Newell delivered the argument as the first presidential address of the American Association for Artificial Intelligence in August 1980. The archival Carnegie Mellon report is dated July 1981; a magazine version appeared in 1981; the journal article governing this canon appeared in 1982. That publication history explains why different bibliographies attach different years to what is recognisably the same address.

The starting problem is a familiar ambiguity in artificial intelligence. Designers say that a programme knows a rule, a location or a repair procedure. Sometimes this means that a sentence occurs in memory. Sometimes it means that the programme can behave appropriately because some data structure and process together make the information effective. Newell wants a disciplined version of the second use.

His model is the hierarchy already used in computer engineering. A machine may be described at device, circuit, logic, register-transfer and symbol levels. Each level has an autonomous vocabulary and behavioural laws, yet must reduce to systems at the level below. A gate-level account and a circuit-level account are not rival stories; they expose different regularities and must agree where they meet.

The new proposal splits what practitioners had treated as one symbolic description. Below, the symbol level contains structures, programmes, search and access. Above, the knowledge level contains an agent, goals, actions and knowledge, bound by a rationality principle. Intelligent behaviour normally needs both accounts. The upper level says what action the content warrants; the lower one says how the machine manages to produce it.

This is meant as a hypothesis about a real regularity, not merely a convenient manner of speaking. That ambition is important. Newell knows the proposed level violates several features of familiar engineering levels. Rather than conceal the strain, he organises the central argument around four surprises: almost no structure, a global behavioural law, radical indeterminacy and no passive physical medium.

At the logic-circuit level, bits provide the medium, gates provide components, wiring supplies composition and Boolean relations describe behaviour. At the symbol level, expressions form the medium, operations manipulate them and programmes assemble those operations into larger systems. A specification at one level determines behaviour according to that level’s laws while a reduction explains how the same system is realised below.

Levels are neither arbitrary views nor claims that higher descriptions float free of hardware. A particular lower-level system may fail to realise any instance of the higher level. When it does realise one, the two accounts must match. Technologies make this regularity practical: stable devices realise bits, gates realise registers, and programmable machines realise symbolic operations.

Newell extracts four common features from familiar levels. A completely specified system yields definite behaviour. Overall behaviour arises from local component effects. Variety comes from differences in structure. The medium is passive and state-like. These expectations make the knowledge level look almost perverse, because it relaxes every one of them.

The manoeuvre is scientific rather than terminological. If knowledge-level talk supports reliable distal prediction, design and comparison while reducing: however imperfectly: to symbolic mechanisms, then the proposed level earns its place. If it merely redescribes outcomes after the fact, hides mechanisms or licenses any convenient attribution, the hypothesis fails.

This testable framing keeps the paper from becoming a metaphor about minds. Newell is asking what an observer can compute from ascribed goals and world-directed content, which failures are principled, and what correspondence the upper vocabulary has to an implemented symbol system.

The body connects the agent to an environment. Newell simplifies it as a set of possible actions, while allowing that the underlying physical system may be arbitrarily complex. The knowledge-level account treats the body by its powers: move, speak, inspect, grasp, transmit, and so on. Detailed mechanics belong below the level.

The body of knowledge resembles memory only loosely. Capacity, encoding and retrieval organisation are deliberately absent. Once a proposition is ascribed, the pure knowledge-level model treats it as available. Calling this component a memory would import precisely the structural questions the level is designed to abstract away.

A goal describes a desired state of affairs. Goals are distinguished so they can guide action, but their internal medium is again knowledge. Relations among goals: priority, compatibility, subordination: are themselves matters of content rather than wires connecting modules.

There are no interesting composition rules for constructing elaborate agents out of these components. Every agent simply has knowledge, goals and action possibilities. That absence is the first surprise. At lower levels, different structures generate behavioural variety. Here the same sparse anatomy can describe a chess player, a repair system or a traveller because the variety resides in what each knows and seeks.

The gain is distal prediction. An observer need not open the agent to infer an action when the relevant world facts, goals and powers are accessible. The cost is fragility: the description cannot explain delays, search failures, memory interference or arbitrary tie-breaking without importing symbol-level detail.

The principle connects all components globally. It does not specify a search procedure, a message-passing path or any local mechanism by which knowledge reaches action. This is the second surprise. Newell compares such a global law to extremum principles in physics, but immediately removes an easy misunderstanding: his rationality principle contains no built-in notion of optimality.

