Rethinking AI Intelligence: The REACT Framework

Author: Shawn Goodin

Abstract

This paper introduces the REACT AI Intelligence framework, a novel approach to categorizing and assessing artificial intelligence systems based on their cognitive intelligence capabilities. While existing AI frameworks primarily focus on autonomy and human-AI collaboration, REACT addresses a critical gap by examining the quality and sophistication of AI reasoning itself. By providing a cognitive perspective on AI maturity, REACT offers researchers, developers, and organizations a valuable tool for understanding, assessing, and advancing AI intelligence in a thoughtful, strategic, and responsible manner.

1. Introduction: The Need for a Cognitive Perspective on AI

The rapid advancement of artificial intelligence has created an urgent need for frameworks that can meaningfully categorize and assess AI capabilities. As AI systems become increasingly sophisticated, organizations and researchers require clear ways to understand, evaluate, and communicate about AI intelligence.

Current AI maturity models predominantly focus on three aspects:

• The degree of autonomy AI systems possess

• The nature of human-AI collaboration

• Organizational adoption and implementation processes

These frameworks provide valuable insights into how AI operates within organizational contexts and how humans interact with AI systems. However, they leave a critical gap in our understanding: they do not address the cognitive intelligence capabilities of AI systems themselves.

Consider the following questions that existing frameworks struggle to answer:

• How do we assess the quality and sophistication of AI reasoning?

• What distinguishes a cognitively advanced AI system from a basic one?

• How does AI intelligence evolve through distinct developmental stages?

• What are the internal cognitive mechanisms that enable increasingly sophisticated AI capabilities?

The REACT framework was developed specifically to address these questions by providing a cognitive perspective on AI intelligence maturity.

2. The Landscape of Existing AI Frameworks

2.1 Autonomy-Focused Frameworks

Most prominent AI frameworks focus primarily on autonomy and human-AI collaboration rather than cognitive intelligence:

Microsoft’s AI Maturity Model

• Assisted Intelligence: AI gives insights, humans decide

• Augmented Intelligence: AI boosts human decision-making

• Autonomous Intelligence: AI makes decisions on its own

PwC’s AI Augmentation Spectrum

• Advisor: AI suggests

• Assistant: AI helps

• Co-Creator: AI works with you

• Executor: AI handles tasks

• Decision-Maker: AI decides

• Self-Learner: AI improves itself

Deloitte’s Augmented Intelligence Framework

• Automate: AI takes over repetitive tasks

• Augment: AI enhances human decisions

• Amplify: AI scales human capabilities

Gartner’s Autonomous Systems Framework

• Manual: Humans only

• Assisted: AI supports

• Semi-Autonomous: AI does most, humans step in if needed

• Fully Autonomous: AI runs the show

MIT’s Human-in-the-Loop Model

• AI Automation: AI handles everything

• Human-in-the-Loop: Humans review or decide

• Human Override: Humans can step in when needed

HBR’s Human-AI Teaming Model

• Tool: AI provides insights

• Collaborator: AI shares tasks

• Manager: AI handles admin tasks

2.2 Common Patterns and Limitations

These frameworks share several common patterns:

• Focus on Autonomy: They primarily measure the degree of autonomy or independence from human oversight.

• Human-AI Relationship: They define progression in terms of changing relationships between humans and AI systems.

• Task Execution: Progress is measured by which entity (human or AI) performs tasks or makes decisions.

• External Behavior: They focus on observable behaviors rather than internal cognitive processes.

• Organizational Perspective: Most frameworks are designed from an organizational adoption perspective rather than a technical capability perspective.

While these frameworks effectively measure autonomy, human-AI collaboration, and organizational adoption, they do not address:

• Cognitive Sophistication: The quality and complexity of AI reasoning

• Intelligence Types: Different forms of intelligence exhibited by AI systems

• Cognitive Development: How AI intelligence evolves through distinct stages

• Internal Processes: The mechanisms by which AI systems process information

• Intelligence Quality: Measures of how well AI systems think, not just what tasks they perform

This gap highlights the need for a framework that focuses specifically on AI intelligence from a cognitive perspective, which is precisely what the REACT framework addresses.

