The Functional Turn in AI Classification: Understanding ABAF

The Functional Turn in AI Classification: A Critical Analysis and Strategic Assessment of the Approach-Based AI Framework (ABAF)

From the Desk of BeResponsibleAI

Author - Dr. Sharad Maheshwari, imagingsimplified@gmail.com

Executive Summary

Artificial intelligence has outgrown the vocabulary used to describe it. The long-standing divide between machine learning and deep learning—once a convenient pedagogical shorthand—has become conceptually exhausted.

Modern systems blend symbolic reasoning, statistical learning, multimodal perception, and adaptive deployment in ways that the ML/DL dichotomy can no longer capture. This paper introduces the Approach-Based AI Framework (ABAF), a functional taxonomy that classifies AI systems by how they operate and what they require rather than by how many layers they contain.

ABAF defines five core functional categories—Rule-Guided, Representation-Driven, Hybrid Reasoning, Resource-Adaptive, and Autonomous Learning—supplemented by four contextual modifiers: data dependency, learning mode, knowledge source, and deployment environment. Together these elements create a descriptive coordinate system linking algorithmic behaviour to operational realities such as compute constraints, interpretability, and governance risk.

We argue that this functional turn provides a stable and interdisciplinary foundation for the next phase of AI practice. It bridges the gap between developers who speak in architectures and decision-makers who need to understand behaviour, reliability, and cost.

The framework also offers regulators a principled basis for risk-proportionate oversight, replacing one-size-fits-all rules with category-specific obligations. Finally, it serves as a pedagogical lens for AI literacy, allowing non-technical professionals to reason about intelligent systems in terms of function and dependency rather than algorithmic jargon.

ABAF’s contribution is threefold: it reframes AI classification around operational behaviour, establishes a language that aligns research, engineering, and governance, and proposes a quantitative extension capable of empirical validation. By articulating AI in functional rather than architectural terms, ABAF positions the field for a more coherent, responsible, and future-resilient evolution.

I. The Crisis in AI Classification: Beyond Algorithmic Depth

The lexicon used to describe and categorize artificial intelligence systems is in a state of crisis. The prevailing distinction between "machine learning" (ML) and "deep learning" (DL), once a useful shorthand, has become conceptually exhausted and a source of profound confusion, impeding effective communication, sound strategic decision-making, and coherent governance.

This conceptual stagnation arises from a fundamental mismatch between a classification system based on algorithmic architecture and the reality of modern AI, which is increasingly defined by its functional behavior, deployment context, and socio-technical integration. The proposal of a new taxonomy, the Approach-Based AI Framework (ABAF), is a direct response to this crisis, signaling a necessary shift from architectural formalism to functional pragmatism.

This section establishes the intellectual and practical necessity for such a framework by examining the erosion of the ML/DL dichotomy and the manifest failure of architectural labels in applied, high-stakes domains.

1.1 The Erosion of the ML/DL Dichotomy

The distinction between machine learning and deep learning has its roots in a period of AI development when algorithmic families were relatively discrete. Classical machine learning encompassed a range of techniques—such as linear regression, support vector machines (SVMs), and decision trees—that typically relied on structured data and manually engineered features. Deep learning, a subfield of machine learning, distinguished itself through the use of neural networks with multiple layers (hence "deep") capable of learning hierarchical feature representations directly from raw, unstructured data.

For a time, this distinction was clear and functionally meaningful: one approach required significant domain expertise to craft inputs, while the other shifted that burden to the algorithm and the availability of vast computational resources and data.

However, the last decade of AI research and development has systematically dismantled the walls separating these categories. The boundary has been eroded not by a single breakthrough but by a pervasive trend of hybridization and architectural convergence. Modern AI systems frequently blend techniques from both domains, rendering a binary classification obsolete and misleading.

This erosion is evident in several key areas:

First, deep learning models are increasingly applied to tabular, structured data—the traditional bastion of classical ML. In these applications, deep architectures are often combined with explicit feature engineering, where domain-specific knowledge is used to create inputs that enhance model performance. The model may be "deep," but its success is contingent on "classical" feature-crafting techniques.
Second, hybrid architectures that fuse deep backbones with traditional algorithms have become commonplace. A prominent example involves using a deep convolutional neural network (CNN) not for end-to-end classification but as a powerful, automated feature extractor. The rich vector representations generated by the CNN's intermediate layers are then fed into a simpler, more traditional classifier like an SVM or a gradient boosting machine for the final prediction. Is such a system "deep learning" or "machine learning"? The question itself reveals the inadequacy of the terminology.
Third, the proliferation of multimodal systems represents a significant challenge to the old taxonomy. These models are designed to process and integrate information from heterogeneous data sources, such as images, text, and sensor readings, simultaneously. A medical AI might, for instance, combine a deep network for analyzing a chest X-ray with a natural language processing (NLP) model for parsing a radiologist's notes and a classical model for interpreting structured lab results. Labeling such an integrated system as merely "deep learning" ignores the critical role of its other components and the complex fusion strategies it employs.

The core issue is that the "depth" of an architecture—the number of layers in its neural network—has ceased to be a reliable proxy for its capability, complexity, or, most importantly, its operational requirements. As algorithmic boundaries blur, clinging to a distinction based on architectural depth obscures more than it clarifies, creating a pressing need for a new classification paradigm. This ongoing confusion is evident in recent academic literature, which still attempts to classify systems using overlapping categories like 'Rule-based, Deep learning-based, Machine learning-based and the hybrid ways'^[18].

1.2 The Failure of Architectural Labels in Applied Contexts

The shortcomings of the ML/DL dichotomy extend beyond academic pedantry; they have significant practical consequences, particularly for non-technical stakeholders in high-stakes fields such as healthcare, finance, and law. For clinicians, hospital administrators, financial regulators, and corporate executives, the internal architecture of an AI system is of secondary importance. Their primary concerns are functional and operational. The architectural labels of "ML" and "DL" fail to answer the critical questions that drive adoption, trust, and governance.

These stakeholders need to understand a system's dependencies and behaviors. For instance, a hospital's Chief Information Officer needs to know the data and resource requirements of a proposed AI tool.

Will it require a massive, centralized data lake and a cluster of high-performance GPUs, or can it operate efficiently on-device with a smaller, curated dataset? The label "deep learning" suggests the former, but advancements in model compression and efficiency are making the latter increasingly viable. The label itself provides no actionable information about resource planning.

Similarly, a clinician evaluating a diagnostic aid is primarily concerned with interpretability and trust. Is the model's reasoning process transparent and auditable, allowing a human expert to verify its logic, or is it an opaque "black box" whose failures are unpredictable and inexplicable?

While deep learning models are often less interpretable, this is not a universal rule, and the entire field of Explainable AI (XAI) is dedicated to mitigating this issue. Conversely, a "classical" ML model, such as a complex ensemble of thousands of decision trees, can be just as inscrutable as a neural network. The architectural label is a poor predictor of interpretability, a crucial factor in domains where accountability is paramount.

