Weak AI
Weak AI is the position that computers and programs are valuable instruments for studying and modeling the mind, without any claim that a suitably programmed machine actually has a mind, understands, or is conscious. The label comes from the philosopher John Searle, who introduced the weak/strong distinction in his 1980 paper "Minds, Brains, and Programs" in Behavioral and Brain Sciences. On his reading, weak AI says the principal worth of the computer in the study of the mind is that it gives researchers a powerful tool: a hypothesis about, say, how language is parsed can be written as a program, run, and tested. The simulation is a model, in the same way a computer simulation of a rainstorm is a model and leaves everything dry.
Searle's weak AI is a claim about the status of a simulation, not about how capable the machine is. In the same paper he defined strong AI as the view that an appropriately programmed computer literally is a mind, in the sense that it can be said to understand and have other cognitive states (that thesis has its own entry, strong-ai). His Chinese room thought experiment was aimed only at the strong claim: a person who follows a rulebook to manipulate Chinese symbols can produce fluent answers without understanding Chinese, so, he argued, running a program is not sufficient for understanding. Weak AI survives the argument untouched, because it never asserted that the program understands anything. A system could in principle match human performance on every test and still be, on Searle's view, a weak-AI artifact.
The industry phrase "narrow AI" is a different distinction that is often confused with Searle's. Practitioners later began contrasting narrow systems, built to do one job such as spam filtering, route planning, or image classification, with the hoped-for general systems, a vocabulary strengthened in the early 2000s when the term artificial general intelligence was adopted, notably through the work of Ben Goertzel and Shane Legg. Narrow versus general concerns breadth of competence; weak versus strong concerns whether anything is understood. The two axes are independent. A narrow chess engine is both narrow and, on nearly everyone's view, weak in Searle's sense, but a very broad language model could still be weak in Searle's sense, and a hypothetical narrow system could still be said by a strong-AI believer to understand its single domain.
Textbook usage adds a third meaning, which is why the same words circulate with drifting senses. Stuart Russell and Peter Norvig, in Artificial Intelligence: A Modern Approach, describe the weak AI hypothesis as the claim that machines can act as if they were intelligent, and the strong AI hypothesis as the claim that machines that do so are actually thinking rather than simulating thinking. That reading moves the weak position from a stance about research method to a stance about possibility: intelligent behavior is achievable. They then observe that most AI researchers take the weak hypothesis for granted and do not much care about the strong one, because progress on building useful systems does not wait on settling the metaphysics. This pragmatic shrug is probably the dominant working attitude in the field.
The weak stance has real scientific content, and it was productive long before it had a name. Newell and Simon treated programs as theories of human problem solving, Marr's computational level of analysis in vision treated an algorithm as a hypothesis, and cognitive scientists still write models whose success is judged by how well they reproduce human errors and reaction times. In each case the program earns its keep by being precise, runnable and falsifiable, not by being conscious. The same attitude underlies modern machine learning engineering: a classifier is judged by its measured accuracy on held-out data and its failure modes, and nobody needs a theory of inner experience to ship it.
A common misconception is that weak means mediocre, or that weak AI is merely the primitive precursor of strong AI. In Searle's usage weak describes the philosophical claim, and a weak-AI system can be arbitrarily powerful. Another confusion treats the labels as a roadmap, as if enough narrow systems stacked together would become strong AI; the distinction implies no such progression, and neither does it settle whether more scale or new architectures would change the picture. A further pitfall is the reverse slide, from impressive behavior to the conclusion that understanding is present. Behavioral tests such as the Turing test measure performance and cannot by themselves decide between the weak and the strong reading, which is exactly the gap Searle's argument tried to exploit.
The debate reappeared with large language models, which produce fluent text without any agreed account of what, if anything, they understand. Critics describe them as statistical pattern completers, while others point to systematic generalization and internal representations as evidence of something more than surface mimicry. Weak AI remains the neutral engineering position in that argument: treat the system as an instrument whose outputs are to be measured, audited and used carefully, and leave open whether any inner life exists. That stance has practical value because it keeps safety, reliability and evaluation questions separate from questions about consciousness.
Read weak AI as the modest, instrument-first reading of artificial intelligence: build the program, test the model, use the tool, and do not mistake a faithful simulation of understanding for the thing itself.