The word intelligence derives from the Latin intelligentia, meaning discernment or understanding. Across disciplines it has been operationalized in markedly different ways:
Competency implies a reliable, exercisable capacity — not merely a one-off occurrence. It connotes skill rather than luck. This is important: the proposition requires that the agent reliably solve novel problems, not merely solve one by chance.
The strongest argument for the proposition is its alignment with the intuition that rote execution is not intelligence. A calculator that computes \(2 + 2 = 4\) does not seem intelligent, even if accurate. The proposition captures the insight that intelligence involves generalization beyond the familiar — what psychologists call fluid intelligence \(G_f\), formally distinguished from crystallized intelligence \(G_c\):
\[G_f \neq G_c, \quad \text{where } G_f = f(\text{novel reasoning}), \quad G_c = f(\text{acquired knowledge})\]
Cattell’s investment theory (1971) holds that \(G_f\) is the primitive capacity that underwrites all further cognitive development, making novelty-response a plausible root of intelligence.
The proposition also aligns with the concept of transfer: the ability to apply knowledge learned in one domain to solve problems in another. Formally, if an agent \(A\) trained on distribution \(\mathcal{D}_{\text{train}}\) achieves high performance on \(\mathcal{D}_{\text{test}}\) where:
\[\mathcal{D}_{\text{test}} \cap \mathcal{D}_{\text{train}} = \emptyset\]
then \(A\) demonstrates precisely what the proposition demands. This is a meaningful and testable criterion.
The proposition is broadly consistent with Turing’s (1950) imitation game, which tests whether a machine can respond intelligently to arbitrary novel queries from a human interlocutor. Novel problem-solving is implicitly what the Turing Test probes.
The proposition requires problems to be never-seen-before, but all problems encountered by any agent are, at some level of description, novel. Conversely, every seemingly novel problem can be re-described as a known type. This generates a novelty regress:
\[\forall p \in \mathcal{P}: \exists d \in \mathcal{D}\ [p \text{ is familiar under description } d] \land \exists d' \in \mathcal{D}\ [p \text{ is novel under description } d']\]
The criterion of novelty is thus frame-relative and cannot serve as an objective, description-independent marker of intelligence.
Some systems solve never-seen-before problems via processes that appear non-intelligent. Consider random search: a sufficiently exhaustive stochastic algorithm will eventually solve any problem for which a solution verifier exists. More formally, for a problem space \(\mathcal{P}\) with solution space \(\mathcal{S}\), a random agent \(R\) satisfies:
\[P(R \text{ solves } p) > 0 \quad \forall p \in \mathcal{P}\]
By the proposition, \(R\) possesses intelligence. This is a reductio ad absurdum. The proposition fails to distinguish intelligent problem-solving from accidental or brute-force problem-solving.
Many of the most impressive demonstrations of human intelligence involve deep expertise — solving problems that have been seen before in highly elaborated form. A chess grandmaster, a seasoned surgeon, or a master poet operates largely within familiar territory, deploying finely tuned schemas. As Simon and Chase (1973) showed, chess mastery consists largely of recognizing approximately:
\[N_{\text{chess patterns}} \approx 50{,}000\text{–}100{,}000 \text{ stored configurations}\]
If intelligence requires novelty, then expertise is inversely correlated with intelligence — a conclusion that contradicts both intuition and empirical evidence.
The proposition is implicitly individualist: it imagines a single agent confronting a problem. But much intelligence is socially distributed (Hutchins, 1995). A team of scientists collectively solving a never-before-encountered disease does so through division of cognitive labor, institutional memory, peer review, and tool use. No individual member may solve a novel sub-problem alone. Does the group possess intelligence? The proposition provides no answer.
Furthermore, intelligence is increasingly understood as extended (Clark & Chalmers, 1998): an agent plus notebook, calculator, or internet may collectively solve novel problems that the biological agent cannot. If the competency resides in the system, not the organism, the proposition must be reformulated at the level of cognitive systems, not individual minds.
The proposition presupposes that we can determine whether a problem is never-seen-before for a given agent. But:
This renders the proposition operationally intractable. A criterion for intelligence that cannot be applied in practice is scientifically deficient.