Furthermore, the lifecycle and adaptability of an AI system are critical operational concerns. A regulator needs to know if a model is static—trained once and deployed indefinitely—or if it is designed to learn and adapt to new data over time. The latter, which involves concepts from continual and lifelong learning, introduces unique risks and requires different monitoring and validation protocols. These behavioral characteristics are orthogonal to the ML/DL distinction.

Finally, questions of governance and risk, especially concerning data privacy, are central to deployment in sensitive environments. Can the system be trained and operated within a privacy-preserving federated learning architecture, where data remains decentralized?. Can it function on an edge device to minimize data transmission and latency? These deployment constraints are defining features of modern AI engineering, yet they are completely unaddressed by a classification scheme focused on algorithmic structure.

This disconnect between technical labels and practical needs creates a conceptual vacuum. It also fosters a misleading and potentially harmful perception of AI, rooted in linguistic inertia. The term "deep learning" is often interpreted by non-experts as a synonym for "advanced" or "superior" AI, establishing a false hierarchy of sophistication.

This linguistic bias can lead decision-makers to favor complex, data-hungry, and opaque DL models even when simpler, more interpretable, and more robust classical ML systems are better suited to the task and the context. For example, a highly intelligible logistic regression model for predicting pneumonia risk, which allows clinicians to scrutinize its reasoning, may be a far safer and more effective choice than a slightly more accurate but completely opaque deep learning model^[7]. A framework that moves beyond this false hierarchy is essential for promoting sound technological stewardship.

The crisis in AI classification is therefore not merely academic. It is a symptom of the field's maturation from an algorithmic science to an engineering discipline. Early AI research was defined by its algorithms, a hallmark of a developing computational science. Today's AI, however, is increasingly defined by its deployment constraints, its integration into complex socio-technical systems, and its real-world impact.

Just as software engineering developed concepts like design patterns and architectural styles to manage complexity beyond the level of individual algorithms, AI now requires a new vocabulary. The ABAF's focus on "approach and operational requirement" rather than "architecture" is a formal recognition of this transition. It proposes a language for the engineering of AI systems, not just the science of their algorithms, providing a necessary foundation for the next phase of AI development and adoption.

II. A Deep Analysis of the Approach-Based AI Framework (ABAF)

The Approach-Based AI Framework (ABAF) is a direct response to the inadequacies of architecture-centric classification. It proposes a taxonomy grounded in functionalism, contextual adaptability, and educational clarity. The framework is composed of two key elements: five core categories that describe the fundamental operational logic of an AI system, and a set of secondary modifiers that provide essential context about its dependencies and deployment environment.

This section provides a detailed deconstruction and enrichment of the ABAF, analyzing each component to reveal its theoretical underpinnings, practical instantiations, and the nuanced distinctions that define its conceptual boundaries.

2.1 The Five Core Categories: A Functional Deconstruction

The five core categories of ABAF shift the primary question from "What is the algorithm?" to "How does the system work?". They classify systems based on their dominant mode of reasoning, learning, or operation, providing a more intuitive and functionally relevant descriptive language.

The five categories are summarized below before a detailed deconstruction:

Category	Defining Question	Core Principle	Typical Examples
Rule-Guided	Does it rely on explicit human or domain rules?	Symbolic or expert-defined logic ensuring interpretability.	Clinical scoring systems; rule-based decision support; deterministic safety controllers.
Representation-Driven	Does it learn features directly from raw data?	Autonomous feature emergence from unstructured input.	CNN/Transformer models for imaging or text; large multimodal foundation models.
Hybrid Reasoning	Does it combine learned representations with explicit knowledge?	Integration of neural and symbolic components for verifiable reasoning.	Retrieval-Augmented Generation; neuro-symbolic question answering; knowledge-grounded LLMs.
Resource-Adaptive	Is its design shaped by compute, energy, privacy, or bandwidth limits?	Engineering optimisation for constraint—edge, federated, or efficient inference.	On-device diagnostics; federated hospital networks; quantised model deployments.
Autonomous Learning	Can it evolve after deployment through feedback or interaction?	Closed feedback loops enabling continual or reinforcement learning.	Self-tuning recommender systems; adaptive ICU monitors; RL-based control agents.

Visual Representation: The 5 ABAF Categories

Rule-Guided

Explicit human logic

Representation-Driven

Learns from raw data

Hybrid Reasoning

Learned + Explicit Knowledge

Resource-Adaptive

Constraint-driven design

Autonomous Learning

Evolves post-deployment

2.1.1 Rule-Guided Systems

Defining Question: Can the system operate using explicit human or domain rules?

This category encompasses what is often considered the simplest and most foundational form of artificial intelligence: the rule-based expert system. These systems, rooted in the tradition of Symbolic AI or "Good Old-Fashioned AI" (GOFAI), operate on a set of explicit, human-crafted rules, typically in an "if-then" format, to mimic the decision-making of a human expert.

The primary logic is derived directly from a knowledge base of prescribed rules rather than learned from data. Clear, practical examples from healthcare include clinical decision support tools like the 'Contrast Planner,' which uses hardcoded rules based on medical guidelines to advise on contrast agent use. This category also includes other models where behavior is intentionally constrained by domain knowledge for transparency, such as decision trees and certain linear regression models.

This category finds strong resonance with the concept of "Informed Machine Learning," a research area focused on systematically integrating prior knowledge into learning systems to improve data efficiency, robustness, and interpretability. By formalizing the role of expert knowledge, Rule-Guided systems directly address the critical need for intelligible models in high-stakes domains like healthcare. The work of Caruana et al. on creating a highly accurate yet fully understandable model for predicting pneumonia risk is a canonical example of a Rule-Guided system, where the final model consists of a set of simple, additive risk factors that clinicians can directly inspect and validate^[7]. The value of this category lies in its explicit acknowledgment of systems designed for human oversight and trust.

2.1.2 Representation-Driven Systems

Defining Question: Does it learn directly from raw, unstructured data?

This category is the closest analogue to the conventional understanding of "deep learning," but its definition is crucially functional rather than architectural. A Representation-Driven system is characterized by its ability to perform end-to-end learning, discovering hierarchical features (representations) directly from raw data such as pixels, audio waveforms, or text sequences, without the need for manual feature engineering. The emphasis is on autonomous pattern discovery from large-scale datasets.

In healthcare, this includes the vast majority of modern deep learning models used for perception, such as CNNs that analyze medical images like X-rays and CT scans to help detect cancer. It also includes Transformers for understanding clinical notes and generative models for synthesizing medical data.

The power of this approach is exemplified in its ability to tackle complex perceptual tasks where the relevant features are too intricate or numerous to be defined by hand. Multimodal learning, which fuses representations learned from different data types to achieve a more holistic understanding, is a prime example of the sophistication of Representation-Driven systems.

However, this power comes at a cost. The autonomous nature of feature learning in these systems is the primary source of the "black box" problem, which has catalyzed the entire field of Explainable AI (XAI) dedicated to interpreting their internal workings and decisions^[21]. By classifying these systems based on their function of representation learning, ABAF focuses attention on the consequences of this approach—namely, the need for large datasets, high computational power, and specialized techniques for ensuring transparency and trust.

