Margaret Boden and AI as a Science of Mind

AA program trained on a large corpus of poetry generates a stanza no human wrote, scanning correctly, using an unexpected but apt metaphor, submitted anonymously to a competition and shortlisted before anyone involved learns how it was produced. Is it creative? The question splits immediately into several smaller ones the single word "creative" was quietly carrying: was the output novel, was it valuable, did anything resembling exploration or transformation happen in producing it, and does any of that depend on who — or what — gets credit. Margaret Boden spent much of her career insisting these questions be kept separate rather than settled by the word alone.

Why Boden, Not Just the Topics FoundationalKnowledge that endures for decades — core principles

Boden'sBoden's distinctive contribution is less any single claim than a working method: treat building a computational model as a way of doing cognitive science, not merely engineering dressed in psychological language. A program that reproduces some aspect of thought is simultaneously an artefact and a precise, falsifiable hypothesis about how that aspect of thought might be organised — if the program fails to reproduce the target behaviour, that failure is informative about the hypothesis, not just about the code. This method runs across her work on symbolic AI, connectionism, and computational creativity alike, which is part of why her edited collection on the philosophy of AI functions as a genuine bridge across computing, psychology, and philosophy rather than a survey assembled after the fact¹.

Symbolic and Connectionist Traditions, Not a Single Lineage FoundationalKnowledge that endures for decades — core principles

BodenBoden treats symbolic AI (reasoning over explicit rules and representations) and connectionism (learning distributed patterns across many simple, weighted units) as genuinely different hypotheses about cognitive organisation, each illuminating some phenomena and struggling with others, rather than one being a primitive precursor the other superseded. A symbolic system's explicit rules make its reasoning inspectable in a way a trained network's distributed weights typically aren't; a connectionist system's graceful degradation under noise or damage, and its capacity to learn from examples rather than require hand-written rules, are correspondingly harder for a purely symbolic system to reproduce without being told how. Neither tradition is simply right, and treating the history of AI as a straight line from one to the other — the same oversimplification this site's How Machines Learn series was built to correct for machine-learning architectures specifically — misses exactly what Boden's own account preserves: real, ongoing disagreement about which organisation actually explains which phenomena.

What Else Must Be True Before "Creative" Applies? FoundationalKnowledge that endures for decades — core principles

ReturnReturn to the opening case. Boden's own analysis of creativity distinguishes combinatorial creativity (new combinations of familiar ideas), exploratory creativity (finding new possibilities within an existing conceptual space), and transformational creativity (changing the space's own rules so that previously impossible ideas become reachable) — and treats only some novel, valuable outputs as evidence of the deeper two². A system that recombines existing patterns in a statistically surprising way has a real claim to combinatorial creativity. Whether it has explored or transformed a conceptual space depends on further facts about its process, not just its output — did it generate and reject a range of candidates against some evaluative criterion, or did it produce the one output its training made most likely in a single pass? The poem being good is evidence for novelty and value. It is not, by itself, evidence for exploration or transformation, and treating it as though it were collapses a four-part distinction into a single verdict.

Intention, Evaluation, and Attribution FoundationalKnowledge that endures for decades — core principles

AA further question the single word "creative" obscures: who is doing the evaluating, and does the system's own assessment of its output count for anything? A human artist typically rejects many more attempts than they publish, applying their own evaluative standard before an audience ever sees the result. Whether a generative system has anything playing that role internally — rather than an external human curator selecting the one output worth keeping from many generated — changes what's actually being credited with creativity: the system, the curator, or some combination that no single party fully owns.

Strongest Objection: This Makes Creativity Unfalsifiable FoundationalKnowledge that endures for decades — core principles

AA fair complaint about Boden's four-part scheme is that it can be used to move the goalposts indefinitely — any system's impressive output can be downgraded from "transformational" to "merely combinatorial" by someone determined not to credit it, with no decisive test available to settle the dispute either way. This is a weakness of a scheme built on process rather than output alone: process is harder to observe from outside than output is, especially for a system whose internals aren't transparent even to its own designers. The distinction still does work, but applying it responsibly requires evidence about process — training regime, generation-and-selection pipeline, what counts as the system's own evaluative step versus an external curator's — that is often simply unavailable. That absence of evidence counsels caution about confident attribution in either direction, downgrading as readily as upgrading.

Provisional Conclusion FoundationalKnowledge that endures for decades — core principles

TreatingTreating AI as a science of mind, in Boden's sense, means every working system is also a specific, checkable claim about cognitive organisation — which is why this series keeps returning to architecture rather than resting on output alone. "Is it creative" is at minimum four questions wearing one word, and a system's output can answer some of them without touching the others.

Questions for Further Thought

• What evidence, short of full transparency into a system's internals, could distinguish exploratory from merely combinatorial output?
• If a generative system's own evaluative step selected the published poem from many rejected drafts, does that change what "creative" is being attributed to?
• Is Boden's symbolic/connectionist framing still the right pair of traditions to contrast, given architectures (like the attention-based models covered on Attention and Transformers) that don't map cleanly onto either?

Related Topics

• What Is the Philosophy of Artificial Intelligence? — the construction/explanation distinction this page develops through Boden's own working method.
• Neural Networks and Alternative Models — the symbolic/connectionist contrast applied across specific architectures, not just in the abstract.

Further Reading

• Boden, M. A. (Ed.). (1990). The Philosophy of Artificial Intelligence. Oxford Readings in Philosophy. Oxford University Press.
• Boden, M. A. (2004). The Creative Mind: Myths and Mechanisms (2nd ed.). Routledge.

References