The Advice Taker is a conceptual artificial intelligence system proposed by John McCarthy in his seminal 1958 paper "Programs with Common Sense." It was envisioned as a hypothetical program that could accept advice in the form of logical statements (axioms) and use deductive reasoning to improve its behavior over time, thereby exhibiting common sense. Unlike later expert systems, the Advice Taker was meant to learn and adapt from general knowledge rather than from hand-coded rules. Though never implemented in full, the concept profoundly influenced knowledge representation, automated reasoning, and the development of AI planning systems.

1 Introduction

1.1 Historical context: early AI and the Dartmouth conference

The 1956 Dartmouth Summer Research Project on Artificial Intelligence is widely considered the birthplace of AI as a field. Participants including John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon gathered to explore how machines could simulate human intelligence. Early efforts focused on problem-solving, theorem proving, and game playing, but these systems were narrowly specialized. McCarthy, dissatisfied with the ad-hoc nature of these programs, sought a more principled approach—one that could endow machines with general, flexible reasoning ability.

1.2 McCarthy's vision of a common-sense program

In his 1958 paper, McCarthy argued that existing AI programs lacked "common sense"—the ability to draw obvious conclusions from everyday knowledge. He proposed the Advice Taker as a program that would start with a small set of fundamental axioms and incorporate new advice (declarative statements) from human users. By deduction, it would infer consequences and adjust its behavior accordingly. The goal was not to pre-program every rule but to enable the machine to learn and reason from general principles, much like a person absorbing advice.

2 Theoretical foundations

2.1 Logical formalism and first-order predicate calculus

The Advice Taker relies on first-order predicate calculus (FOL) as its representational language. In FOL, knowledge is expressed as well-formed formulas (wffs) using predicates, quantifiers, logical connectives, and variables. For example, "All humans are mortal" might be written as ∀x (Human(x) → Mortal(x)). McCarthy chose FOL because it is declarative, unambiguous, and supports rigorous deduction via rules of inference such as modus ponens and universal instantiation.

2.2 The notion of "advice" as declarative knowledge

"Advice" in the Advice Taker is any declarative statement given to the program—not a procedure or algorithm, but a fact or rule that the system can use. A user might say, "If you are cold, you should turn on the heater." The program would first translate this into a logical axiom (e.g., Cold(x) → TurnOnHeater(x)) and then integrate it into its knowledge base. Over time, accumulating such advice builds a rich body of common-sense knowledge.

2.3 Distinction from procedural programming

Traditional programs specify *how* to do something step by step. The Advice Taker breaks from this by focusing on *what* is true. McCarthy emphasized that the system should be able to accept new facts without reprogramming its logic. This declarative approach foreshadowed later symbolic AI paradigms—expert systems, logic programming (e.g., Prolog), and knowledge bases—where the burden shifts from writing code to stating facts.

3 Proposed architecture

3.1 Knowledge base: axioms and facts

The core of the Advice Taker is its knowledge base, a store of axioms representing both general truths (e.g., "Every vehicle has a driver") and specific facts (e.g., "The car is in the garage"). Axioms are expressed in FOL and can be added incrementally. McCarthy imagined that the program would start with a minimal set of "basic knowledge" about time, space, and causality, and expand through advice.

The inference engine applies logical deduction to the knowledge base to answer queries and decide actions. Given a goal, the system searches for a proof that the goal follows from its axioms. For planning, it would derive sequences of actions that achieve desired states. McCarthy described a method called "situation calculus" (introduced in the same paper) to represent actions and their effects within the logical framework, enabling the program to reason about change over time.

3.3 Learning from advice: modifying behavior

Learning in the Advice Taker is not statistical but logical: when new advice is accepted, the system updates its knowledge base and then uses the updated axioms to reason differently. For example, if it learns that "pressing the brake stops the car," it incorporates that axiom and thereafter deduces that braking is appropriate when stopping is desired. This form of learning is monotonic in the sense that adding axioms never invalidates past deductions—a property that later proved problematic for handling default knowledge (see §4.1.2).