2.1.3 Hybrid Reasoning Systems

Defining Question: Does it combine learned patterns with explicit knowledge?

This category marks the convergence of the two preceding approaches, representing a sophisticated fusion of symbolic AI (knowledge representation and logic) and connectionist AI (learning from data). A Hybrid Reasoning system combines the powerful pattern-recognition capabilities of Representation-Driven models with the precision and verifiability of explicit, structured knowledge. This is not merely about pre-training a model on a large text corpus; it involves systems that can actively access, query, and reason over an external knowledge base during inference to ground, correct, or augment their outputs.

This category directly reflects the rapid growth of Neurosymbolic AI, a field systematically reviewed by^[9], which aims to build more robust and trustworthy models by integrating learning with reasoning.

Prominent contemporary examples include Retrieval-Augmented Generation (RAG). In a healthcare RAG architecture, a large language model (a Representation-Driven component) does not rely solely on its internal knowledge. Instead, it retrieves relevant information from a curated, external source—such as a medical knowledge graph or a clinical guideline database (a Rule-Guided component)—and uses this retrieved context to formulate its response^[11]. A practical implementation of this approach is 'RadIQPro.in,' which combines LLM capabilities with structured radiological guidelines through context prompting to provide evidence-based imaging recommendations^[17]. This approach allows the system to provide more accurate, up-to-date, and verifiable answers, as the source of its information can be explicitly cited. This combination of emergent and expert knowledge is a critical step toward building AI systems that are both powerful and accountable.

2.1.4 Resource-Adaptive Systems

Defining Question: Can it adapt to compute, privacy, or bandwidth limits?

This category represents ABAF's most significant departure from traditional taxonomies. It classifies systems not by their method of reasoning but by their ability to operate under significant operational constraints. The defining characteristic of a Resource-Adaptive system is that its design and implementation are fundamentally shaped by limitations in compute power, memory, energy consumption, network bandwidth, or data privacy requirements. This is a pragmatic, engineering-driven classification.

This single category synthesizes a vast and diverse body of research that is critical for real-world AI deployment. In a medical context, it includes AI algorithms running directly on portable or bedside devices like smart stethoscopes, portable ultrasound machines, or patient monitors, where computational power and battery life are limited.

It also encompasses the suite of techniques for making large models more efficient, such as quantization, pruning, and knowledge distillation, which fall under the umbrella of resource-efficient deep learning. Crucially, this category also includes architectures designed for privacy, with Federated Learning being the most prominent example. In Federated Learning, models are trained decentrally on local data (e.g., at different hospitals) without the data ever leaving its source, addressing critical privacy and data governance challenges^[12]. The system's architecture is dictated by the privacy constraint. This category acknowledges that in many practical applications, the deployment environment and its constraints are the primary drivers of AI system design.

The inclusion of this non-functional category is a powerful strategic move. While this represents a shift from a purely *functional* axis to an *operational* one, this move is ABAF's most critical contribution, as it reframes engineering constraints as a primary design approach, not an afterthought.

While a functional taxonomy should, by definition, classify based on function (e.g., reasoning), resource constraints are an operational boundary condition. A tumor segmentation model (a Representation-Driven function) can be deployed in the cloud or on an edge device. The core function remains the same, but the engineering choices required to make it work on the edge—model compression, integer quantization—fundamentally alter the system's properties, including its performance, reliability, and cost.

By creating a primary category for this, ABAF elevates deployment engineering from a secondary concern to a first-class citizen in AI classification. It formally recognizes that how a model is deployed is often as important as what it does, thereby bridging the conceptual gap between AI science and operational reality.

2.1.5 Autonomous Learning Systems

Defining Question: Can it evolve via feedback or interaction?

The final category describes systems capable of adapting their behavior and improving their performance after initial deployment through ongoing interaction with their environment. The key feature is a closed feedback loop that enables adaptation and refinement without requiring a full, manual retraining cycle supervised by developers. This category is defined by its dynamic, evolving nature.

This classification encompasses a range of learning paradigms. The most well-known is Reinforcement Learning (RL), where an agent learns to make optimal decisions by receiving rewards or penalties for its actions. It also includes the broader fields of online learning and Continual Learning, which focus on developing models that can incrementally learn from a continuous stream of data without catastrophically forgetting previously learned knowledge.

This concept has deep historical roots in AI, tracing back to early work on lifelong learning algorithms. In a modern healthcare context, an Autonomous Learning system could be a smart ICU monitoring system that learns a patient's unique physiological baseline and adapts its alerts over time to become more specific, reducing alarm fatigue for clinical staff. These systems offer the promise of continuous improvement but also introduce significant challenges related to safety, stability, and long-term behavioral predictability.

Together, these five categories provide a comprehensive functional map of the AI landscape. They also implicitly create a spectrum of autonomy and interpretability. The progression from Rule-Guided to Autonomous Learning systems charts a course from maximum human control and transparency to maximum machine autonomy and emergent behavior.

Visual Representation: Spectrum of Autonomy & Interpretability

Rule-Guided

Max Control

High Interpretability

Autonomous

Max Autonomy

Low Interpretability

Interpretability typically decreases and governance risk increases along this axis.

A Rule-Guided system is maximally constrained by human knowledge and is therefore highly interpretable. A Representation-Driven system cedes feature discovery to the algorithm, increasing autonomy at the cost of innate interpretability, thus necessitating XAI techniques. A Hybrid Reasoning system attempts to strike a balance by re-introducing human-curated knowledge to guide and constrain the less-interpretable component. Finally, an Autonomous Learning system represents the highest degree of autonomy, as it evolves its own behavior post-deployment, posing the greatest challenges for governance and long-term alignment.

ABAF thus provides a language not just for classification, but for discussing the critical trade-offs between human oversight and machine autonomy—a central dialogue for the future of responsible AI. The framework thus doubles as a conceptual map for oversight intensity.

2.2 The Secondary Modifiers: Adding Contextual Dimensions

If the five core categories define the primary "mode" of an AI system, the secondary modifiers provide the essential contextual dimensions that transform ABAF from a simple classification scheme into a rich, multi-dimensional descriptive framework.

