to be overcome with parallel processing in AI, or is the bottleneck itself a necessary feature for effective, focused reasoning?
A Roadmap for Future Research
As a position paper defining a new research direction, this work does not present empirical data or specific datasets. Instead, it provides a structured set of 10 Research Questions intended to galvanize the community. These questions focus on:
- Defining metrics for hybrid System 1/System 2 AI.
- Modeling the governance mechanisms that manage the switch between systems.
- Exploring "Society of Mind" architectures where multi-agent systems represent specialized cognitive skills.
Conclusion
The authors conclude that the "integration route"—combining the strengths of connectionist and symbolic AI—represents the most viable path toward Artificial General Intelligence (AGI). While the implementation of a seamless governance mechanism remains a significant engineering challenge, moving toward "consciousness-inspired" AI that builds subjective models of the world is the next frontier in achieving human-like intelligence.
Summary updated