The target is not a theorem prover whose formal system is merely the object being studied. It is a programme that manipulates formalised, practical statements: where things are, what actions are feasible, what those actions change and what somebody wants. A basic routine receives premises and writes down immediate conclusions. Some conclusions assert facts. Others are imperatives, and the system obeys them by printing, moving sentences between lists or restarting deduction over a newly selected list.

This architecture turns “common sense” from praise into a behavioural criterion. McCarthy proposes calling a programme common-sensical when it automatically obtains a sufficiently wide class of immediate consequences from new information together with old knowledge. The qualifier “sufficiently wide” matters. The paper does not claim complete first-order consequence, nor does it equate common sense with omniscience. It asks for enough readily available consequence to support flexible instruction and ordinary action.

The contrast with Newell, Shaw and Simon’s Logic Theory Machine and Gelernter’s geometry work is architectural. In those earlier systems, according to McCarthy’s comparison, the formal system is the subject matter while heuristic guidance is embedded in the programme. In the Advice Taker, procedures and especially heuristics are to be described, as far as possible, inside the same language that contains ordinary knowledge. Advice can then alter how the system reasons, not only the facts upon which it reasons.

That phrase “as far as possible” prevents a stronger but tempting reading. A basic interpreter, storage machinery, observation routines and deduction routine still exist as code. The proposal is to move improvable, domain-sensitive guidance into inspectable sentences. It is an early design for separating a general reasoning substrate from knowledge and control information that a user can state.

McCarthy compares imperative and declarative instruction without pretending that one simply dominates. Imperatives arrive already ordered and can execute quickly. They can begin from a known basic state, assuming little prior knowledge. That is their engineering strength: the procedure says what to do next. But its effect is sensitive to order and to the precise state from which execution begins.

Declaratives offer four stated advantages. They can exploit previous knowledge. They have logical consequences that can be combined with earlier facts. Their meanings are less dependent on order, making afterthoughts easier. And their effects depend less on the machine’s previous state, so the instructor needs less knowledge of that state. The user describes circumstances and desired results; the reasoner supplies an applicable procedure.

The paper’s learning ambition rests on this difference. Contemporary learning programmes could adjust anticipated numerical weights or remember lost game positions. McCarthy wants a machine able to acquire a high-level abstraction that its programmer did not individually anticipate, such as a strategic principle in checkers. Before a programme can discover an abstraction, he argues, it needs a simple way to represent it. Language is offered as the known vehicle for expressing such abstractions.

The celebrated maxim follows: for a programme to learn something, it must first be capable of being told it. This is neither a definition of all learning nor a claim that human teaching is exclusively linguistic. It is a staging decision. First build a system whose behaviour can be specifically improved through intelligible statements; later, perhaps, tell that system how to learn from experience. Teachability becomes an intermediate test of representational adequacy.

McCarthy grants that a universal machine can represent arbitrary behaviour by simulating neural nets, Turing machines or calculator programmes. Universality is not enough. In a low-level space, interesting behaviours may be exceedingly sparse, and a conceptually small improvement can require a complicated structural change. Searching arbitrary behaviours then resembles looking for a meaningful organism by random alteration of blueprints: most changes break details rather than express a coherent new ability.

The proposed system should therefore satisfy several demands. Every behaviour must be representable, perhaps through a general-purpose language or arbitrary automata. Interesting behavioural changes must be simply expressible. Everything except the most routine behaviour should be improvable, including the improvement mechanism. The machine needs concepts of partial success because decisive outcomes on hard problems are rare. It should also form subroutines and recognise a potentially useful one before a particular context proves it good.

The paper concentrates mainly on the second demand: make an interesting change easy to say. This focus explains why symbolic language, lists, properties and formulas dominate the construction. McCarthy is not yet supplying the search process that invents high-level concepts. He is designing a representational landing place for them, and a channel through which a person can install one without intimate knowledge of code.

The consequence is a demanding notion of modularity. A statement should retain its intelligible content when combined with new statements, and a heuristic should be separately modifiable. Yet the system must still turn those independent pieces into timely action. Representation and control cannot be separated completely: what is easy to express affects what can be retrieved, matched and deduced.

First, the system recursively represents expressions. A term is an expression, and a sequence of expressions is also an expression; list structures provide the machine representation. Some expressions count as declarative sentences in a formal system resembling a universal Post canonical system. McCarthy anticipates a single inference rule that combines substitution for variables with modus ponens, avoiding a flood of separately stored instances of general propositions.