Learning Mode (Static / Incremental / Interactive): This provides crucial information about the model's post-deployment lifecycle, a key component of AI safety and evaluation frameworks^[22]. "Static" describes a model that is trained once and does not change. "Incremental" refers to a system that can be periodically updated with new data, a key feature of Continual Learning systems. "Interactive" describes a system that learns continuously from user feedback or environmental interaction, aligning with the Autonomous Learning category. This modifier is critical for planning maintenance, monitoring, and validation strategies.
Knowledge Source (Expert / Learned / Combined): This modifier explicitly defines the provenance of the system's "intelligence," a key factor for trust, auditing, and accountability. "Expert" knowledge comes from human-defined rules or features, characteristic of Rule-Guided systems. "Learned" knowledge is emergent, discovered from data by Representation-Driven systems. "Combined" is the hallmark of Hybrid Reasoning systems, which leverage both.
Deployment Environment (Cloud / Edge / Hybrid): This is the framework's most explicit link to operational reality. It captures the context of resource availability and data governance. "Cloud" deployment implies access to significant computational resources but may introduce latency and data privacy concerns. "Edge" deployment signifies that the model runs on a local device (e.g., a smartphone, a medical scanner), which is essential for low-latency, offline functionality and privacy preservation. This directly connects to the Resource-Adaptive category and the technologies that enable it.
Data Dependency (Low / High / Mixed): This modifier clarifies the system's reliance on data, moving beyond the simple structured-versus-unstructured binary. "Low" might describe a Rule-Guided clinical scoring system that operates on a few dozen variables, while "High" would characterize a Representation-Driven foundation model trained on petabytes of internet data. "Mixed" could describe a Hybrid Reasoning system that uses both. This modifier is a direct proxy for the cost, infrastructure, and governance overhead associated with a model.

By combining a primary category with these modifiers, one can create a detailed "ABAF profile" of an AI system. For example, a federated learning system for diagnosing pneumonia from chest X-rays on hospital servers could be classified as: Resource-Adaptive (primary category) with the modifiers High Data Dependency, Incremental Learning Mode, Learned Knowledge Source, and Hybrid Deployment Environment.

This profile is vastly more informative for a hospital administrator or regulator than the simple label "deep learning." It immediately communicates that the system is designed for privacy, requires large amounts of data, can be updated over time, learns features automatically, and operates across both local servers and a central coordinator. This multi-dimensional approach is what gives ABAF its practical power.

2.3 Re-contextualizing Traditional Methods within ABAF

A common question is where traditional concepts like "supervised learning" or specific algorithms like "CNNs" fit into the ABAF framework. This is a critical point of clarification: ABAF does not *replace* these terms, but rather *re-contextualizes* them.

Traditional taxonomies focus on *how an AI model is built* (e.g., its training method or its architecture). ABAF focuses on *what the resulting system does* and *what it needs to operate* (e.g., its function and dependencies). Think of "supervised learning" as a *training recipe* and "CNN" as an *ingredient*; the ABAF category describes the *finished dish* and its functional properties.

Visual Representation: Mapping Old Terms to ABAF Categories

Old Terms (How it's built)

Supervised Learning
Unsupervised Learning
Reinforcement Learning
CNNs, Transformers
Decision Trees, SVMs
Federated Learning (Arch)
RAG (Arch)

→

ABAF Categories (What it does)

Rule-Guided: (SVMs, Trees, Regression)
Representation-Driven: (CNNs, Transformers, GANs, Unsupervised Pre-training)
Hybrid Reasoning: (RAG, Neurosymbolic)
Resource-Adaptive: (Federated Learning, Quantized Models)
Autonomous Learning: (Reinforcement Learning, Online Learning)

ABAF shifts focus from implementation details to functional behavior and operational needs.

Mapping Learning Methods (Supervised, Unsupervised)

These are training processes, not ABAF categories. A single method can be used to build systems in different categories.

Supervised Learning (using labeled data): This method can be used to create:
- A Rule-Guided System: For example, training a logistic regression model where the inputs are explicit, human-defined features (e.g., "patient age," "smoker_status"). The model learns *weights* for those pre-defined rules.
- A Representation-Driven System: The most common use case. Training a CNN on labeled X-rays to detect tumors is a classic example. The system learns the features and the classification all at once from the raw data.
Unsupervised Learning (using unlabeled data): This method is most often used to create:
- A Representation-Driven System: This is its most powerful application. Clustering algorithms that find new patient sub-groups from lab data, or generative models (GANs) that learn to create synthetic X-rays, are Representation-Driven. Their *function* is to discover and learn the underlying patterns of the data distribution. The pre-training phase of most large language models is also unsupervised.
Reinforcement Learning (RL):
- This method maps most directly to a single category. It is the primary *engine* for building Autonomous Learning systems. The *function* of an RL system is to learn and evolve its behavior by interacting with an environment, which is the definition of the Autonomous Learning category.

Mapping Algorithmic Architectures (CNNs, Transformers, etc.)

These are the *tools* used to build a system within an ABAF category.

CNNs (Convolutional Neural Networks): These are quintessential tools for building Representation-Driven systems for visual or spatial data. Their entire *function* is to learn hierarchical features from raw pixels.
Transformers (e.g., BERT, GPT): These are the primary tools for building Representation-Driven systems for sequential data like text. Their *function* is to learn deep contextual relationships from sequences of tokens.
RAG (Retrieval-Augmented Generation): This is a *system architecture* that is the perfect example of a Hybrid Reasoning system. It combines a Representation-Driven component (the LLM) with a Rule-Guided component (the explicit, verifiable knowledge base it retrieves from).
Federated Learning: This is a *deployment architecture* that defines a Resource-Adaptive system. The *underlying model* at each node might be a Representation-Driven CNN, but the *system as a whole* is classified as Resource-Adaptive because its design is fundamentally shaped by the constraint that data cannot be centralized.

By making this mapping explicit, ABAF translates implementation details into functional and strategic identities. The fact that a system uses a "CNN" is a technical detail; the fact that it is Representation-Driven is its functional identity, which immediately surfaces critical questions about data appetite, opacity, and governance.

Comprehensive Algorithm and Technique Mapping

To make this transition concrete, the following table maps a wide range of common AI algorithms, techniques, and architectures to their primary ABAF category.

Algorithm / Technique	Primary ABAF Category	Rationale (Why it fits here)
--- Classical Machine Learning ---
Linear / Logistic Regression	`Rule-Guided`	Relies entirely on human-defined, explicit input features. The model's "rules" are the features it's given.
Decision Trees / Random Forests	`Rule-Guided`	Creates an explicit, human-readable (in theory) set of `if-then` rules, but operates on human-defined features.
Support Vector Machines (SVMs)	`Rule-Guided`	Finds the optimal boundary between data points that are defined by human-engineered features.
k-Means Clustering	`Rule-Guided` or `Rep-Driven`	(Rule-Guided) when clustering on human-defined features (e.g., age, income). (Rep-Driven) when clustering on learned features (e.g., document embeddings).
--- Deep Learning (Representation-Driven Tools) ---
CNNs (Convolutional Neural Nets)	`Representation-Driven`	The quintessential example. Its entire function is to learn hierarchical spatial features (representations) from raw pixels.
Transformers (BERT, GPT, etc.)	`Representation-Driven`	Learns deep, contextual representations and patterns directly from raw text sequences (tokens).
Autoencoders / VAEs	`Representation-Driven`	An unsupervised method to learn a compressed representation (a "code") of raw input data.
GANs (Generative Adversarial Nets)	`Representation-Driven`	Learns the entire underlying distribution of a raw dataset (e.g., images) to generate new, synthetic samples.
--- Hybrid Systems ---
RAG (Retrieval-Augmented Gen)	`Hybrid Reasoning`	The canonical system. It combines a `Rep-Driven` model (the LLM) with an external `Rule-Guided` component (the knowledge base).
Neurosymbolic AI (general)	`Hybrid Reasoning`	Any system that explicitly fuses neural networks (for pattern learning) with symbolic logic (for reasoning).
--- Adaptive Systems & Architectures ---
Federated Learning	`Resource-Adaptive`	A system architecture whose primary design driver is a constraint (data privacy/decentralization), not its learning method.
Quantized/Pruned Models (e.g., GPTQ)	`Resource-Adaptive`	The technique creates a new model whose primary functional characteristic is its adaptation to resource limits (memory, compute).
Reinforcement Learning (Q-Learning, PPO)	`Autonomous Learning`	The model's function is to learn and evolve its behavior after deployment by interacting with an environment to maximize a reward.
Online Learning Systems	`Autonomous Learning`	Designed to be incrementally updated in real-time from a continuous stream of new data, thus evolving its behavior.

