Types of AI

01. What It Is

"Types of AI" is an umbrella phrase covering several distinct ways researchers and practitioners sort AI systems into categories. The confusion in popular coverage usually comes from treating these as a single ranked list, when they are actually orthogonal lenses. A large language model (LLM) like GPT-4o or Claude is simultaneously a narrow AI (by capability), a limited-memory machine (by functionality), a machine-learning system (by approach), and a generative or conversational AI (by function). Each classification tells you something different and none of them contradicts the others.

02. Why It Matters

Knowing which axis a claim is using prevents common errors. "AI will surpass human intelligence" is a capability claim, not a functional one. "This model remembers your conversation" is a functionality claim, not a capability one. Mismatched axes produce nonsense comparisons and feed the hype cycles that make it harder to reason clearly about what AI can and cannot do.

03. By Capability (ANI, AGI, ASI)

This axis asks: how broad and transferable is the system's intelligence?

Artificial Narrow Intelligence (ANI)

Also called weak AI, ANI describes a system that performs well within a specific domain but cannot generalize to tasks it was not designed or trained for. A chess engine cannot write code. A speech recognizer cannot drive a car. A protein-folding model cannot compose music.

Every AI system deployed commercially as of June 2026 is ANI. This includes the most capable frontier models. IBM puts it plainly: narrow AI is intelligence "demonstrated only in specialized domains." The term does not imply the systems are simple or unimpressive. AlphaFold predicts protein structures better than any human scientist. Deep Blue defeated the world chess champion in 1997. These are achievements of narrow intelligence operating at superhuman levels within constrained problem spaces.

Artificial General Intelligence (AGI)

AGI refers to a hypothetical system that can match or surpass human cognitive capabilities across virtually any task, generalizing knowledge between domains and solving novel problems without task-specific reprogramming. Wikipedia's definition: "a hypothetical type of artificial intelligence that matches or surpasses human capabilities across virtually all cognitive tasks."

The term was used early by Mark Gubrud in 1997 and popularized in the early 2000s (around 2002) by Shane Legg and Ben Goertzel. The 2007 book "Artificial General Intelligence" was an edited volume by Ben Goertzel and Cassio Pennachin. There is no consensus on what exactly would count as AGI. A 2023 Google DeepMind paper (Morris et al., published at ICML 2024) identified at least eight competing frameworks for defining AGI, ranging from passing the Turing Test to performing economically valuable work to demonstrating flexible, cross-domain capability. The authors proposed a levels-of-AGI framework based on both performance depth and generality breadth.

OpenAI defines AGI as "highly autonomous systems that outperform humans at most economically valuable work," which critics including Meta's chief AI scientist Yann LeCun have called too narrow. LeCun has argued that current LLMs lack common sense, cannot plan hierarchically, and cannot learn from embodied experience, making them far short of AGI regardless of benchmark scores.

A survey of 2,778 AI researchers published in January 2024 (Grace et al.) estimated a 50% probability of "unaided machines outperforming humans in every possible task" by 2047. That figure is 13 years earlier than the same survey found a year prior, reflecting the acceleration seen since ChatGPT's 2022 launch. Treat all such forecasts with skepticism. Expert predictions about their own field have a poor historical track record.

Artificial Superintelligence (ASI)

ASI is a further hypothetical in which an AI system outperforms the best humans across every cognitive domain by a significant margin, not just on average. IBM frames it clearly: AGI is a prerequisite for ASI, but ASI is not required for AGI. A system with average human-level general intelligence would be AGI without being ASI.

Nick Bostrom's book "Superintelligence" (2014) introduced ASI to a wide audience and made the case that a sufficiently advanced ASI could pose existential risks. This position remains genuinely contested in the research community. Critics such as Gary Marcus and Melanie Mitchell argue the path from current systems to anything approaching ASI is far less linear than Bostrom's arguments assume. No credible technical roadmap exists for achieving ASI.

04. By Functionality (The Four Types)

This axis, widely attributed to AI researcher Arend Hintze in a 2016 essay, asks: what cognitive mechanisms does the system use? It describes a rough developmental progression, though the later types remain speculative.

Type 1: Reactive machines

Reactive machines perceive their current environment and respond to it, but maintain no memory of past interactions and build no internal model of the world. IBM's Deep Blue is the canonical example: it evaluated chess positions and selected moves without storing anything between games or learning from past opponents. IBM's Watson playing Jeopardy! fits this description at the retrieval level.

