Why Review 40-Year Old Research on AI?

Any assessment of the abilities of AI needs to take into account the work of Hubert and Stuart Dreyfuss 40 years ago. Interestingly, Hubert Dreyfuss’s field was philosophy, which he taught at M.I.T. His brother, Stuart, was a computer specialist and a skilled chess player. For a time, Stuart worked at RAND, where Herbert Simon also worked. This may seem an insignificant detail, but many of the early AI programs were designed to play chess. There were all kinds of news stories about the computer beating chess champions at the game. With his chess skill, Stuart was uniquely qualified to understand what was involved in programming a computer to win at chess. But that still leaves the question of why the work they did 40 years ago is so important today.

While the world was trying to demonstrate that computers could think like humans, the Dreyfus brothers studied human expertise and found that the highest levels of human brain functioning, which include judgement, cannot be captured by computer programs as data. This put the highest level of human thinking beyond the analysis of computers and AI, even if computers could beat master chess players at their own game. I am not suggesting AI has no legitimate roll to play in data analysis. Only that judgment and expertise are not AI skills and those who think AI can think like human experts have missed a very important feature of human expertise.

Herbert Simon, also at RAND at that time, believed computers could be programmed to use intuition to solve problems. From Herbert Simon’s perspective:

“…we now have the elements of a theory of heuristic (as contrasted with algorithmic) problem solving); and we can use this theory both to understand human heuristic processes and to simulate such processes with digital computers. Intuition, insight, and learning are no longer exclusive possessions of human beings: any large high-speed computer can be programmed to exhibit them also.”
Man and Machine, p. 7

The Dreyfuss brothers found themselves on the opposite sides of the argument that computers could be programmed to think like humans. Hubert Dreyfuss, with his knowledge of the history of philosophy, understood that classic philosophy concentrated on how to make rational decisions by looking at evidence. Hubert Dreyfuss agreed that the AI group at RAND had provided solutions to the computer’s ability to solve problems, but the research did not support the claim that computers understood intuition.

With the COVID pandemic, we saw a of suppression of scientific opinion which disagreed with the information being supplied to the public about the vaccine. Interestingly, the Dreyfuss brothers were subject to similar suppression of their opinion by Herbert Simon and RAND. The Dreyfuss brothers’ paper questioning the claim computers could process intuition was suppressed for a year because it questioned the politically correct assumptions about human expertise. Stuart Dreyfuss, as a skilled chess player, knew from his experience:

“…human decisionmaking was an inscrutable business, a mysterious blending of careful analysis, intuition, and the wisdom and judgment distilled from experience.”
Man and Machine, p. 8

Hubert Dreyfuss, as a student of philosophy, recognized that the well known philosophers of history understood that perception could not be reduced to a set of rules:

“Human understanding was a skill akin to knowing how to find one’s way about in the world, rather than knowing a lot of facts and rules for relating them. Our basic understanding was thus a knowing how rather than a knowing what.”
Man and Machine, p.4

A difference covered in RuralDocAlan’s Substack article What to Think vs. How to Think.

The Dreyfuss brothers set out to document what kind of behaviors can be associated with various levels of expertise, from novices to experts. They studied the skill acquisition process of airplane pilots, chess players, car drivers, and adults learning a second language and observed five levels of skill acquisition. They studied physical activities such as those of car drivers because they believed that automatic driving of a car became something that was learned, but became second nature, something you didn’t have to think about as you were doing it. They maintained the same kind of being able to do something without thinking about it applied to bodily activities such as driving as well as the mental activity of judgment and expertise.

From their research, the Dreyfuss brothers represented their findings on expertise in a table of five levels:

1. Novice

2. Advanced Beginner

3. Competence

4. Proficiency

5. Expertise

The novice follows instructions to learn objective facts related to skill. The advanced beginner applies the rules learned to real situations. Eventually, according to their research, the Dreyfuss brothers suggested the experiences themselves become more important than the rules. A competent person with a goal sees the situation as a set of facts. The proficient performer intuitively understands the task, but still thinks analytically about the goal. The first three levels of performance are wedded to making decisions by conscious deliberation, analysis of the data, the “facts.”

In contrast, the levels of proficiency and expertise make judgments based upon “prior concrete experiences in a manner that defies explanation….there is more to intelligence than calculative rationality.” The skills of those at the proficiency and expertise level make decisions “characterized by a rapid, fluid, involved kind of behavior that bears no apparent similarity to the slow, detached reasoning of the problem solving process.” (p. 27)

To me, the whole business of “evidence-based” medicine and the current fixation on data and the denigration of physician judgment and expertise has been in the making since the 1980s. Even 40 years ago, the Dreyfuss brothers recognized the danger of confining problem-solving to analyzing data:

“Should we become savants of expert systems and demanding of our experts their rules and facts, become careless of the intuitive powers that fall outside our stunted vision, we will in one generation lose our professional expertise and confirm those expectations.”
Man and Machine, p. 206

Expertise involved being able to immediately see a gestalt of a situation years ago without having to analyze the parts of the mental image. As the Dreyfuss brothers allege, this is judgment, something lacking from computer-based analytical results of data alone.

For a more recent analysis of this problem with representing judgment and expertise, please see C. Thi Nguyan’s The Limits of Data:

“In transforming understanding into data, we typically eliminate or reduce evaluative methods that require significant experience or discretionary judgment in favor of methods that are highly repeatable and mechanical. And if policymakers insist on grounding their policy in large-scale public datasets, then they are systematically filtering out discretion, sensitivity, and contextual experience from their decisionmaking process.”

The current insistence that good medicine is contained only with evidence-based data while physician expertise based upon years of experience is denegrated is a blatant attempt to control the practice of medicine with rules. Those ensconced in the sanctity of the rules of data manipulation do not understand that their data can’t capture true expertise, whether that of a physician or a chess master.

Yes, there is good reason to look back 40 years to the work of the Dreyfuss brothers. The notion that everything a physician does must be controlled by the rules of evidence-based medicine confines the practice of medicine to the competency level of performance—and puts large entities manipulating healthcare for profit in control of medicine.

As we’ve said many times before, let physicians be physicians. It’s the only way to restore the dysfunctional U.S. healthcare system to the level which allows experienced physicians to practice medicine with the judgment and expertise their patients deserve.

---