Lilly Tech Systems
Types of Artificial Intelligence
AI can be categorized by capability (how much it can do) and by functionality (how it processes information). Understanding these categories clarifies what today's AI can and cannot do.
Classification by Capability
Narrow AI (Artificial Narrow Intelligence — ANI)
Narrow AI is designed and trained for a specific task. It can perform that task extremely well — often better than humans — but cannot transfer its skills to other domains. All current AI systems are narrow AI.
- Siri and Alexa (voice assistants)
- Google Search and recommendation algorithms
- Self-driving car perception systems
- ChatGPT, Claude, and other LLMs (despite their versatility, they are still narrow AI)
- AlphaGo and chess engines
General AI (Artificial General Intelligence — AGI)
AGI would possess human-level intelligence across all cognitive tasks. An AGI system could learn any intellectual task that a human can, transfer knowledge between domains, and reason about novel situations without specific training.
- Does not exist yet
- Would require understanding, common sense, and flexible reasoning
- Timeline estimates vary wildly: some researchers predict 10–30 years, others say it may never be achieved
Super AI (Artificial Superintelligence — ASI)
ASI would surpass human intelligence in every conceivable way — creativity, problem-solving, social intelligence, and scientific reasoning. This is purely theoretical and raises profound existential questions.
Classification by Functionality
An alternative taxonomy classifies AI by how it processes information and what it can remember:
| Type | Description | Memory | Examples | Status |
|---|---|---|---|---|
| Reactive Machines | Respond to inputs with no memory of past interactions | None | IBM Deep Blue, basic spam filters | Exists |
| Limited Memory | Can use recent past data to inform decisions | Short-term | Self-driving cars, chatbots with conversation history, LLMs | Exists |
| Theory of Mind | Can understand emotions, beliefs, and intentions of others | Social | Would understand that a human is frustrated or confused | Theoretical |
| Self-Aware | Has consciousness, self-awareness, and sentience | Full | Science fiction (HAL 9000, Data from Star Trek) | Theoretical |
Reactive Machines
The simplest form of AI. These systems perceive the current situation and respond based on predefined rules or learned patterns. They have no concept of the past or future.
Example: IBM's Deep Blue analyzed chess positions and calculated optimal moves. It had no memory of previous games and no understanding of chess as a concept — just raw computation of board positions.
Limited Memory
These systems can store and use past experiences for a limited time. Most modern AI falls into this category.
Example: A self-driving car observes the speed and direction of surrounding vehicles over the last few seconds to predict their future positions and make safe driving decisions. LLMs use conversation history (context window) to maintain coherent dialogue.
Theory of Mind
This level of AI would understand that other entities have their own thoughts, emotions, and beliefs. It could adjust its behavior based on what it infers others are thinking or feeling.
Some researchers argue that advanced LLMs show rudimentary theory of mind capabilities, but this remains a topic of active debate. True theory of mind would require genuine understanding of mental states, not just pattern matching.
Self-Aware AI
The most advanced (and purely hypothetical) type. A self-aware AI would have consciousness — it would know it exists, have subjective experiences, and understand its own internal states. This raises profound philosophical questions about the nature of consciousness itself.