--- title: "Forecasting and Assessing the Impact of Artificial Intelligence on Society" authors: - "Oscar Firschein" - "Martin A. Fischler" - "L. Stephen Coles" - "Jay M. Tenenbaum" doi: "" url: "https://ijcai.org/proceedings/1973/15" pdf: "https://ijcai.org/proceedings/1973/0015.pdf" venue: "Proceedings of the Third International Joint Conference on Artificial Intelligence" year: 1973 pages: "105-119" track: "Applications and Implications of A. I." keywords: --- # Forecasting and Assessing the Impact of Artificial Intelligence on Society Oscar Firschein, Martin A. Fischler, L. Stephen Coles, Jay M. Tenenbaum *Proceedings of the Third International Joint Conference on Artificial Intelligence*, pages 105-119. ## Abstract At the present stage of research in artificial intelligence, machines are still remote from achieving a level of intelligence comparable in complexity to human thought. As computer applications become more sophisticated, however, and thus more influential in human affairs, it becomes increasingly important to understand both the capabilities and limitations of machine intelligence and its potential impact on society. ## Full text ## 1. Introduction In the near future, intelligent machines will re­ place or enhance human capabilities in many areas pre­ viously considered strictly within the human domain. Such machines will affect the everyday lives of people in ways unique in the history of mankind, and will change the face of society as we now know it (c.f., Weizenbaum, 1972). Although many philosophers, science fiction writers, and others have conjectured about robots and other products related to artificial intel­ ligence, as f·ar as we know there has been little effort to examine the AI field as a whole (in its relation to society) in an organized manner. At the request of the Institute of Electronic and Electrical Engineers (IEEE) to conduct a technical forecasting and assessment study in one of its various fields of interest, the San Francisco Chapter of the IEEE Systems, Man, and Cyberdecisions. To put this study into perspective, it is useful to indicate the present and predicted state of the art in computer hardware, and to indicate the AI-like devices that are currently on the market. It has been recently indicated by Foster (1972), that before 1980 a general­ purpose microcomputer complete with central processor and internal working memory (but no peripherals) will be available on a single chip in small quantities for a cost of between $1.00 and $10.00. About this time frame, magnetic bubble technology is expected to bring the cost of memory from the present $0.01 a bit to $0.001 a bit. In a recent article, Broers and Hatzakis (1972) predict netics Society selected the field of artificial intel100,000 transistors and similar devices on a silicon ligence; this paper reports on the results of this study. chip a few milimeters square. In 25 years it is preAfter an initial. framework was established, the coopdieted that a single chip computer will be available eration of other organizations, such as the ACM Special Interest Group on Artificial Intelligence (SIGART), the World Future Society, and the Institute for the Future, was obtained. capable of 20 million instructions a second with 65K of internal memory selling for about~OO! Even if this estimate were off by an order of magnitude, the social significance is enormous. ## FORECASTING AND ASSESSING THE IMPACT OF ARTIFICIAL INTELLIGENCE ON SOCIETY* Oscar Firschein Martin A. Fischler Lockheed Research Laboratory Palo Alto, California L. Stephen Coles Jay M. Tenenbaum Stanford Research Institute Menlo Park, California The present paper describes the derivation of the Delphi questionnaire which forms an important part of the study, an associated detailed examination of the AI field, the Delphi study itself, the results obtained in the study, and finally, the implications of the study. ## Forecasting and Assessing the AI Field Why bother to make an effort to predict future de­ velopments in AI? On the one hand it hardly seems worth­ while to plan further ahead than ten years. Things are changing so rapidly that any attempt to predict con­ ditions more than a decade hence are likely to be seriously wrong in at least several important respects. to make short-term forecasts overly optimistic while long-term forecasts are observed to be overly pessimis­ tic. Moreover, the unanticipated appearance of satu­ ration effects can radically alter the basis of a fore- On the other hand, we tend to make decisions prac­ tically every day that must reach beyond a decade, or occasionally even a century! For example, new buildings are normally designed for an occupancy of one hundred years . Because computer systems also have the potential for long term effects on human life and society, it is important that the implications of such systems be under­ stood. It is our hope that this study can be useful to those who are given responsibility for future-based * The research reported here was sponsored in part by a grant from the National Science Foundation (GJ-37696) to the Institute of Electrical and Electronics Engineers. 