Second, an immediate-deduction routine takes a set of premises and returns one-step consequences. Its initial version is syntactic rather than subject-sensitive. It should not embody, for example, a special trick for airport travel. The paper locates intelligence in procedures that choose which lists of premises receive this routine. Applying deduction simultaneously to everything the programme knows would take too long.

Third, expressions have several roles. Not every formula is a proposition. Some name individuals, functions or programmes. Objects have property lists containing specific things the system has to say about them. Even a structurally derivable property may be cached if recomputation was difficult. McCarthy illustrates objecthood pragmatically: a number about which a person knows a historically specific fact functions as an object, while an unremarkable number may be understood only through its numeral structure.

Finally, the programme operates cyclically. It applies immediate deduction to selected premises and individuals. Some conclusions are imperatives, which it obeys. Among the possible imperatives is one that invokes deduction and obedience again. Reasoning can therefore reorganise the context for further reasoning. The loop joins symbolic assertion, control and external action without claiming that every imperative is safe merely because it is derivable.

A large knowledge base turns logically simple inference into an attention problem. The desired premises may already be stored, yet deduction fails if they are not brought together. McCarthy separates two questions in the airport example: where the seventeen premises come from, and how deductions proceed once those premises are found. He explicitly leaves both incomplete and notes that retrieval and deduction may interleave.

The tentative control structure begins with a main list L containing the goal. An observation routine describes the form of that list: for example, that its only statement resembles a desire for a relation to hold. The system combines this meta-level observation with a small standing-orders list. A derived imperative then tells it which property lists to inspect. The content of reasoning helps decide where reasoning should look next.

Retrieval moves from specific to general. The exact goal expression may have no property list because it has not become an object. A pattern such as wanting oneself to be at some location may store travel rules and an instruction to observe the current location. A still more abstract pattern, wanting any proposition, can store generic rules about actions that ultimately achieve desired conditions. Search climbs abstraction levels until it finds reusable knowledge.

The route problem also requires working backwards and upwards through place relations. Property lists for origin, destination and action patterns expose candidates. A driving rule requires being at the car; that new subcondition triggers retrieval of a walking rule. Side effects of search supply containment facts. This is only a sketch, but it identifies a lasting problem: reasoning power depends not just on consequence rules, but on how memory indexes them and how goals generate queries.

The initial place facts put the person at a desk, the desk and car at home, and both home and airport in a county. A transitivity rule for the stipulated place-within-place relation yields broader locations. Two feasibility rules say, roughly, that movement between places inside a walkable region can be walked, while movement inside a drivable region can be driven if the person and car are co-located. Home is walkable; the county is drivable.

An action-effect rule states that after going from x to y, the person is at y. The problem is introduced as a desire that the person be at the airport. Generic achievement rules relate a condition, an action possible under that condition and a resulting condition. If one achievable transition ends where another begins, their actions can be composed into a programme. If a desired result is achievable by a programme whose first action is y, an imperative orders y.

The intended derivation first recognises that, from the desk, walking to the car is possible. That action would place the person at the car. From there, driving from home to the airport is possible and would place the person at the airport. Each action becomes an “ultimately achieves” relation. Transitivity of achievement composes them. The goal then activates the rule whose conclusion is to do the first action.

The notation is provisional. McCarthy remarks that identifying a single action with a one-step programme still needs formal work. The derivation order is “more or less” as displayed, and the larger issue of collecting its premises is unresolved. The example is best read as an architecture test: can facts, generic action knowledge, a goal and a small amount of control knowledge jointly produce an appropriate first step?

The laboratory uses forward chaining because the paper’s immediate-deduction routine writes one-step consequences from a selected premise list. It begins with ground facts. At each pass it fires every rule whose antecedents are already present, records new conclusions and stops only when a full pass adds nothing. The displayed order is therefore an audit trail for one fair scheduling of enabled rules.

The reconstruction distinguishes possibility, effect knowledge, composition and execution. “Walk possible” does not assert that walking happened. Instead, an achievement link packages an applicable action with its stated effect. Two links compose only when the result condition of the first matches the starting condition of the second. A goal plus the composed plan enables the imperative “do walk”.

Three exact invariants are evaluated on every scenario. Every derived atom must cite a rule whose antecedents are all present. An imperative must have both the airport goal and a composed achievable plan as direct support. Finally, the reported closure must be a fixed point: no rule remains enabled with an omitted conclusion. These are tests of the reconstruction, not empirical evidence that a car journey is safe or optimal.