III. A Critical Evaluation and Comparative Analysis of ABAF

While the Approach-Based AI Framework (ABAF) presents a compelling alternative to architecture-centric taxonomies, a thorough assessment requires a critical evaluation of its strengths and weaknesses, as well as a comparative analysis that situates it within the broader landscape of AI classification paradigms. This section undertakes such an evaluation, examining the framework's intellectual contributions, identifying its conceptual ambiguities, and contrasting its approach with existing methods of categorization.

3.1 Strengths and Intellectual Contributions

ABAF’s primary contribution lies in re-anchoring AI classification around function and operation, a shift from architectural essentialism to behavioural pragmatism. Three strengths define its intellectual coherence:

Deployment-centric clarity. ABAF foregrounds what most taxonomies ignore—the engineering and environmental realities that determine reliability and cost. By elevating Resource-Adaptive and Learning Mode to first-class descriptors, it bridges laboratory models and fielded systems. This is the vocabulary operations teams and regulators actually need.
Interdisciplinary translation. Traditional taxonomies—architectural or mathematical—serve developers but alienate decision-makers. ABAF replaces jargon with conceptual handles: “expert vs. learned knowledge,” “cloud vs. edge environment.” These are intelligible to clinicians, lawyers, or policymakers without technical dilution, creating a common semantic substrate across disciplines.
Forward-compatibility. Because ABAF classifies by approach rather than architecture, it remains stable as paradigms evolve. Whether diffusion models, world-model agents, or neurosymbolic hybrids dominate, their behaviour can still be located within the five functional axes. In this sense ABAF is time-robust—a rare property in AI frameworks that usually expire with each algorithmic cycle.

3.2 Weaknesses, Ambiguities, and Areas for Refinement

No taxonomy is immune to overlap, and ABAF’s conceptual power introduces its own tensions. Despite its significant strengths, a critical analysis reveals several areas of conceptual ambiguity and potential for misapplication that must be addressed.

(a) Boundary Ambiguity. Modern systems routinely combine multiple functional modes. The five core categories, while presented as distinct, exist on a continuum. A retrieval-augmented language agent, for instance, is simultaneously Representation-Driven (neural encoder), Hybrid Reasoning (knowledge retrieval), and potentially Resource-Adaptive (edge caching). Assigning a single “primary” label can be arbitrary and lacks clear guidelines for classifying such multi-faceted systems.
(b) Risk of Semantic Drift. The term Representation-Driven could be misinterpreted as a fashionable synonym for "deep learning." If practitioners use it as a simple drop-in replacement without embracing the functional mindset that underpins the entire framework, the problem of misleading terminology will persist under a new name. Its value depends on users focusing on the *implications* of representation learning (e.g., data appetite, opacity) rather than just using it as a label.
(c) Need for Empirical Grounding. ABAF is conceptually rigorous but presently qualitative. Adoption will depend on validation through systematic classification exercises: multiple experts independently rating real-world AI systems, followed by inter-rater reliability analysis. Without such evidence, critics may dismiss ABAF as descriptive theory rather than a usable instrument.
(d) Perspective Hierarchy. Different stakeholders may prioritise different primary labels for the same system—clinicians emphasising interpretability (Rule-Guided), CIOs emphasising deployment (Resource-Adaptive), regulators emphasising adaptivity (Autonomous Learning). Future iterations of the framework should clarify a perspective hierarchy—a method for selecting the dominant category relative to the specific use-case context.

3.3 Comparative Analysis with Other TaxonomIES

To fully appreciate ABAF's unique contribution, it is essential to situate it within the context of other major AI classification paradigms.

Paradigm	Primary Axis of Classification	Key Question Answered	Examples of Categories	Strengths	Limitations for Applied Domains
Architectural	Internal structure and algorithm family.	"What is it made of?"	Neural Networks, SVMs, Decision Trees, Transformers.	Precise for developers; algorithmically clear.	Opaque to non-experts; poor predictor of operational needs; quickly outdated.
Learning Method	How the model is trained.	"How does it learn?"	Supervised, Unsupervised, Reinforcement Learning.	Foundational for education; describes training process well.	Doesn't describe the deployed model's behavior or requirements; many systems use multiple methods.
Symbolic vs. Connectionist	The nature of knowledge representation.	"How does it think?"	Logic-based systems vs. Neural networks.	Captures a fundamental philosophical divide in AI.	"Increasingly obsolete due to the rise of hybrid/neurosymbolic systems."
AI Safety Taxonomy	Potential failure modes and safety concerns.	"How can it fail?"	Specification, Robustness, Assurance.	Essential for risk management; focuses on reliability.	Primarily for safety engineers; doesn't classify the AI's function or purpose.
ABAF (Proposed)	Functional approach and operational requirements.	"What does it need, and how does it behave?"	Rule-Guided, Representation-Driven, Hybrid, Resource-Adaptive, Autonomous.	Interdisciplinary; deployment-centric; future-compatible; links function to governance.	Potential for ambiguity at category boundaries; requires contextual knowledge to apply correctly.

This comparative analysis reveals the distinct niche that ABAF aims to fill. Our framework's 'functional turn' aligns with parallel efforts in major standards bodies. For example, the **OECD *Framework for the Classification of AI Systems*** \[16] and the **NIST *AI Use Taxonomy*** \[20] are concurrent proposals to create classification systems independent of technical architecture and focused on policy, context, and "human-AI tasks." This shows a clear and necessary convergence toward functional, human-centric classification.

While other taxonomies are valuable for their specific purposes, they are tailored to particular audiences. The Architectural paradigm serves developers and computer scientists. The Learning Method paradigm is essential for educating students on training methodologies. The historical Symbolic vs. Connectionist debate is important for understanding the intellectual history of AI but is less relevant for classifying modern hybrid systems. Finally, AI Safety Taxonomies \[1, 19] are indispensable for risk and reliability engineers but do not describe what an AI system does , only how it might break.

ABAF, in contrast, is explicitly designed for the high-level, multi-stakeholder dialogue that is essential for responsible AI adoption and governance in the real world. Its key question—"What does it need, and how does it behave?"—is precisely the question that a hospital board, a regulatory agency, or a corporate strategy team needs to answer.