3. Introducing the REACT Framework

The REACT framework represents a novel approach to categorizing and assessing AI systems based on their functional intelligence capabilities rather than their autonomy or human-AI collaboration patterns. It examines how AI systems themselves manifest different levels of cognitive capability, providing a framework for understanding what AI systems can actually do at different stages of intelligence maturity.

REACT stands for:

• Recognition (or Replicate in the original version)

• Evaluation (or Enhance in the original version)

• Analysis (or Assemble in the original version)

• Correlation (or Create in the original version)

• Thinking (or Transfer in the original version)

These five progressive levels represent a developmental path of AI cognitive capabilities, from basic pattern recognition to sophisticated abstract reasoning and cross-domain innovation.

Theoretical Foundations

The REACT framework is grounded in established cognitive science theories, particularly Marr’s three levels of analysis. As Yamins and DiCarlo (2016) note in their paper Using goal-driven deep learning models to understand sensory cortex, Marr’s framework provides a powerful approach for understanding intelligent systems by distinguishing between:

• Computational Level (what the system does and why): Maps to the purpose of each REACT level

• Algorithmic Level (how the system does it): Maps to the implementation approaches at each level

• Implementation Level (physical realization): Maps to the technical requirements for each level

This theoretical foundation aligns with recent research in cognitive science approaches to understanding AI systems. In Using the Tools of Cognitive Science to Understand Large Language Models at Different Levels of Analysis (Mahowald et al., 2023), the authors argue that cognitive science frameworks provide valuable perspectives for understanding AI capabilities and limitations.

By mapping each REACT level to Marr’s framework, the model gains theoretical coherence and academic validity. This alignment helps explain why certain capabilities naturally precede others and provides a principled basis for the progressive structure of the framework.

3.2 A Multidimensional Approach to Intelligence

Rather than defining intelligence by a single capability, REACT recognizes that intelligence manifests across multiple dimensions:

• Knowledge Representation: How information is structured and stored

• Reasoning Mechanisms: How the system processes information

• Learning Capacity: How the system improves over time

• Autonomy: Degree of independent operation

• Generalization: Ability to apply learning to new contexts

This multidimensional approach aligns with contemporary research on AI capabilities. As noted by Chollet (2019) in On the Measure of Intelligence, meaningful assessment of intelligence requires evaluating systems across multiple dimensions rather than on narrow task performance.

This approach enables a more nuanced assessment of AI systems and helps identify specific areas for improvement.

4. The Five Levels of the REACT Framework

4.1 Level 1: Recognition / Replicate

Cognitive Characteristics:

• Pattern Recognition: Identifying recurring structures in data

• Classification: Categorizing inputs based on learned features

• Imitative Learning: Reproducing behaviors or outputs seen in training data

• Supervised Learning: Learning from labeled examples

Real-World Examples:

• Auto-completion systems

• Template-based chatbots

• Basic content summarization

• Image classification systems

• Customer service response generators

Key Insight:

At this level, AI systems can recognize patterns and replicate known information but lack understanding of underlying concepts or the ability to adapt to new contexts.

4.2 Level 2: Evaluation / Enhance

Cognitive Characteristics:

• Comparative Analysis: Assessing relative quality or fit

• Criteria-Based Judgment: Evaluating against defined standards

• Optimization: Improving outputs based on feedback

• Preference Learning: Understanding what constitutes “better”

Real-World Examples:

• Email subject line optimization

• Content quality assessment

• Code improvement suggestions

• Marketing message refinement

• Product recommendation systems

Key Insight:

At this level, AI systems can evaluate and enhance existing content but are limited by predefined evaluation frameworks and may struggle to understand why certain options are better.

4.3 Level 3: Analysis / Assemble

Cognitive Characteristics:

• Decomposition: Breaking complex systems into components

• Relationship Identification: Recognizing connections between elements

• Causal Reasoning: Understanding cause-effect relationships

• Synthesis: Combining elements into coherent wholes

Real-World Examples:

• Sales funnel creation from multiple data sources

• Campaign orchestration

• Product design tools

• Diagnostic systems

• Multi-step planning systems

Key Insight:

At this level, AI systems can analyze complex information and assemble components into new structures but may miss subtle relationships or struggle with truly novel combinations.

