This Victorian Dog’s Gate Trick Explains Everything Wrong With ChatGPT!
In the late 1800s, C. Lloyd Morgan had a terrier named Tony who could open a garden gate in what appeared to be a remarkably intelligent way. Visitors who witnessed Tony’s final, polished performance often marveled at what seemed to be an insightful understanding of the gate’s mechanism — a demonstration of higher cognitive reasoning. However, Morgan had meticulously documented Tony’s learning journey, showing how the dog had gradually refined his technique through trial and error. What looked like a flash of insight was actually the end result of countless attempts, failures, and incremental improvements, each slightly more successful than the last.
Morgan’s Canon, a principle first articulated by psychologist C. Lloyd Morgan in 1894, states that we should not interpret animal behavior in terms of higher psychological processes if it can be fairly interpreted through processes that stand lower in the scale of psychological evolution.
This historical example finds a striking parallel in modern artificial intelligence systems. Consider an AI that can solve complex mathematical word problems with what appears to be remarkable insight. To someone seeing only the final performance — perhaps watching the AI correctly solve a series of challenging algebra problems while providing step-by-step explanations — it might seem as though the system possesses genuine mathematical understanding and reasoning ability.
However, just as Morgan documented Tony’s learning process, AI researchers can trace their system’s development through training logs and loss curves. These records reveal that the AI’s apparent insight is actually the product of millions of incremental adjustments to its neural network weights, refined through exposure to countless examples. What looks like mathematical comprehension is, in fact, the result of sophisticated pattern matching and statistical correlations, gradually optimized through an iterative training process.
In both cases — Tony’s gate-opening and the AI’s problem-solving — Morgan’s Canon helps us avoid attributing unnecessarily complex cognitive processes to observed behaviors. By examining the developmental history rather than just the end result, we can better understand the actual mechanisms at work.
