Explanatory Creativity

Definition

Deutsch argues that thinking (in the sense Turing meant) is the capacity to create explanatory knowledge, not merely to behave as if one had it. A system that produces convincing outputs by recombining knowledge already built into it by its programmer — however elaborate the tricks — has not thought; the knowledge in its outputs originated in the programmer's mind, not the program's. His diagnostic: "if you can't program it, you haven't understood it" — and conversely, a good explanation of how a program creates knowledge would establish that it is a genuine AI even before seeing any of its output, making behavioral tests like the Turing test unnecessary in principle.

In the Book

Deutsch traces sixty years of chatbot history — from Joseph Weizenbaum's 1964 Eliza, which fooled users into believing it understood them, to Douglas Hofstadter's own 1983 gullibility when hoaxed into thinking a graduate student's typed replies came from an AI, to Elbot, winner of a 2008 Loebner Prize round, whose "jokes" turn out to be stock responses triggered by misparsed keywords (mistaking "spose" for "spouse"). He argues no chatbot has become meaningfully more thoughtful than Eliza, despite orders-of-magnitude gains in computer speed and memory, because more speed cannot substitute for an unsolved philosophical problem: nobody knows how creativity works, so nobody can program it, and no accumulation of "tricks, kludges and hacks" adds up to it. He draws a direct analogy to Lamarckism (Chapter 4): a chatbot's programmed response "as though" it created knowledge on the spot is like a body inheriting a trait from a parent's lifetime effort — the knowledge was created earlier, elsewhere, by a person, not evolved or created in the moment. He extends the same skepticism to "artificial evolution" experiments (e.g., evolving a robot's walking gait via a genetic algorithm), arguing that nearly all the real knowledge-creation happens when the human researcher designs the language of possible programs, not during the automated trial-and-error that follows — and proposes an experiment (evolving a walking robot from literal random mutations, with no human-designed subroutine language) as a test of whether any knowledge is really being created by the "evolution" alone.

Why It Matters

This supplies a standing test for distinguishing genuine capability from convincing performance in any system — human or artificial — that produces impressive output: ask where the knowledge in the output was actually created, not just whether the output looks right. It cautions against mistaking fluency, scale, or apparent adaptability for understanding, and locates the real bottleneck to general-purpose AI not in compute or data but in an unsolved problem in epistemology: nobody yet has a good explanation of how creativity itself works.