4.4 Level 4: Correlation / Create

Cognitive Characteristics:

• Cross-Domain Connection: Linking information from different fields

• Pattern Recognition at Scale: Identifying higher-order patterns

• Generative Capability: Creating novel outputs

• Divergent Thinking: Exploring multiple possible solutions

Real-World Examples:

• Designing new visual brand identities

• Writing original creative content

• Composing music

• Cross-domain innovation

• Novel product design

Key Insight:

At this level, AI systems can create novel outputs and identify connections across domains but may generate plausible yet incorrect connections or exhibit variable output quality.

4.5 Level 5: Thinking / Transfer

Cognitive Characteristics:

• Abstract Reasoning: Thinking beyond concrete examples

• Analogical Transfer: Applying concepts from one domain to another

• Meta-Cognition: Awareness of own reasoning processes

• Autonomous Problem-Solving: Independent goal-setting and strategy

Real-World Examples:

• Using strategies from ant colony behavior to optimize cloud resource allocation

• Applying psychological models to UX design

• Scientific research assistants making novel connections

• Strategy development across disciplines

• Cross-domain innovation hubs

Key Insight:

At this level, AI systems can think abstractly and transfer knowledge across domains. This represents the most complex and challenging level to implement, with significant ethical considerations.

4.6 Progression Through REACT Levels

The REACT framework represents a developmental progression, with each level building upon the capabilities of previous levels. This progression is grounded in cognitive science principles that explain why certain capabilities naturally precede others:

Recognition → Evaluation:

• Development of internal quality metrics

• Integration of multi-dimensional criteria

• Shift from “what” to “how good”

Evaluation → Analysis:

• Development of causal understanding

• Ability to decompose complex systems

• Shift from assessment to explanation

Analysis → Correlation:

• Integration of knowledge across domains

• Development of transfer capabilities

• Shift from depth to breadth

Correlation → Thinking:

• Development of abstract reasoning

• Integration of creative capabilities

• Shift from specific to general

This progression aligns with cognitive development patterns observed in both human cognition and artificial intelligence systems, providing a principled basis for the framework’s structure.

5. Practical Applications of the REACT Framework

5.1 Assessment and Benchmarking

Organizations can use REACT to:

• Assess current AI capabilities against a clear developmental framework

• Benchmark systems against industry standards

• Identify specific cognitive capabilities to develop

• Create roadmaps for AI intelligence advancement

The multidimensional nature of the framework enables precise identification of capability gaps across different aspects of intelligence. Organizations can assess their systems across knowledge representation, reasoning, learning, autonomy, and generalization to pinpoint areas for improvement.

5.2 Strategic Planning and Development

REACT provides a structured approach to AI development by:

• Defining clear capability targets for each level

• Establishing measurable milestones for progress

• Guiding resource allocation based on capability gaps

• Informing make-vs-buy decisions for AI capabilities

By understanding the cognitive requirements of each level, organizations can make more informed decisions about AI development priorities and approaches.

5.3 Communication and Expectation Setting

The framework facilitates clearer communication about AI capabilities by:

• Providing a common language for discussing AI intelligence

• Setting realistic expectations about what systems can and cannot do

• Enabling more precise requirements definition

• Supporting more accurate marketing and positioning

This clarity helps bridge the gap between technical and non-technical stakeholders and reduces the risk of capability misrepresentation.

6. Ethical Considerations Across REACT Levels

Recognition Level Ethics

• Data Bias: Systems may perpetuate biases present in training data

• Privacy Concerns: Pattern recognition may reveal sensitive information

• Transparency Requirements: Need for clarity about what patterns are being recognized

Evaluation Level Ethics

• Value Alignment: Ensuring evaluation criteria reflect appropriate values

• Fairness in Assessment: Preventing discriminatory evaluations

• Explainability: Providing rationale for evaluative judgments

Analysis Level Ethics

• Causal Misattribution: Risk of identifying false causal relationships

• Responsibility for Recommendations: Determining accountability for analysis-based actions

• Complexity and Transparency: Balancing sophisticated analysis with understandable explanations

Correlation Level Ethics

• Spurious Correlations: Risk of identifying meaningless or misleading connections