105 Implications of Artificial Intelligence Table 1 indicates some AI products which have been reported in the trade press. Although some of these products are relatively unsophisticated compared to the corresponding AI products postulated in our study, they do indicate current trends and the state of the art. ## 2. Setting Up The Study In assessing the effects of the artificial intel­ ligence field on society, we found it convenient to break the effort into two parts: - (1) A technical portion in which the AI field is examined, capabilities derived, and products postulated . - (2) A more judgmental portion in which appli­ cations and implications for society are postulated, and their effects on society are judged. It is in the latter portion of the study that the Delphi Forecasting technique is most useful. The overall methodology used in the study is shown in Fig. l. - (1) The topic and subtopic areas of AI were first derived using such sources as the ACM Com­ puting Reviews Index, the IJCAI Call-For­ Papers, and through discussions with experts in the field. ## 3 . Current AI Capabilities and Derived Products In order to postulate products that could result from advances in the artificial intelligence field, we first partitioned the field into basic subject cate­ gories, and then analyzed the capabilities currently available in each category . The following categories and the capabilities associated with each category were used in the study: !--Language understanding. The ability to "under­ stand" natural language and to respond in natural lan­ guage. Ability to translate from spoken language to a written form. Ability to translate from one natural language to another. 2--Problem solving. Ability to formulate a problem in a suitable representation, to plan for its solution, and to know when new information is needed and how to obtain it. - 3--Perception (visual). The ability to analyze a sensed scene by relating it to an internal model which represents the perceiving organism's "knowledge of the world." The result of this analysis is a structured set of relationships between entities in the scene. 4--Modeling. The ability to develop an internal representation and set of transformation rules which can be used to predict the behavior and relationship between some set of real-world objects or entities. 5-- Learning and adaptive\_systems. The ability to adapt behavior based on previous experience, and to develop general rules concerning the world based on such - (2) These topic areas were examined individually to determine what capabilities might be available if success were achieved in present- experience. day efforts. - (3) From the capabilities available, products were postulated based on these capabilities . Steps (2) and (3) were iterated several times, since conceptualizing products often leads one to a closer investigation of the availability of capabilities for that product . - (4) Examining the product list, we organized the items according to the social areas on which the products will have a major impact . - (5) The applications and implications of the AI products on .society were postulated . - (6) A Delphi study was performed to solicit expert opinion concerning the items postu­ lated. - (7) The responses to the Delphi questionnaire were analyzed . We then iterated back to step (3) as required . The results of the more technical portion of the study are given in Section 3, while the judgmental portions are given in Section 4. 6--Robots. A combination of most or all of the above abilities with the ability to move over terrain and manipulate objects. 7--Games. The ability to accept a formal set of rules for games such as Chess, Go, Kalah, Checkers, etc., and to translate these rules into a representation or structure which allows problem-solving and learning abilities to be used in reaching an adequate level of performance. The areas of artificial intelligence and their subtopics are shown in Table 2. A representati.ve col­ lection of early work in AI is given in Fe igenbaum and Feldman (1963), and in Minsky (1968) . We have focused our attention on current and recent research projects in AI and have summarized in Table 3 those efforts which we feel indicate currently obtainable levels of compe­ tence for each of the AI categories postulated. By examining these AI capabilities, we were able to postulate a set of hypothetical products that might impact on society . * The products were discussed with members of the AI community, and eventually a list of twenty-one items was compiled, as given in Table 4. * Note that the postulated products were based on what could, not what would or should be developed. The would and should aspects are dealt with in the Delphi portion of the stucty . Table 1 ## SOME COMMERCIAL PRODUCTS USING AI TECHNOLOGY | Description | Name of Product | Company or Source | Function | |-------------------------------------------------------------|-------------------|----------------------------------------------------------------------------------|--------------------------------------------------------------------------------------------------------| | Unique personal identification | Identimat 2000 | Identimation Marketing, Northvale, N.J. | Device compares hand measurements with data coded on user's ID card. | | Automatic voice verification | | Bell Telephone Labs | Will be used to identify credit card users by having user repeat previously recorded phrase. | | Medical monitoring | IMBLMS | Lockheed, Sunnyvale, Calif. | A "biobelt" which allows transmission of user's bio-activity to remote console for analysis. | | Industrial robot {first generation) | Versatran Unimate | American Machine and Foundry, New Rochelle, N.Y. Unimation, Inc. Danbury, Conn . | "Blind" assembly line operation | | Industrial robot {1-1/2 generation) | HIVIP, Mark I | Hitachi Ltd., Japan | Robot with TV vision can pick up objects from a moving conveyor belt. | | Automated warehouse | | Mobility Systems, Santa Clara, Calif . | Computer-controlled positioning of load­ ing devices to correct storage bin. | | Voice operated (a) supermarket checkout {b) baggage routing | VIP- 100 | Threshold Technology, Cinaminson, N.J. | (a) clerks dictate the prices of the i terns (b) baggage handlers dictate flight number or destination | | Voice response system | | Quantel Corp . | 16,000 word vocabulary, "complex" sen­ tence response. | Table 2 | 1. 