Introduction
Writing in the early 1970s, Dr. P. H. Millar provided a critical evaluation of the foundational definitions of Artificial Intelligence. At a time when the field was still nascent, Millar sought to move beyond the anthropocentric benchmarks of the era—most notably the Turing Test—to establish a more rigorous, engineering-oriented framework. His work, On Defining the Intelligence of Behaviour and Machines, critiques the "black box" approach to intelligence and proposes a definition rooted in decision theory and goal-oriented efficiency.
The Limitations of the Turing Test
In the 1970s, the Turing Test was widely considered the gold standard for machine intelligence. Millar, however, identified two fundamental flaws that rendered it insufficient for scientific progress:
- The "Prima Donna" Problem: Millar argued that the Turing Test is inherently "gameable." A machine could potentially pass the test by employing deceptive tactics—such as feigning a difficult personality or acting like a "prima donna"—to steer a conversation away from its cognitive weaknesses. Such behavior mimics human quirks without demonstrating genuine intelligence.
- Lack of Constructive Guidance: For engineers, the Turing Test provides a destination but no map. It defines intelligence by its output (mimicking a human) rather than its internal processes, offering no "step-by-step approach" for building the constituent skills required for complex cognition.
Furthermore, Millar challenged the prevailing notion that intelligence should be measured by the ability to perform "intellectual tasks" like theorem proving. He noted that what humans perceive as "difficult" is often computationally trivial (such as high-dimensional mathematics), making "intellectual difficulty" a subjective, human-centric metric rather than a universal standard for intelligence.
Conceptual Framework: Selection vs. Execution
To deconstruct the nature of intelligence, Millar utilized conceptual analysis and thought experiments. One of his most compelling arguments involves the "Humanoid Controller" experiment. He posits a scenario where a humanoid robot is controlled by a human operator via electrodes. While the resulting behavior is undeniably intelligent, the intelligence does not reside in the robot (the executor) but in the operator (the decision-maker).
This distinction is central to Millar’s thesis: Intelligence is a property of the selection mechanism, not the execution mechanism. A machine that proves a complex theorem is not necessarily intelligent if it was hard-coded for that singular purpose. True intelligence is found in the "task-selector"—the system that decides which skill to deploy in a given context to achieve a specific aim.
A Formal Definition of Intelligence
Drawing on neurological analogies—specifically the work of neurosurgeon W. G. Penfield regarding "automatic" versus "conscious" actions—Millar proposed that AI should be structured as a hierarchy. In this model, successful routines are eventually pushed down into "automatic" sub-processors, allowing the higher-level intelligence to focus on novel task selection.
Millar concluded his analysis by offering a formal definition intended to serve as a functional roadmap for researchers:
"Intelligence is displayed when tasks for execution are selected from a range of alternatives in such a way that the aims of the agent are likely to be furthered with maximum economy of effort."
This definition shifts the focus from mimicry to utility. It suggests that intelligence is relative to an agent's goals and must be measured by the efficiency with which those goals are pursued.
Implications and Legacy
Millar’s insights predate several modern AI concepts. His suggestion that AI research should focus on the "spontaneous labeling of statistically-detected recurring routines" mirrors contemporary developments in hierarchical reinforcement learning and automated macro-action discovery.
However, Millar was also cognizant of the limitations of his own framework. He acknowledged the difficulty of defining a machine's "aims" without falling back into anthropocentrism. Unlike humans, machines do not possess innate biological drives, making the assignment of goals a subjective exercise by the programmer. Additionally, he noted that terms like "economy of effort" remain abstract and require specific, case-by-case interpretations to be useful in practice.
Despite these challenges, Millar’s work remains a significant historical contribution, advocating for a transition from "black box" mimicry toward a transparent, decision-theoretic understanding of machine behavior.
Summary updated