Removing a premise shows causal dependence. Without the car-location fact, the walking method cannot be selected. Without walkability, the first link fails. Without drivability, the second link fails. In each case the engine still derives whatever other consequences have support, but it withholds action because the complete chain to the goal is absent.

Bar-Hillel questions treating “at” as transitive when ordinary spatial neighbourhood is not. If A is near B and B near C, A need not be near C. He also argues that walking and driving are not the only reasonable choices: a taxi, a cancelled flight or many other actions could intervene. Time, distance, speed and deadlines are absent. A practically sound model would require far greater complexity.

The most consequential objection concerns relevance. Saying that a human-level navigator may possess the needed facts does not explain how a machine selects them from millions. Bar-Hillel regards this division of the universe into relevant and irrelevant facts as much harder than the propositional deduction heuristics then available. That challenge strikes precisely where McCarthy has placed the system’s intelligence.

McCarthy replies that “at” is a mnemonic for a stipulated relation between place and sub-place, not a formalisation of every English use of “at”. The airport story is not offered as a practical navigation problem; it is a simplified case for thinking about what reasoning and premise collection require. He agrees that a practical model needs more complexity but believes the conceptual difficulties already appear at the simplified level.

His reply also clarifies the epistemic status of the result. Logic is intended to be faultless, but premises cannot be guaranteed. The conclusion is the command to perform the first walking action, not the desired fact that the person is at the airport. Deduction can recommend an action under a model; only execution and observation can establish what subsequently holds.

The paper proposes a language of expressions, declarative and imperative interpretations, a substitution-plus-modus-ponens style inference rule, an immediate-consequence routine, property lists, meta-level observations, standing orders and a recurrent deduce-and-obey cycle. It gives a compact operational criterion for common sense and connects teachability to the representation required for more ambitious learning.

It does not report an implemented Advice Taker, benchmark results, timing data or a proof of completeness. It does not solve premise relevance, natural-language understanding, inconsistent knowledge, default reasoning, belief revision, uncertainty, perception or safe physical action. The particular logical formalism and the representation of many heuristic statements remain unsettled. McCarthy says he hopes to formalise the heuristics in another paper before programming the system.

The airport chain assumes a curated model. Its walking and driving rules ignore obstacles, ownership, fuel, schedules and costs. Its relation symbols receive stipulated formal meanings. Its goal is singular and unqualified. These are not hidden empirical findings; they are abstractions selected to expose the interfaces between stored knowledge, premise retrieval, deduction, programme composition and action.

A fair reading therefore holds two ideas together. The proposal is remarkably ambitious because it demands that domain and heuristic knowledge become tellable, composable and revisable. It is also candidly incomplete. The central bottleneck is not disguised by a successful toy derivation: the paper explicitly asks how the relevant premises are assembled and postpones a complete answer.

Generality comes from variables, abstractions and declarative statements. A rule about wanting to be at an arbitrary location can serve many journeys. Transparency comes from representing reasons as premises and deductions, so an action can be traced to a goal, feasibility facts and effect rules. Control comes from main lists, observations, standing orders and property-list retrieval, which restrict attention to a manageable context.

Improving one dimension can damage another. Highly general rules match many situations and increase search. Aggressive retrieval improves speed but can omit a decisive premise. More procedural shortcuts improve control yet move knowledge back into opaque code. A larger declarative store is easier to extend locally but harder to keep consistent and relevant. The paper does not dissolve these trade-offs; it gives them an architecture in which they can be separately discussed.

Later work changed the vocabulary and machinery. McCarthy and Hayes developed a richer philosophical and logical account of situations and actions. Non-monotonic logics addressed conclusions that should be withdrawn when exceptions appear. Knowledge-representation systems built explicit taxonomies and roles. Contemporary systems can emulate multi-step reasoning over natural-language facts and rules, while modern commonsense research emphasises scale, uncertainty and social or physical knowledge. None makes the original relevance question disappear.

The most durable lesson is methodological. A capable agent needs more than stored propositions and more than a valid inference rule. It needs a representational language in which useful improvements are compact, a memory organisation that can expose relevant knowledge, control rules that decide which consequences to seek, and a disciplined bridge from reasons to actions. These components should be judged separately even when a modern system implements them with very different techniques.