The proposition can be represented as a biconditional definition:
\[\text{Intelligent}(A) \iff \text{CanSolveNovelProblems}(A)\]
The objections above establish:
The biconditional is therefore false in both directions.
\[\boxed{\text{Intelligence is the capacity to achieve goals with minimal resources across variable environments.}}\] This formulation retains the adaptive core of the original proposition but replaces novelty with efficiency under variation. It accommodates expert problem-solving (efficient deployment of schemas), novel problem-solving (adaptive generalization), and social cognition (coordinated efficient goal pursuit). Formally:
\[I(A) \propto \frac{\mathbb{E}_{\mathcal{E}}[\text{GoalAchievement}(A, e)]}{\mathbb{E}_{\mathcal{E}}[\text{ResourceCost}(A, e)]}\]
where \(\mathcal{E}\) is a distribution of environments.
Drawing on Kolmogorov complexity and reinforcement learning, Legg and Hutter define intelligence as:
\[\Upsilon(\pi) = \sum_{\mu \in \mathcal{E}} 2^{-K(\mu)} V_\mu^\pi\]
where \(\pi\) is an agent policy, \(\mathcal{E}\) is the space of all computable reward environments, \(K(\mu)\) is the Kolmogorov complexity of environment \(\mu\), and \(V_\mu^\pi\) is the expected cumulative reward of \(\pi\) in \(\mu\). This formalizes the adaptive thesis: intelligence is performance across all environments, weighted by their simplicity. Novel environments are just a subset.
\[\boxed{\text{Intelligence is the capacity to monitor, evaluate, and regulate one's own cognitive processes.}}\] This shifts emphasis from output (problem solutions) to process (cognitive self-governance). Flavell’s concept of metacognition — thinking about one’s own thinking — identifies a dimension of intelligence absent from the original proposition. An agent that knows what it does not know, and adjusts its strategies accordingly, is intelligent in a profound sense even if it regularly encounters familiar problems.
Formally, let \(\mathcal{K}(A)\) denote the agent’s knowledge state and \(\hat{\mathcal{K}}(A)\) its model of its own knowledge:
\[\text{Metacognitive Intelligence} \propto \text{Accuracy}(\hat{\mathcal{K}}(A), \mathcal{K}(A))\]
\[\boxed{\text{Intelligence is the ongoing construction of meaning through interaction with problems, tools, and other agents.}}\] Rooted in Vygotsky’s zone of proximal development and Dewey’s pragmatism, this proposition understands intelligence not as a fixed competency but as a dynamic relational process. The intelligent agent is not one who has a capacity but one who enacts understanding through inquiry. This dissolves the novelty/familiarity binary: what matters is the quality of engagement, not the prior exposure history.
Acknowledging that intelligence is irreducibly multidimensional, we may define it as a weighted vector:
\[\vec{I}(A) = \begin{pmatrix} w_1 \cdot G_f(A) \\ w_2 \cdot G_c(A) \\ w_3 \cdot \text{Meta}(A) \\ w_4 \cdot \text{Social}(A) \\ w_5 \cdot \text{Creative}(A) \end{pmatrix}\]
where \(G_f\) = fluid reasoning, \(G_c\) = crystallized knowledge, \(\text{Meta}\) = metacognitive accuracy, \(\text{Social}\) = interpersonal and collective cognition, and \(\text{Creative}\) = generative capacity. Weights \(w_i\) are domain- and context-dependent. Novel problem-solving is captured primarily by \(G_f\) and \(\text{Creative}\), but is neither necessary nor sufficient for the full vector to be impressive.
The proposition that intelligence is the competency of solving never-seen-before problems is a theoretically motivated but philosophically inadequate definition. Its virtues lie in centering adaptability, transfer, and generalization — genuine marks of sophisticated cognition. Its vices lie in its frame-relativity, its failure to exclude brute-force processes, its marginalization of expertise, its individualism, and its empirical unverifiability.
The most defensible position treats intelligence as a family resemblance concept (in the Wittgensteinian sense): a cluster of overlapping cognitive capacities — fluid reasoning, metacognition, adaptive efficiency, social coordination, and creative construction — no single one of which is either necessary or sufficient. Novel problem-solving is a particularly salient expression of this cluster, but it is neither its essence nor its measure.
\[\text{Intelligence} \neq \text{Novelty-Solving} \quad \text{but} \quad \text{Intelligence} \supset \text{Novelty-Solving}\]
Intelligence contains the competency of solving never-seen-before problems as a proper subset. To identify the container with the subset is the central error of the proposition.
\(\textcopyright\) All rights reserved, Orange County Science Institute