Textbook of AI

Narrow AI, also called weak AI or applied AI, refers to artificial-intelligence systems designed and trained for a particular, well-defined task. The term was coined by philosopher John Searle in his 1980 paper Minds, Brains, and Programs, the same paper that introduced the Chinese Room argument, in deliberate contrast to strong AI, the hypothesis that a suitably programmed machine could possess genuine understanding and consciousness.

Examples

Almost every AI system that has ever been deployed in production is narrow AI:

• A spam filter that classifies email as spam or not-spam
• A face-recognition system that matches images to identities
• A chess engine such as Stockfish or AlphaZero
• A medical-image classifier that flags suspicious mammograms
• A recommendation system that suggests films or products
• A machine-translation system that converts French into English

Each of these can be highly effective within its domain, often surpassing human performance, but none has any understanding of, or capability outside, that domain. The Go-playing AlphaGo cannot translate French; the French-translating Google Translate cannot drive a car; the car-driving Waymo cannot answer medical questions.

Are large language models narrow AI?

Large language models occupy a contested position in this taxonomy. Superficially, models like GPT-4, Claude, and Gemini appear general , they produce fluent text on virtually any topic, write code, solve mathematical problems, and engage in apparent reasoning. They are sometimes characterised as a step beyond pure narrow AI ("artificial general intelligence-lite", or proto-AGI).

Most researchers, however, classify them as narrow systems whose narrow task happens to be next-token prediction over web-scale text. Their capabilities are bounded by the distribution of text they were trained on and the patterns they extracted from it. They lack persistent memory across interactions, embodied grounding in the physical world, and the goal-directed agency that proponents of AGI consider essential. The 2023 emergence of agent frameworks (AutoGPT, Devin, Claude Code) that wrap LLMs with tool use and persistent memory is a step towards more general systems but does not by itself transform the underlying model.

Why the distinction matters

The narrow/general distinction is central to debates about AI's future trajectory and current risk. Narrow systems pose concrete present-day risks and benefits, job displacement, algorithmic bias, privacy erosion, improved diagnostics, scientific acceleration, while general AI raises more speculative but potentially more severe questions about alignment, control, and the long-term future of humanity. Conflating the two muddles policy debates: critics of "AI doomerism" sometimes accuse AGI safety researchers of treating narrow AI risks as proof of imminent superintelligence; conversely, AGI researchers sometimes accuse narrow-AI critics of dismissing existential concerns by pointing to the obvious limitations of today's systems.

Understanding which kind of system one is discussing is essential for reasoning clearly about AI policy, AI safety, and AI's likely social impact.