• Creative Outputs: Questions of ownership and attribution

• Deception Potential: Ability to generate convincing but false information

Thinking Level Ethics

• Autonomy Boundaries: Determining appropriate limits to AI decision-making

• Unpredictability: Managing systems with emergent behaviors

• Human-AI Power Dynamics: Addressing potential shifts in control and authority

7. Future Research Directions

7.1 Measurement and Assessment Tools

• Development of standardized tools to assess AI systems against the REACT framework, including:

• Benchmark tasks for each level

• Standardized evaluation protocols

• Quantitative metrics for each dimension of intelligence

7.2 Developmental Pathways

• Research into how AI systems progress through REACT levels, including:

• Identifying critical capabilities that enable level transitions

• Understanding developmental bottlenecks

• Exploring alternative developmental sequences

7.3 Cognitive Architecture Integration

• Exploration of how the REACT framework can inform AI system design, including:

• Architectural requirements for each level

• Integration of multiple intelligence types

• Hybrid systems that combine different approaches

7.4 Cross-Domain Applications

• Investigation of how the REACT framework applies across different AI domains, including:

• Language models

• Computer vision systems

• Robotics

• Decision support systems

• Creative AI

8. Conclusion

The REACT framework represents a significant contribution to our understanding of AI intelligence maturity. By providing a cognitively grounded, multidimensional approach to categorizing AI capabilities, it fills an important gap in existing maturity models that primarily focus on autonomy and human-AI collaboration.

As AI systems continue to advance in sophistication, frameworks like REACT become increasingly important for guiding development, setting appropriate expectations, and ensuring responsible implementation. By focusing on the cognitive capabilities of AI systems themselves, REACT provides a valuable complement to existing frameworks and a foundation for more nuanced discussions about AI intelligence.

The framework’s grounding in cognitive science principles provides both theoretical validity and practical utility, making it a valuable tool for researchers, developers, and organizations navigating the complex landscape of artificial intelligence.

References

• Yamins, D. L., & DiCarlo, J. J. (2016).

  • Using goal-driven deep learning models to understand sensory cortex.

  • Nature Neuroscience, 19(3), 356-365.

  • DOI: 10.1038/nn.4244

• Mahowald, K., Ivanova, A. A., Blank, I. A., Kanwisher, N., Tenenbaum, J. B., & Fedorenko, E. (2023).

  • Using the Tools of Cognitive Science to Understand Large Language Models at Different Levels of Analysis.

• Chollet, F. (2019).

  • On the Measure of Intelligence.

• Bhatt, S., Jain, P., Niu, Y., Geyik, S. C., Kenthapadi, K., & Wang, H. (2024).

  • PyRIT: A Framework for Security Risk Identification and Red Teaming in Generative AI Systems.*

• Raji, I. D., Smart, A., White, R. N., Mitchell, M., Gebru, T., Hutchinson, B., … & Barnes, P. (2023).

  • Auditing Large Language Models: A Three-Layered Approach.

GitHub Repositories

• KlotzJesse/neuralnetwork

  • Description: Java implementation of neural network architecture

• P-Programist/AI-ML-Projects

  • Description: Collection of AI and machine learning projects

• Aryia-Behroziuan/Overview

  • Description: Repository focused on knowledge representation in AI

Organizational Sources

• Microsoft. (2023). AI Maturity Model.

  • Description: Microsoft’s framework for assessing organizational AI maturity

• PwC. (2022). AI Augmentation Spectrum.

  • Description: PwC’s framework for human-AI collaboration

• Deloitte. (2023). Augmented Intelligence Framework.

  • Description: Deloitte’s framework for AI’s role in human productivity

• Gartner. (2023). Autonomous Systems Framework.

  • Description: Gartner’s framework for categorizing AI involvement in work

• MIT Sloan Management Review. (2023). What’s Your Company’s AI Maturity Level?

  • Description: MIT’s framework for assessing organizational AI maturity

• Harvard Business Review. (2022). Human-AI Teaming Model.

  • Description: HBR’s model for AI as a teammate rather than a replacement

• Accenture. (2023). AI Maturity and Transformation.

  • Description: Accenture’s framework for AI maturity assessment