If two actions are known to reach the same goal, both are acceptable at the level. Equipotence does not rank them by speed, elegance or cost. A general goal such as minimising effort can narrow the set only when the agent also has knowledge linking effort to the alternatives.

For several goals, an auxiliary preference for joint satisfaction takes the intersection of the actions selected for each goal. This extension often yields a unique candidate. It can also yield none. Further extensions: goal preferences, expected values, minimax reasoning: cover additional cases, but Newell refuses to pretend that one universal rule determines every real action.

Physical possibility is another boundary. An action cannot occur if the body cannot perform it. Yet an unknown constraint affects selection only when it enters the agent’s knowledge. An observer may know a bridge is closed while the traveller does not. Predicting the traveller’s intended route and predicting the eventual movement then require different descriptions.

The principle is normative only in a limited functional sense. It states the regularity that makes knowledge attribution predictive. It is not a moral theory, a full account of human reasons or a guarantee of computational perfection. A system counts as approximately knowledge-level to the extent that its behaviour follows content and goals in this way.

Newell dramatises this with Frank Stockton’s story of a princess whose lover must open one of two doors. One hides a tiger, the other a woman the lover would marry. The princess knows which is which and signals a door. Her goals: save the lover and avoid the rival: pull in opposite directions. Complete information about the doors does not resolve what she wants most.

The dilemma is not ignorance on the observer’s part alone. No additional task fact need determine the choice. The princess acts, but the knowledge-level account may be silent about how conflict, impulse or symbolic processing settles the moment. A complete programme-level state could in principle determine the trajectory even when the upper model cannot.

This is the third surprise. Ordinary computer-system levels promise a definite behavioural trajectory from a full state and boundary conditions. The knowledge level is an approximation whose domain can be broad yet patchy. It predicts well when relevant knowledge, compatible goals and action affordances align; it loses purchase when they do not.

Several acceptable actions also produce underdetermination without tragedy. If two buttons both start the machine and the agent has no reason to prefer one, either is rationally licensed. An observer who predicts the exact hand movement is using unmentioned information: habit, motor bias, physical position or randomness.

The lesson is methodological. “The model cannot decide” is sometimes the correct result. Forcing a unique answer disguises the boundary between a content-level explanation and an implementation-level cause. A responsible interface should return the selected set and explain why it is plural or empty.

The definition reverses the usual order. Instead of identifying a physical item and declaring it knowledge, an observer begins with successful action prediction. If ascribing content K and goals G makes the rationality principle correctly constrain the agent, K earns the role of knowledge in that account.

This makes knowledge competence-like. The agent may be able to use a fact across an open-ended family of situations even though no finite table lists every relevant means–ends relation. A chess position supports unboundedly many questions: whether a piece can be captured now, after a pin, following an exchange, or before an attack develops. What the agent can derive cannot be read off as a static inventory from afar.

Finite machines do not literally contain an infinite extensional table. They contain generative symbol systems. Procedures operate on finite structures to produce relevant consequences as tasks demand. At the knowledge level, the observer abstracts over those procedures and treats their potential contribution as a body of content.

This is the fourth surprise. Lower-level media: voltages, bits, bit vectors, symbolic expressions: can be located as passive states. Knowledge has no matching lump of matter. It is realised only by an active combination of structures and processes, and only approximately. The physical carrier is a complete symbol system, not an isolated sentence.

The observer is an agent too. To predict another, the observer ascribes goals and knowledge, attends to the shared environment and may simulate taking the other’s role. Newell uses this to explain why the apparently abstruse level arises naturally in everyday social prediction: a general-purpose symbolic mechanism can compute what the other ought to do without reconstructing the other’s internal code.

Logic is an obvious representation because its structures, inference rules and semantics can be stated explicitly. If a set of formulas is used, the attributed knowledge includes consequences that the symbol system can in principle bring to bear. Yet a person or programme cannot enumerate every implication. The pure knowledge-level attribution is therefore an approximation, corrected by a symbol-level account of search and time.

Other representations include models, diagrams, remembered scenes, tables, scientific notation and language. None is privileged merely because its marks look less formal. What matters is the complete system: which content an observer can extract, through which processes, for which actions.

Access is not a neutral pipe. An indexed table may answer one query instantly and make a relational query expensive. A graph may reverse that profile. Two structures can realise the same knowledge extension while differing radically in the time and space required to deliver particular consequences. Those differences are symbol-level facts with behavioural effects.

The reduction is consequently a many-part mapping. The knowledge-level agent corresponds to the total symbol system; actions correspond to processes with input and output transducers; knowledge and goals correspond to data structures plus extraction processes; the rationality principle corresponds to the total problem-solving activity.

