Even the first-generation computer was able, if not to give advice, then at least to signal abnormal occurrences like overflow, low voltage, etc. A second-generation computer, developed in the 1960s with much more “intelligen- ce” embodied in its software, gave a user with two clocks, one broken and the other losing four minutes per day, the advice to use the broken one; for it indicated the exact time two times a day, as opposed to two times a year for the retarder. The computers of the third-generation, using operating and virtual memory systems, oriented the user in the construction and debugging of programs via many-choice menus and progressive explanations. With the actual fourth-generation, a sophisticated planning system like PANDORA can deal with requests such as, “How can I save this file temporarily if I have no space left and can’t contact the system manager?” The answer is: mail it to yourself.

The problem is now to improve the operating, planning, and expert systems in the light of the engineering knowledge principles, in order to give the next generation computers that dreamed-of capacity of entertaining long and substantial dialogues with users, analyzing situations, and offering valuable advice, so that even Turing would not be able to distinguish between natural and artificial minds. Is that possible? The present article and its 69 references seem to point in the affirmative direction. But the problem is not simple, for the natural languages were constructed during millions of years (see Hockett [1]). Hockett cites Edward Sapir, who wrote in 1921, “the lowli- est South African Bushman speaks in the forms of a rich symbolic system that is in essence perfectly comparable to the speech of the cultivated Frenchman.” Such a level is far from being attained by the most sophisticated artificial languages of the 20th century.

The authors, after reviewing current approaches to the provision of on-line advisory interaction with users (including Coombs and Alty; Naiman’s WordStar; Stallman’s EMACS; the “help key” of Foderaro’s Franz LISP; Klemperer’s MELVYL; Cullingford’s CADHELP; Schrager’s DCL-VMS; Faletti’s PANDORA; Davis’ TEIREISIAS; and Clancey’s GUIDON), enunciate the principal problems facing an advisor for open-ended tasks. These include: establishing what a user is trying to do, helping him to plan ways to reach his goals, allowing for a two-way commentary, giving feedback on the progress or side-effects of commands, and deciding when to volunteer advice or to question the user. (All this seems to be good common sense, but the common sense is not so universally distributed as Descartes had thought.)

The man-machine dialogue, considered as a manifestation of a physical symbol system, is then analyzed in terms of speech acts, plans, and meta-level inference (interpreting user inputs, supporting planful behavior, and reasoning about control). The analysis of context-dependent interpretation of questions (“p?” = “how do I make p true?” or “how do I do p?”) and commands (“p” = “make p true]” or “do p]”) entails aspects of plan formation, plan revision, plan recognition, and plan generalization. An advice-giving program would have to be capable of a good deal of introspection. It would need access to a representation of its own theory of representation of the user’s model of itself. But this goes beyond reasoning about control, for it bears the very creation of an artificial conscience.

The proposed application of rule-based systems to the interface between user and advice-giving program asks for “metahelp,” needs continuous global representation, entails increasing system complexity, and gives rise to a lot of difficult questions (e.g., what constitutes a good mental model of a system?). Nevertheless, the approach can bring some valuable contribution through the very kind of analysis carried out and required by the codification of expertise, through the use of meta-rules allowing such program to reason about the way in which they are being used, and through the modularization of rule-like representations, lending extensions to principled systems.