Reactive machines are predictable and auditable. Their behavior is entirely a function of current inputs and fixed rules or weights.

Type 2: Limited memory

Limited-memory systems can access a window of recent information, either from a rolling context buffer, a retrieval system, or training data, and use it to inform current decisions. Most modern machine learning systems fall here, including every deployed LLM as of 2026.

The "limited" qualifier matters. An LLM's context window is not long-term memory; it is ephemeral state that disappears when the session ends. Retrieval-augmented generation (RAG) systems extend this by querying external databases, but the system still cannot independently form and retain new memories in the way a human brain does. Self-driving vehicles, recommendation engines, and predictive maintenance systems all belong in this category.

Type 3: Theory of mind

Theory-of-mind AI would understand that other agents (humans, other AIs) have their own beliefs, intentions, and mental states, and would reason about those states to guide its behavior. This capability exists partially and unevenly in frontier models as of 2026. LLMs can pass some theory-of-mind benchmarks while failing others that require tracking nested mental states over time. No deployed system reliably maintains accurate models of interlocutor intent across extended, varied interactions. This remains an active research area rather than a solved problem.

Type 4: Self-aware AI

Self-aware AI would have genuine consciousness, subjective experience, and an internal model of its own cognitive processes. This type exists only as a philosophical and science-fiction concept. There is no agreed definition of machine consciousness, no empirical test for it, and no credible engineering path toward it. The debate around the Chinese Room argument (John Searle, 1980) illustrates that the philosophical questions remain unresolved even in principle.

05. By Approach (Symbolic, Machine Learning, Hybrid)

This axis asks: how does the system represent knowledge and arrive at conclusions?

Symbolic AI (GOFAI, rule-based, expert systems)

Symbolic AI, sometimes called Good Old-Fashioned AI (GOFAI), represents knowledge as explicit human-readable symbols, rules, and logic. Programs encode "if-then" production rules, ontologies, semantic networks, or formal logic. Examples include MYCIN (medical diagnosis, 1970s), XCON (computer configuration, 1980s), and modern computer algebra systems.

Symbolic AI dominated the field from the mid-1950s through the mid-1990s. Its strengths are interpretability and the ability to apply formal reasoning guarantees. Its weaknesses are brittleness outside the designed domain, the "knowledge acquisition bottleneck" (encoding human expertise is expensive and slow), and poor handling of uncertainty and ambiguous perception data. Two AI winters, in the 1970s and late 1980s, resulted partly from over-promises and the practical limits of purely symbolic approaches.

Symbolic AI has not disappeared. Automated theorem provers, formal verification tools, and knowledge graphs remain active and useful. The field continues in hybrid forms.

Machine learning and neural networks

Machine learning systems learn statistical patterns from data rather than following hand-coded rules. Deep learning, the dominant subparadigm since roughly 2012, uses multi-layer neural networks loosely inspired by the structure of biological brains. Backpropagation allows networks to adjust weights across millions or billions of parameters to minimize prediction error on training data.

The 2012 breakthrough, when Krizhevsky, Sutskever, and Hinton's AlexNet dramatically outperformed prior methods on image classification, began the current era. Subsequent advances in transformer architectures (Vaswani et al., 2017) enabled the LLMs at the center of the 2020s AI boom.
For a deeper look at LLMs specifically, see What Is a Large Language Model? and AI, ML, Deep Learning, LLMs, and Algorithms: The Differences.

Machine learning is powerful but opaque. Models learn correlations that may not correspond to causal structure, can fail unexpectedly on out-of-distribution inputs, and are difficult to audit. These are active areas of research in interpretable ML and AI safety.

Neuro-symbolic and hybrid approaches

Neuro-symbolic AI attempts to combine the pattern-recognition strengths of neural networks with the structured reasoning and interpretability of symbolic systems. This is not a single architecture but a research direction. Examples include systems that use neural components for perception and symbolic components for planning, or LLMs augmented with external tools, databases, and formal reasoners.

As of 2026, most frontier AI products are implicitly hybrid in practice: an LLM (neural) is paired with web search, code execution, or structured data retrieval (symbolic or rule-based). Pure approaches at either extreme are relatively rare in deployment.

06. By Function (Predictive, Generative, Agentic, Conversational)

This axis asks: what task does the system actually perform for users? This is the most practically useful framing for understanding deployed AI products.