The seamless surface above becomes a pastiche below. There need be no single knowledge store, uniform encoding or central rationality module. Specialised indexes, search routines, caches, productions and sensors may cooperate. The level boundary lets designers specify what a component should make possible before deciding how to implement it.

The resolution-theorem-proving enthusiasm of the 1960s fused two claims: first-order logic could describe tasks, and uniform proof search could efficiently control intelligent action. When search disappointed, some researchers reacted against logic itself. Newell diagnoses a level confusion. A language may be excellent for knowledge-level analysis while a poor direct encoding for a time-critical programme.

A logical representation externalises both sentences and inference rules. That makes it scientifically tractable. Natural language, pictures and remembered scenes can also support knowledge attribution, but if their extraction procedures remain unknown they are research problems rather than ready engineering tools.

Logical closure creates a familiar objection: finite agents cannot consciously possess every remote consequence of their axioms. Newell accepts the force and locates it. At the pure knowledge level, closure describes competence; the symbol level explains which consequences search can produce under resources. The upper approximation is good for ordinary arithmetic competence and poor for a novel theorem under a five-second deadline.

Contradiction and inefficiency still matter. They motivate new logics, non-monotonic systems and representations embedded in larger control architectures. But every candidate has processing costs and failure modes. Logic is one member of the representational crowd for implementation, while remaining central to precise analysis of content.

This distinction also sharpens evaluation. Adding domain facts or a causal model is chiefly a knowledge-level contribution. Inventing a faster indexing, search or control method is chiefly a symbol-level contribution. Real systems contain both; identifying the scientific novelty prevents a successful demonstration from receiving the wrong explanation.

Everyday psychology already does this. We predict another person from goals and knowledge, then add rough qualifications: they may forget, have no time to think, or perform worse under stress. These additions are untidy because they borrow from the symbol level without a complete process model. They are nevertheless better than pretending perfect access.

AI designers also describe internal components as knowledge sources, messages as transferring knowledge and memories as holding content. Hearsay II’s blackboard architecture made this style especially vivid: specialised sources contribute phonetic, syntactic and other constraints to a shared working structure. The design approximates a community whose knowledge is brought to bear cooperatively.

The advantage is specification before construction. A module can be assigned the competence “supply plausible word boundaries” before its data structures and procedures are fixed. Analysis can likewise summarise a complex subsystem by how it responds to its internal environment.

The danger is the homunculus. Naming a component “the planner” or “the understander” does not explain how planning or understanding occurs. An implemented system must progressively replace every knowledge-level component description with symbol-level structures and processes. Limited goals and a controlled local environment can make a simple mechanism a good approximation to a rational sub-agent.

Mixed explanations should therefore carry level labels. “The module knows K” is a behavioural specification. “The index retrieves record r in O(1)” is a mechanism claim. “The module failed because its queue expired” crosses from the first to the second. Explicit labelling prevents metaphor from laundering an unfinished design.

MYCIN is his example of a knowledge-level contribution. The central achievement was capturing medical expertise and the relations between findings, diagnoses and therapies. Backward-chaining control mattered, but it was comparatively familiar. The scientific gain lay in articulating what competent action required.

Hans Berliner’s SNAC procedure for backgammon sits nearer the symbol-level pole. It addressed processing pathologies created by discontinuities in a global evaluation assembled from local functions. The advance improved how existing evaluative content guided search rather than adding a new body of backgammon knowledge.

Conceptual dependency offers a subtler case. Newell interprets its main contribution as a causal model of a fragment of language meaning: states, objects, actions and mental acts with enabling and causal relations. The implementation used much ordinary AI machinery. Critics who looked only for a novel parser missed the knowledge-level claim.

The distinction resembles McCarthy and Hayes’s epistemological versus heuristic adequacy but is not identical. For Newell, logic is not the only adequate content representation, and the knowledge level is a system hypothesis, not simply a checklist for formal languages.

Good modern evaluation should report both axes. A system that knows more but searches no better differs from one that uses the same content faster. Benchmark gains can otherwise be misattributed to architecture when they came from added data, or to knowledge when they came from retrieval engineering.

Newell finds a close neighbour in Daniel Dennett’s intentional stance: predict a system by treating it as rational and attributing beliefs and desires. The important difference is ontological emphasis. Dennett’s “stance” belongs to an observer’s strategy; Newell insists that system levels are real regularities, although knowledge attribution still occurs through an observer.