It subsumes elements of the other paradigms but reframes them in a functional context. By focusing on the functional implications rather than the technical implementation, ABAF provides a uniquely powerful tool for strategic decision-making, distinguishing itself as the first major taxonomy designed primarily for the "engineering" phase of AI's maturation.

IV. Strategic Implications: From Hospital to Boardroom

The adoption of the Approach-Based AI Framework (ABAF) has consequences that extend far beyond classification semantics.

By reframing how systems are described, ABAF has the potential to rewire decision-making at three critical levels of the AI ecosystem: governance, strategy, and education.

This section articulates how a functional taxonomy can translate directly into improved regulation, procurement, and literacy—three domains currently constrained by architecture-centric thinking.

4.1 Functional Classification as a Foundation for Risk-Based Governance

Current regulatory frameworks, from the EU AI Act to emerging U.S. and Indian guidelines, create a complex landscape for organizations \[6, 10]. These frameworks tend to group AI by sector or intended use rather than by functional behaviour, though some, like \[3] and \[5], are beginning to build domain-specific taxonomies.

The result is either overbroad, one-size-fits-all rules or fragmented sub-regulations that lag behind innovation.

ABAF introduces a more stable and proportionate alternative: a category-based governance model in which obligations scale with the autonomy and opacity of the system. This aligns with the growing need for practical, standardized tools for AI assurance and governance, such as the LLM-based RATS (Resiliency, Accountability, Trust, and Safety) score proposed by initiatives like *beresponsibleai.com* \[4].

Proposed alignment between ABAF category and regulatory focus:

ABAF Category	Principal Regulatory Concern	Recommended Oversight Approach
Rule-Guided	Accuracy and provenance of expert knowledge	Audit of rule base, bias assessment, and version control
Representation-Driven	Opacity and data bias	Mandatory robustness testing, explainability protocols, post-market surveillance
Hybrid Reasoning	Verifiability of knowledge integration	Logging and traceability of retrieved sources; validation of symbolic interface
Resource-Adaptive	Distributed data governance and security	Certification of federated aggregation, edge-device integrity, privacy guarantees
Autonomous Learning	Behavioural drift and alignment risk	Continuous monitoring, human-in-the-loop override, real-time audit trails

This approach allows regulation to become adaptive and risk-proportionate, rather than reactive.

It also lends itself to tier-based governance, where oversight intensity grows with the degree of autonomy and environmental sensitivity, providing a blueprint for institutions such as medical device regulators, financial authorities, or national AI councils.

4.2 Strategic Transformation in AI Procurement and Portfolio Management

AI adoption in enterprises and healthcare systems is still dominated by algorithmic marketing—vendors promoting “deep learning” superiority rather than demonstrating contextual fit. Business-oriented taxonomies \[8] are emerging to combat this, and ABAF provides a rigorous, cross-domain standard.

ABAF realigns procurement around operational suitability.

By requiring vendors to specify an ABAF profile—primary functional category plus modifiers (Data Dependency, Learning Mode, Knowledge Source, Deployment Environment)—organizations can compare solutions based on tangible dimensions:

Infrastructure Fit: Does the model’s data dependency and deployment mode align with available compute and governance policies?
Maintainability: Does its learning mode (static vs. incremental) match the organization’s capability for monitoring and retraining?
Trustworthiness: Does its knowledge source enable traceability and domain validation?

Such profiling replaces rhetorical claims (“AI-powered” or “deep-learning based”) with transparent, auditable descriptors.

In strategic terms, ABAF also facilitates portfolio mapping.

A hospital or technology firm can chart all its AI assets across the five axes to identify concentration risk—e.g., over-reliance on opaque representation-driven tools—and rebalance investments toward interpretable or resource-efficient systems.

This mirrors how cybersecurity frameworks map risk posture across enterprise assets.

If adopted widely, ABAF could shift competitive dynamics in the AI marketplace: from performance benchmarks toward metrics such as interpretability, adaptability, and lifecycle cost—dimensions that directly impact value and safety.

4.3 Building AI Literacy and Institutional Capacity

A pervasive obstacle in responsible AI deployment is the literacy gap among non-technical stakeholders who are nonetheless accountable for AI oversight—clinicians, executives, policymakers, and educators.

ABAF offers a pedagogical scaffolding to bridge this gap.

Because it abstracts away from mathematical implementation and focuses on function, it allows educators to teach how AI systems behave without requiring coding fluency.

An ABAF-based curriculum could be structured around six diagnostic questions:

What kind of data does it need? (Data Dependency)
How does it update or learn? (Learning Mode)
Where does its “intelligence” originate? (Knowledge Source)
How is it deployed? (Deployment Environment)
How transparent is it? (Rule-Guided vs. Representation-Driven)
Does it change over time? (Autonomous Learning)

These questions correspond directly to governance, safety, and ethics considerations, giving professionals a conceptual handle for accountability.

Institutionally, the framework can also inform the organization of AI ethics or responsibility teams. Taxonomies of trustworthiness \[14] can be mapped directly to ABAF's functional categories.

Each ABAF category carries a distinct ethical risk:

Rule-Guided systems risk embedding human bias in explicit rules.
Representation-Driven systems risk data-driven bias and opacity.
Hybrid Reasoning systems risk provenance confusion and traceability loss.
Resource-Adaptive systems introduce privacy and decentralization risks.
Autonomous Learning systems pose alignment and oversight challenges.

Dividing governance responsibilities along these lines creates functional specialization within ethics teams, improving focus and reducing ambiguity in accountability.

4.4 Synthesis: Institutionalising the Functional Turn

Adopting ABAF is not a semantic exercise; it constitutes a paradigm shift in institutional reasoning about AI.

It moves organizations from algorithm-first to function-first thinking—aligning technological assessment with the dimensions that determine safety, scalability, and cost.

For policymakers, it enables proportional regulation.

For executives, it operationalises strategy.

For educators, it scaffolds comprehension.

In each case, the shift from “what architecture?” to “what behaviour?” replaces jargon with actionable insight.

Ultimately, ABAF’s strategic power lies in its ability to unify the language of AI’s three domains—research, operations, and governance—under a single functional lexicon.

It does not discard existing paradigms but connects them, creating an interpretable grammar for the applied AI era.

V. Future Trajectory and Quantitative Evolution of ABAF

A viable framework must classify not only the present but also anticipate the future.

The long-term value of the Approach-Based AI Framework (ABAF) will depend on its capacity to evolve from a descriptive taxonomy into a quantitative, empirically testable system capable of mapping next-generation AI architectures whose boundaries are increasingly fluid.

5.1 From Qualitative Taxonomy to Quantitative Coordinate System

To move beyond interpretation toward measurement, ABAF can be formalized as a multi-dimensional coordinate model, an approach that aligns with recent calls for multi-dimensional "matrix" models of classification \[13]. We propose this model consists of two distinct vectors:

First, a five-axis Behavioral Signature representing the system’s degree of expression along each primary functional dimension:

Behavioral Signature = (RG, RD, HR, RA, AL)

where RG = Rule-Guidedness, RD = Representation-Drivenness, HR = Hybrid Reasoning, RA = Resource-Adaptiveness, AL = Autonomous Learning.