4 Implementation challenges

4.1 Representing common-sense knowledge

McCarthy's vision assumed that common sense could be captured in a finite set of logical axioms. However, real-world knowledge is vast, nuanced, and often uncertain. Representing concepts like intentions, beliefs, or exceptions (e.g., "birds usually fly but penguins do not") requires extensions to pure FOL, leading to research in non-monotonic logic, default reasoning, and probabilistic representations.

4.1.1 The frame problem

The frame problem is the difficulty of efficiently specifying what does *not* change when an action is performed. In FOL, one must explicitly list all unaffected facts—for every action, an enormous number of "frame axioms." The Advice Taker's situation calculus suffered from this explosion. Later work (e.g., by Hayes, Kowalski, and Reiter) proposed solutions such as explanation closure, circumscription, and the successor-state axiom, but the problem remains central to formal reasoning about change.

4.1.2 Non-monotonic reasoning

The Advice Taker assumed monotonicity: adding axioms never retracts old conclusions. But common sense often requires defeasible inferences; for instance, "Tweety is a bird" normally implies "Tweety can fly," but if we later learn "Tweety is a penguin," we must retract the earlier conclusion. This led to the development of non-monotonic logics (e.g., McCarthy's circumscription, Reiter's default logic) that allow conclusions to be revised as new information arrives.

4.2 Computational tractability of deduction

Full first-order deduction is semi-decidable—there is no algorithm that always finishes and gives a correct answer. For the Advice Taker to run in practice, deduction must be limited in scope. McCarthy recognized this and suggested using heuristics to guide search (like "look first at simple proofs"), but he did not specify how to implement such heuristics effectively. The trade-off between expressive power and computational efficiency remains a central challenge in automated reasoning.

5 Legacy and influence

5.1 Impact on knowledge representation research

The Advice Taker paper launched the field of knowledge representation. Its introduction of situation calculus inspired formalisms for reasoning about action and change (e.g., the event calculus, fluent calculus). The idea of a logical knowledge base that can be updated with new facts became foundational for deductive databases and ontology languages (e.g., OWL, RDF). McCarthy's emphasis on declarative representation later influenced the development of the General Problem Solver (Newell & Simon) and the logic programming paradigm.

5.2 Relationship to later expert systems

In the 1970s and 1980s, expert systems like MYCIN and DENDRAL adopted a rule-based approach (if–then rules) rather than full FOL deduction. While less general than the Advice Taker, they proved practical by restricting reasoning to narrow domains. The Advice Taker's notion of accepting "advice" in a declarative form is echoed in rule-based systems where domain experts provide heuristics. However, expert systems lacked the capacity for common-sense reasoning and learning from general statements—a limitation that McCarthy himself criticized.

5.3 Modern echoes: large language models and neuro-symbolic AI

Recent advances in large language models (LLMs) such as GPT-4 have revived interest in the Advice Taker's goals. LLMs can process natural-language advice and recall vast amounts of common-sense knowledge from training data, but they operate primarily through statistical pattern matching rather than deductive logic. Neuro-symbolic AI seeks to combine neural learning with symbolic reasoning—arguably fulfilling McCarthy's vision by allowing systems to accept logical axioms while learning from data. The Advice Taker remains a touchstone for researchers aiming to create machines with robust, flexible common sense.

6 See also

  • Common sense reasoning
  • Expert system
  • Logic programming
  • Non-monotonic logic
  • Situation calculus
  • Stanford Research Institute Problem Solver (STRIPS)

7 References

McCarthy, J. (1959). "Programs with Common Sense." *Proceedings of the Teddington Conference on the Mechanization of Thought Processes*, 756–791.

McCarthy, J., & Hayes, P. J. (1969). "Some Philosophical Problems from the Standpoint of Artificial Intelligence." *Machine Intelligence 4*, 463–502.

Reiter, R. (2001). *Knowledge in Action: Logical Foundations for Specifying and Implementing Dynamical Systems*. MIT Press.

Russell, S., & Norvig, P. (2021). *Artificial Intelligence: A Modern Approach* (4th ed.). Pearson. (Chapters on knowledge representation and automated reasoning.)