Predictive AI

Predictive AI analyzes historical data to forecast future outcomes, classify inputs, or detect anomalies. It draws on regression models, decision trees, ensemble methods, and neural networks. Applications include demand forecasting, fraud detection, predictive maintenance, patient risk scoring, and credit assessment. IBM describes it as using "statistical analysis and machine learning to identify patterns, anticipate behaviors and forecast upcoming events."

Predictive AI has existed in industrial and financial settings for decades under names like "analytics" and "data mining." The current wave of deep learning expanded its scale and accuracy without fundamentally changing its purpose.

Generative AI

Generative AI produces new content, including text, images, audio, video, and code, from learned distributions over training data. The generative AI category grew dramatically after 2022 with the public release of ChatGPT, image models like Stable Diffusion and Midjourney, and multimodal systems.
For a dedicated treatment of this category, see Generative AI.

The technical mechanisms differ by modality. LLMs predict the next token in a sequence. Diffusion models iteratively denoise random inputs toward target distributions. Both are probabilistic at their core, meaning outputs are samples rather than deterministic lookups.

Agentic AI

Agentic AI describes systems that pursue goals over multiple steps, using tools and taking actions with varying degrees of autonomy. The control flow is typically driven by an LLM that can call external tools, execute code, browse the web, or spawn sub-agents. Wikipedia notes that NIST describes agentic AI as "an emerging area requiring standards for secure operation, interoperability, and reliable interaction with external systems."

As of mid-2025, the primary proven use cases for agents are software development (coding agents like Cursor) and customer service automation. The Financial Times compared agent autonomy to SAE self-driving levels, placing most current applications at levels 2-3 with a small number reaching level 4 in highly constrained scenarios. Level 5 (full autonomy) remains theoretical.
For a fuller treatment, see Agents and Agentic Workflows.

Conversational AI

Conversational AI handles natural language dialogue for information retrieval, task completion, and customer interaction. Chatbots, virtual assistants (Alexa, Siri, Google Assistant), and enterprise support tools fall here. Modern conversational AI is almost always built on LLMs, which gives it broad language understanding and fluent output, but also makes it susceptible to hallucination and inconsistency.

Conversational AI and generative AI overlap substantially in 2026. The distinction is primarily one of design intent: conversational systems prioritize dialogue coherence and task completion, while generative systems prioritize content production.

07. Where Today's Models Fit

A current frontier model such as GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro, or Llama 3 sits in the following positions across the four axes:

• Capability: ANI. These models are impressive but task-constrained in ways that disqualify them from AGI by any rigorous definition. They cannot autonomously learn new domains from experience, maintain persistent memory across sessions without engineering scaffolding, or reliably perform physical tasks.
• Functionality: Limited memory (Type 2), with partial and inconsistent theory-of-mind behavior on benchmarks.
• Approach: Connectionist (deep learning, transformer architecture), often combined in practice with symbolic tools via function calling, RAG, and code execution.
• Function: Generative, conversational, and increasingly agentic when deployed with tool access.

The DeepMind levels-of-AGI paper proposes placing these models at "Emerging AGI" (Level 1) on their scale, noting they perform at or above median human level on a range of tasks but fall well short of expert human performance across all domains. That framing remains contested. LeCun and others argue that performance on language benchmarks does not capture the grounded, embodied, continual-learning capabilities that matter for genuine general intelligence.

08. Summary Table

09. Common Misconceptions

"ChatGPT is AGI."
No credible definition of AGI is satisfied by current LLMs. They do not learn from new experiences at runtime, cannot generalize to arbitrary physical or social tasks, and have no persistent cross-session memory without engineering scaffolding.

"AGI is coming in five years."
Forecasts in this field have been wrong repeatedly and systematically. The 2024 Grace et al. survey median was 2047. Treat any specific near-term date as an opinion, not a fact.

"Narrow AI is weak or primitive."
ANI systems routinely outperform humans at specific tasks by enormous margins. "Narrow" describes scope, not quality.

"The four functionality types are a development roadmap."
Hintze's framework describes a conceptual progression but was not a formal research agenda. There is no guaranteed path from Type 2 to Type 3 and no agreed engineering approach to get there.

"Generative AI and AI are the same thing."
Generative AI is one functional category within a much larger field. Predictive AI in healthcare, logistics, and finance has been running at industrial scale for decades without generating a single sentence or image.

"Symbolic AI is dead."
Formal logic, knowledge graphs, automated theorem provers, and constraint solvers remain in active use. They are often the components that make neural systems more reliable and auditable.