Each axis is scored on a consistent **4-point scale (0–3)** using the following qualitative anchors, with operational proxies derived from observable system attributes:

0 – Not Applicable / Absent: The characteristic is not observed, or irrelevant for the system’s design.
1 – Mild / Minimal Presence: Present in isolated modules or as a secondary trait. Limited functional impact.
2 – Moderate / Evident: Clearly contributes to system behaviour but balanced by other design drivers.
3 – Significant / Defining: Dominant or defining property shaping the system’s behaviour or deployment.

Axis	Quantitative Proxy	Illustrative Metric
RG	Fraction of decision logic directly interpretable or rule-based	% of output variance explained by explicit rules
RD	Degree of autonomous feature emergence	Ratio of learned to engineered features
HR	Reliance on external structured knowledge during inference	Frequency/impact of knowledge-base queries
RA	Efficiency under constraint	FLOPs or Joules per inference normalized to baseline
AL	Capacity for post-deployment adaptation	Magnitude of autonomous performance change over time

Example Signatures:

System	RG	RD	HR	RA	AL
GPT-4-class LLM	1	3	2	1	1
Federated Radiology Network	1	2	1	3	0
Adaptive Reinforcement Agent	0	3	2	2	3

To clarify these illustrative scores:

A GPT-4-class LLM scores highest on RD (3 - Defining) because its core function is learning representations from massive, raw text data. It scores minimally on RG (1 - Minimal, as some rules might exist via RLHF or prompting), RA (1 - Minimal, it's resource-intensive), and AL (1 - Minimal, generally static post-training). Its HR score is Moderate (2) reflecting its ability to use tools or potentially function in a RAG setup.
Conversely, a Federated Radiology Network scores highest on RA (3 - Defining) because its *defining characteristic* is being a system built to adapt to the "resource constraint" of data privacy. The underlying model's RD nature is still evident (2 - Moderate), but secondary. Its RG and HR are minimal (1), and it's typically not adaptive post-deployment (AL=0 - Absent).
Finally, an Adaptive Reinforcement Agent (like a self-driving system) scores highest on AL (3 - Defining), as its entire purpose is to evolve behavior through interaction. It also scores highly on RD (3 - Defining) as it learns from raw sensor data. Its use of internal models or external APIs gives it Moderate HR (2) and potentially Moderate RA (2) if optimized for onboard compute, while explicit rules are usually absent (RG=0).

Plotting systems in this five-dimensional space allows for a direct comparison of their core functional behavior.

Visual Representation: ABAF Behavioral Signatures (Example Radar Charts)

GPT-4 Class LLM

Dominantly Representation-Driven

Federated Radiology

Dominantly Resource-Adaptive

Adaptive RL Agent

Dominantly Autonomous Learning

Each signature provides a unique functional fingerprint.

This 5-axis *Behavioral Signature* tells us *what the AI does*. To complete the profile, we must also quantify its operational context. We therefore propose a second, 4-axis Contextual Signature based on the modifiers:

Contextual Signature = (DD, LM, KS, DE)

where DD = Data Dependency, LM = Learning Mode, KS = Knowledge Source, and DE = Deployment Environment. Each of these axes can also be scored on the same **0-3 scale** using the qualitative anchors defined above, based on appropriate operational proxies:

Axis	Quantitative Proxy	Illustrative Metric
DD	Reliance on data volume and diversity	Ratio of data volume to model parameters
LM	Post-deployment adaptability	Frequency/scale of autonomous model updates
KS	Provenance of decision logic	% of output variance from expert vs. learned rules
DE	Degree of decentralized operation	% of inference computed at the edge vs. cloud

A full ABAF Profile thus consists of both signatures using the consistent **0-3 scale** (e.g., (RG:1, RD:2... | DD:2, LM:1...)), providing a complete 9-dimensional snapshot of the system's function *and* its operational context.

This 9-dimensional profile functions like a "BMI for AI." It does not reveal every internal detail, but it provides a standardized, high-level, and comparative snapshot of a system's behavior and dependencies. This quantitative turn transforms ABAF from a semantic map into a true measurement instrument suitable for comparative analysis, lifecycle monitoring, and risk stratification.

5.2 Stress-Testing the Framework Against Emerging Paradigms

As AI evolves toward greater autonomy and integration, ABAF must remain robust under new conceptual stresses. Three frontier classes illustrate its adaptability.

a. World Models and Generative Physical Simulators
Models that internalize dynamic representations of reality (e.g., DeepMind Gato, Genesis) blur learning and reasoning. Initially Representation-Driven, they develop internal simulacra that function as implicit knowledge bases—effectively Hybrid Reasoning within a learned world. ABAF accommodates this by interpreting reasoning over self-generated internal models as HR behaviour rather than inventing new categories.
b. Autonomous Agents and Tool-Using LLMs
Agentic systems such as AutoGPT or enterprise copilots exhibit composite behaviour: high RD (LLM backbone), moderate HR (knowledge retrieval and API reasoning), substantial AL (adaptation through feedback). The coordinate model expresses this hybridity without semantic strain, eliminating the need for arbitrary “meta-AI” labels.
c. Scientific Foundation Models
Models trained on multimodal scientific data to generate hypotheses or design molecules challenge the boundary between learning and knowledge creation. Their signatures often progress from RD dominance to RG/HR elevation as their discoveries are codified into explicit rules. ABAF thus captures the transition from generative discovery to knowledge formalization within a single functional continuum.

These stress-tests confirm that ABAF’s functional axes are orthogonal and durable: new architectures alter coordinates but do not require new dimensions.

5.3 A Roadmap for Evolution and Standardization

To institutionalize ABAF, three development tracks are recommended:

Empirical Validation Program: Construct an open repository of AI systems annotated with ABAF signatures by independent experts, enabling meta-analysis across domains and measuring inter-rater reliability.
Metric Standardization: Collaborate with standards bodies (e.g., NIST, ISO, OECD) and open frameworks like the AICS \[2] to formalize scoring rubrics and reporting templates \[15].
Integration with Regulatory Schemes: Embed ABAF profiles into documentation required by risk-based frameworks (e.g., EU AI Act conformity assessments), providing regulators a functional descriptor of system behaviour.

5.4 The Functional Turn as a Long-Term Paradigm Shift

Historically, AI has cycled through taxonomies—symbolic vs. connectionist, supervised vs. unsupervised, statistical vs. neural—each dissolving as technology hybridized.

The functional turn proposed by ABAF may represent the field’s first stable abstraction: classification by operational behaviour and constraint rather than internal mechanics.

If developed into a quantitative standard, ABAF could underpin:

Cross-disciplinary comparability (engineering ↔ policy ↔ ethics).
Functional risk indexing for adaptive regulation.
Longitudinal analysis of AI system evolution.

In doing so, it would provide what the AI ecosystem currently lacks—a shared coordinate system for intelligent behaviour, uniting scientific precision with policy relevance.