The paper explicitly says the level does not explain aboutness. How symbols designate external situations must be addressed partly at the symbol level and partly through lower mechanisms of sensing and action. The upper account begins after those relations are available.

Later critics have pressed this opening. Luis Pineda argues, in a much later paper, that machine symbols receive their knowledge-level interpretations from people and uses the gap to challenge claims about machine understanding and consciousness. That is a later philosophical proposal, not Newell’s conclusion. It helps expose what the original functional account settles: and what it leaves unsettled.

Later cognitive-architecture work has also made Newell’s access problem concrete. Antonio Lieto argues that general architectures often contain small, task-specific and overly homogeneous knowledge bases. He proposes heterogeneous external knowledge resources as one response. Again, this is later context: it tests the practicality of rich content and retrieval rather than rewriting the 1982 hypothesis.

The strongest defensible reading is neither deflationary nor mystical. Knowledge-level description is a powerful competence model for rational action. Its reality consists in stable cross-implementation regularities and successful reduction, not in a ghostly substance. Claims about consciousness require additional premises and evidence.

Begin with a knowledge-level contract. List goals, available actions and the means–ends relations that competent behaviour requires. Return a set of warranted actions, not a fictional unique answer when the content leaves alternatives open.

Design several possible representations. Treat a database, vector index, rule engine, causal graph or learned network as a coupled structure-and-process system. Test not only whether the content is encoded but which queries can retrieve it within the operating budget.

Measure content and access separately. Use controlled swaps: fixed content with different mechanisms, and fixed mechanisms with different content. Attribute gains to the correct level and report interactions. This makes architecture claims reproducible.

Allow mixed models openly. Real agents forget, time out and face partial observability. Mark these as lower-level constraints rather than silently redefining what is known. The distinction supports debugging: a missing fact, an inaccessible fact and an ignored fact demand different repairs.

Eliminate homunculi by progressive reduction. Each apparently intelligent module should acquire a bounded environment, narrow goals, explicit structures and executable processes. A label such as “reasoner” is a useful promise only while accompanied by a plan to cash it out.

Finally, resist category inflation. Success at rational prediction does not alone establish grounding, understanding or consciousness. The knowledge level is valuable precisely because it says something strong but bounded: content can organise the explanation of action across different implementations.

Review one: primary fidelity and historical reconstruction. The companion was checked against both complete Carnegie Mellon archival scans, including the title history, level hierarchy, four common level properties, agent anatomy, rationality and its auxiliary principles, the lady-or-tiger limit, functional definition of knowledge, observer–agent simulation, representation solutions, reduction table, access costs, mixed systems, Hearsay II, MYCIN, SNAC, conceptual dependency, predicate-calculus discussion, philosophy, generative rational systems and the conclusion.

The most serious fidelity risks were removed. Rationality does not mean optimal choice. “Selected” denotes candidacy. The joint-goal rule is an auxiliary extension, not the entire theory. Radical incompleteness can persist under complete world knowledge. Logic is privileged for analysis, not mandated as the implementation. Knowledge is realised by structures plus processes, not a static sentence.

Review two: pedagogy, fairness, AI telltales and delivery. Later criticisms are dated and labelled rather than inserted into Newell’s mouth. The treatment avoids claiming that the paper proves consciousness or semantic grounding. Repetitive summary language was reduced, metaphors were tied to formal distinctions, and every worked example states which level supplies the explanation.

All ten bespoke SVG figures have titles, descriptions, captions and source notes; motion is decorative, redundant and disabled by reduced-motion preference. The laboratory covers four scenarios and three encodings. Its exact checks independently recompute expected selected sets, verify table/graph content invariance and ensure a deadline-limited failure is labelled as access rather than knowledge deletion.

Non-monotonic logic sits primarily at the knowledge level when it specifies which conclusions are warranted by a body of information. Its tableaux, truth-maintenance machinery and search belong at the symbol level. Newell’s split therefore clarifies the previous paper without replacing its formal results.

The sequence also exposes a useful tension. McDermott and Doyle make belief states sensitive to assumptions and revision; Newell’s pure level treats knowledge as stably available and pushes access failures below. Practical agents need both: a semantics of changing commitments and a mechanism that brings relevant content to bear.

Canon 0097 moves to Richardson and Domingos’s Markov logic networks. That work will combine first-order formulas with weights, providing a concrete way to soften logical constraints under uncertainty. Newell’s question will remain live: which part of such a system is knowledge-level content, and which part is the machinery that retrieves and computes with it?