Heuristic 'one-purpose' programs in domains like motor design or line balancing are not genuine succ… — Hubert L. Dreyfus

Heuristic 'one-purpose' programs in domains like motor design or line balancing are not genuine successes for artificial intelligence unless they are evaluated against human experts and compared to nonheuristic algorithms on equivalently formalized problems; by contrast, rule-based simulations of stereotyped human activities such as investment banking merely show that some tasks are formalizable and reveal nothing about the intelligence originally used to discover or judiciously apply those rules, so the real challenge for AI lies instead in complex domains (Areas II and III) and especially in flexible, non‑formalizable behavior (Area IV).

By Hubert L. Dreyfus, from What Computers Can't Do

Key Arguments

He notes that 'one-purpose programs that have been developed for motor design, line balancing, integrating, and so forth' are 'relevant to work in artificial intelligence' but 'are not clearly successful programs' until they are held to two standards: '(a) like the chess and checker programs they are tested against human professionals; and (b) the problems attacked by these programs have, if possible, been formalized so that these heuristic programs can be compared with nonheuristic programs designed for the same purpose.'
He points out that 'Wherever such a comparison has been madein checkers, logic, pattern recognition, chessthe nonheuristic programs have proved either equal or superior to their heuristic counterparts,' indicating that heuristic AI programs have not outperformed straightforward nonheuristic methods where fair comparison was possible.
He insists that 'programs which simulate investment banking procedures and the like have no bearing on Cognitive Simulation or Artificial Intelligence at all,' because they only codify already‑established, stereotyped human routines.
He explains that such simulations 'merely show that certain forms of human activity are sufficiently simple and stereotyped to be formalized,' emphasizing their limited significance.
He adds that, although 'Intelligence was surely involved in formulating the rules which investors now follow in making up a portfolio of stocks,' once these rules are formalized 'the formalization of these rules only reveals them to be explicable and unambiguous, and casts no light on the intelligence involved in discovering them or in their judicious application.'
He concludes that 'The challenge for artificial intelligence does not lie in such ex post facto formalizations of specific tasks, but rather in Area II where the system is sufficiently complex to require elegant techniques in order to reach a solution, in Area III where the formal system is so complex that no decision procedure exists and one has to resort to heuristics, and in Area IV in which behavior is flexible and not strictly formalizable,' thus explicitly locating AI’s real difficulties in increasingly complex and finally non‑formalizable domains.

Source Quotes

Conclusion: The Scope and Limits of Artificial Reason 1. It is difficult to classify and evaluate the various one-purpose programs that have been developed for motor design, line balancing, integrating, and so forth. These programs are relevant to work in artificial intelligence, but they are not clearly successful programs, until (a) like the chess and checker programs they are tested against human professionals; and (b) the problems attacked by these programs have, if possible, been formalized so that these heuristic programs can be compared with nonheuristic programs designed for the same purpose.

It is difficult to classify and evaluate the various one-purpose programs that have been developed for motor design, line balancing, integrating, and so forth. These programs are relevant to work in artificial intelligence, but they are not clearly successful programs, until (a) like the chess and checker programs they are tested against human professionals; and (b) the problems attacked by these programs have, if possible, been formalized so that these heuristic programs can be compared with nonheuristic programs designed for the same purpose. (Wherever such a comparison has been madein checkers, logic, pattern recognition, chessthe nonheuristic programs have proved either equal or superior to their heuristic counterparts.) On the other hand, programs which simulate investment banking procedures and the like have no bearing on Cognitive Simulation or Artificial Intelligence at all.

These programs are relevant to work in artificial intelligence, but they are not clearly successful programs, until (a) like the chess and checker programs they are tested against human professionals; and (b) the problems attacked by these programs have, if possible, been formalized so that these heuristic programs can be compared with nonheuristic programs designed for the same purpose. (Wherever such a comparison has been madein checkers, logic, pattern recognition, chessthe nonheuristic programs have proved either equal or superior to their heuristic counterparts.) On the other hand, programs which simulate investment banking procedures and the like have no bearing on Cognitive Simulation or Artificial Intelligence at all. They merely show that certain forms of human activity are sufficiently simple and stereotyped to be formalized.

These programs are relevant to work in artificial intelligence, but they are not clearly successful programs, until (a) like the chess and checker programs they are tested against human professionals; and (b) the problems attacked by these programs have, if possible, been formalized so that these heuristic programs can be compared with nonheuristic programs designed for the same purpose. (Wherever such a comparison has been madein checkers, logic, pattern recognition, chessthe nonheuristic programs have proved either equal or superior to their heuristic counterparts.) On the other hand, programs which simulate investment banking procedures and the like have no bearing on Cognitive Simulation or Artificial Intelligence at all. They merely show that certain forms of human activity are sufficiently simple and stereotyped to be formalized. Intelligence was surely involved in formulating the rules which investors now follow in making up a portfolio of stocks, but the formalization of these rules only reveals them to be explicable and unambiguous, and casts no light on the intelligence involved in discovering them or in their judicious application.

Intelligence was surely involved in formulating the rules which investors now follow in making up a portfolio of stocks, but the formalization of these rules only reveals them to be explicable and unambiguous, and casts no light on the intelligence involved in discovering them or in their judicious application. The challenge for artificial intelligence does not lie in such ex post facto formalizations of specific tasks, but rather in Area II where the system is sufficiently complex to require elegant techniques in order to reach a solution, in Area III where the formal system is so complex that no decision procedure exists and one has to resort to heuristics, and in Area IV in which behavior is flexible and not strictly formalizable. 2.

Key Concepts

It is difficult to classify and evaluate the various one-purpose programs that have been developed for motor design, line balancing, integrating, and so forth.
These programs are relevant to work in artificial intelligence, but they are not clearly successful programs, until (a) like the chess and checker programs they are tested against human professionals; and (b) the problems attacked by these programs have, if possible, been formalized so that these heuristic programs can be compared with nonheuristic programs designed for the same purpose.
Wherever such a comparison has been madein checkers, logic, pattern recognition, chessthe nonheuristic programs have proved either equal or superior to their heuristic counterparts.
programs which simulate investment banking procedures and the like have no bearing on Cognitive Simulation or Artificial Intelligence at all. They merely show that certain forms of human activity are sufficiently simple and stereotyped to be formalized.
The challenge for artificial intelligence does not lie in such ex post facto formalizations of specific tasks, but rather in Area II where the system is sufficiently complex to require elegant techniques in order to reach a solution, in Area III where the formal system is so complex that no decision procedure exists and one has to resort to heuristics, and in Area IV in which behavior is flexible and not strictly formalizable.

Context

Opening paragraph of the 'Conclusion: The Scope and Limits of Artificial Reason', where Dreyfus assesses various applied, task‑specific programs and situates the true challenge for AI in more complex and especially non‑formalizable domains (Areas II–IV).