VI. Conclusion: A Call to Action

The Approach-Based AI Framework (ABAF) represents a critical and timely "functional turn" in AI classification. By shifting the focus from architectural minutiae to operational requirements and functional behavior, it provides a pragmatic, interdisciplinary, and forward-compatible language for understanding, developing, and governing artificial intelligence.

Its five core categories, contextual modifiers, and quantitative coordinate system create a rich descriptive tool that bridges the gap between technical specialists and the diverse stakeholders responsible for AI deployment.

While the research roadmap for validating and standardizing ABAF is now clear (Section V.3), the onus is on practitioners and policymakers to put this functional turn into practice.

6.1 A Call for Industry Adoption

Leaders in technology and regulated sectors must move beyond algorithmic marketing. We call on industry to:

Adopt ABAF for Portfolio Management: Map existing and planned AI initiatives to the framework to identify functional gaps, manage concentration risk, and align investment with strategic objectives.
Integrate ABAF into Procurement: Revise vendor assessment processes to require an "ABAF profile," forcing clarity on function, data dependency, and operational fit over opaque technical jargon.
Use ABAF for Internal Literacy: Build upskilling programs around ABAF's core concepts to create a shared, practical understanding of AI's functional risks and benefits across all corporate levels.

6.2 A Call for Policy Integration

Regulators must move away from monolithic or algorithm-specific rules. We call on policymakers to:

Use ABAF as a Governance Blueprint: Leverage the framework's functional categories as the basis for tiered, risk-based regulatory models that are both effective and innovation-friendly.
Fund Pilot Programs: Test the utility of ABAF as a practical governance tool in high-stakes sectors, such as medical device approval or financial model risk management, to refine its application.
Promote ABAF as a Standard: Champion ABAF as the common lexicon for public-private dialogue, ensuring all stakeholders are engaging in a coherent and constructive conversation.

The "functional turn" is not merely an academic proposal; it is an essential strategic and operational pivot. Adopting a shared, functional language for AI is the first step toward building a more mature, responsible, and impactful technological future.

References

Amodei, D., Olah, C., Steinhardt, J., Christiano, P., Schulman, J., & Mané, D. (2016). Concrete Problems in AI Safety. arXiv:1606.06565. https://arxiv.org/abs/1606.06565
AI Classification Standard (AICS). (n.d.). An Open Framework for Classifying AI Systems. Accessed October 26, 2025. https://4hum-ai.github.io/aics/
Ballester, O. (2021). An Artificial Intelligence Definition and Classification Framework for Public Sector Applications. In Proceedings of the 22nd Annual International Conference on Digital Government Research (dg.o '21). https://doi.org/10.1145/3463677.3463706
BeResponsibleAI. (2025). A Responsible AI Suite. Accessed October 25, 2025. https://beresponsibleai.com
Blaß, M., Gimpel, H., & Karnebogen, P. (2024). A Taxonomy and Archetypes of AI-Based Health Care Services: Qualitative Study. Journal of Medical Internet Research, 26, e51740. https://www.jmir.org/2024/1/e51740
Bradley, W. A. (2025). Global AI Governance: Five Key Frameworks Explained. Bradley Arant Boult Cummings LLP. https://www.bradley.com/insights/publications/2025/08/global-ai-governance-five-key-frameworks-explained
Caruana, R., Lou, Y., Gehrke, J., Koch, P., Sturm, M., & Elhadad, N. (2015). Intelligible Models for HealthCare: Predicting Pneumonia Risk and Hospital 30-day Readmission. In Proceedings of the 21st ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (pp. 1721–1730). https://doi.org/10.1145/2783258.2788613
CITI IO. (2023). Not All AI Are Created Equal – A Taxonomy of Artificial Intelligence Systems and Capabilities. https://www.citigps.com/publication/e30N6H7-Not-All-AI-Are-Created-Equal
Colelough, B. C., & Regli, W. (2025). Neuro-Symbolic AI in 2024: A Systematic Review. arXiv:2501.05435. https://arxiv.org/abs/2501.05435
Hacker, P., Engel, A., & Katic, M. (2024). Can there be responsible AI without AI liability? Incentivizing generative AI safety through ex-post tort liability under the EU AI liability directive. International Journal of Law and Information Technology. https://academic.oup.com/ijlit/article/doi/10.1093/ijlit/eaae021/7758252
Lewis, P., Pérez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Kuncoro, A., Yih, W., & Zettlemoyer, L. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. arXiv:2005.11401. https://arxiv.org/abs/2005.11401
McMahan, B., Moore, E., Ramage, D., Hampson, S., & y Arcas, B. A. (2017). Communication-Efficient Learning of Deep Networks from Decentralized Data. arXiv:1602.05629. https://arxiv.org/abs/1602.05629
Mökander, J., Sheth, M., Watson, D., & Floridi, L. (2024). The Switch, the Ladder, and the Matrix: Models for Classifying AI Systems. Minds and Machines. https://doi.org/10.1007/s11023-024-09670-z
Newman, J. (2023). A Taxonomy of Trustworthiness for Artificial Intelligence. Center for Long-Term Cybersecurity, UC Berkeley. httpsF://cltc.berkeley.edu/2023/12/12/a-taxonomy-of-trustworthiness-for-artificial-intelligence/
NIST. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0). National Institute of Standards and Technology. https://nvlpubs.nist.gov/nist.pubs/ai/NIST.AI.100-1.pdf
OECD. (2022). OECD Framework for the Classification of AI Systems. OECD Digital Economy Papers, No. 323. https://doi.org/10.1787/cb6d9f57-en
RadIQ Pro. (2025). RadIQ Pro AI - Advanced Medical Imaging Planner. Accessed October 25, 2025. https://radiqpro.in/
Shinde, S. M. (2024). A Comprehensive Overview of Artificial Intelligence-Based Classification Techniques. International Journal of Science and Research Archive, 11(1), 40-47. https://ijsra.net/sites/default/files/IJSRA-2024-0387.pdf
Slattery, P., Saeri, A.K., Grundy, E.A.C., et al. (2024). The AI Risk Repository: A Comprehensive Meta-Review, Database, and Taxonomy of Risks from Artificial Intelligence. arXiv:2404.09453. https://arxiv.org/abs/2404.09453
Theofanos, M., Choong, Y.Y., & Jensen, T. (2024). AI Use Taxonomy: A Human-Centered Approach (NIST.AI.200-1). National Institute of Standards and Technology. https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.200-1.pdf
Vilone, G., & Longo, L. (2021). Explainable Artificial Intelligence: A Review of Machine Learning Interpretability Methods. PeerJ Computer Science, 7, e530. https://pmc.ncbi.nlm.nih.gov/articles/PMC7824368/
Xia, B., Lu, Q., Zhu, L., & Xing, Z. (2024). An AI System Evaluation Framework for Advancing AI Safety: Terminology, Taxonomy, Lifecycle Mapping. arXiv:2403.07872. https://arxiv.org/abs/2403.07872

Functional Turn in AI Classification: An Approach-Based AI Framework (ABAF)