1 | Language Understanding 1.1 Speech Understanding 1.2 Semantic Information Processing (Computational Linguistics) . 3 Question Answering 1.4 Information Retrieval 1.5 Language Translation AREAS OF ARTIFICIAL INTELLIGENCE 4 . 2 Modeling Natural Systems (Economic, Sociologi ­ cal, Ecological, Biological, etc.) 4.3 Robot World Modeling (Perceptual and Func­ tional Representations) 5 . Learning and Adaptive Systems 5.1 Cybernetics 5.2 Concept Formation | |-----------------------------------------------------------------------------------|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| | 2 . Proble~ 2.1 Inference Plausible 2.2 Interactive 2.3 Automatic 2 . 4 Heuristic | Solving (Resolution-Based The orem Proving, 6 . Robots Inference, and Inductive Inference) Problem Solving Program Writing Search 6.1 Exploration 6 . 2 6.3 Transportation/Navigation Industrial Automation (e.g., Process Control, Assembly Tasks, Executive Tasks) 6.4 Security 6.5 Other (Agriculture, Fishing, Mining, Sanita­ tion, Construction, etc.) | | 3. Percept ion 3.1 Pattern 3 . 2 Scene 4. | Recognition Analysis 6.6 Military 6. 7 Household | | 4.1 The Solving | Representation Problem for Problem­ Systems 107 7. Games 7 . 1 Particular Games (Chess, Go, Bridge, etc . ) | | Modeling | | ## l. Language Understanding Winograd [ 1971 ] : A system that understands reasonably complex declarative and imperative sentences about a limi ted (simulated robot) environment . Badre [ 1972 ]: Charniak [ 1973]: Comprehension of children ' s stories in a reading text . Thompson [1969] : Woods [ 1970]: Schank [1972]: Coles [1972] : Question-answering systems which can respond to queries posed in an impressive subset of natural English about information stored in a relational data base. Reddy-Vicens [1969]: Connected speech recognition with small vocabulary and limited number of speakers. Reddy [ 1972 ] : Chess program to which moves are provided in the form of spoken chess notation. Uses knowledge of chess to decide phonetic ambiguities in favor of the more probable move. ## 2 . Problem Solving--Automatic Programming Strips, ~. et a l. [1971 ] : Uses resolution theorem prover to se l ect subgoals for GPS-type, means- ends analysis. Used in SRI robot to specify a sequence of operations (Push, Go To, etc .) that accomplish a desired end-result. Planner, Hewitt [ 1971]: QA4, Rulifson, et al . [1972]: Conniver, Sussman, et al. [ 1972 ] : Procedures that ac ­ complish a desired goa l are retrieved by pattern matching . Planner provides a n effective way of organizing ad hoc knowledge about a problem domain. Hacker, Sussman [ 1972]: Hacker is a programming system intended to emulate the programming style (try and debug) of its author. This ongoing work has thus far written successful programs for block-stacking with a simulated robot. ## 3 . Perception MIT, Minsky-Papert [1972] : Stanford, Falk [1972] hand/eye systems : Can understand scenes consisting of white polyhedral blocks, arbitrarily stacked on a black table top. Kelly [1970]: Fisch l er [ 1973]: Can find human faces in a cluttered room environment and recognize a · limited number of those previously seen. SRI mobile robot project, ~. et al. [1973]: Current research intended to understand scenes depicting real-world office environments well enough for a mobile robot to be able to fetch specified objects . ## 4. Modeling Forrester [1969] [ 1971 ] : Urban and wor l d models which indicate effects of present population trends . McCarthy and Hayes [1969]: Hayes [ 1971]: The problem of epistemology for a robot. ## 5. Learning and Generalization Samuel [ 1959-72 ]: Parameter l earning for move - evaluation in the game of checkers . Winston [ 1970 ]: Learns the distinguishing attributes for concepts in the block world (e.g., an arch or tower) by forming structural descriptions of well-chosen examples. Generalization of Strips Plans, Fikes, et a l. [1972]: Generalizes a solution to a specific problem so that it can be used in whole or in part to subsequently solve similar problems (e.g., where the room names or objects differ). ## RESEARCH REPRESENTATIVE OF CURRENT COMPETENCE IN AI ## 6. Robot i cs (Integrated Prob l em- Sol vi ng and Perceptua l Systems ) Stanford, Fe l dman [ 1971] : MI T hand/eye projects, Winston [1973] : Disassembl e a tower of b l ocks and reas ­ semble it elsewhere. Stack b l ocks, Winston [ l 973] ·and thread bo l t, Pau l [ 1973] u sing visual servoing. SRI industrial a u tomation project: Pick up selected objects from a moving conveyor belt . Current work at each of the above projects invol ves the assembly and/or repair of simple structures ( pumps , e l ectronic circu it cards, etc . ) . SRI, Ni l sson [ 1969], robot project: Mobi l e vehicle which can navigate t h rough a simpl e world of rooms and doorways to fetch large pol yhedral objects. Currently attempting to extend these capabi l ities to enab l e the robot to do useful "office boy" tasks in a real, office environment. University of Edinburg4 Amb l er, et al . [l973] : Stationary robot with mobile su rroundings . Assemb l e simpl e toys (e . g ., cars) given a kit of parts and high level instructions about how to recognize and manipulate t he various parts . Japan : Assembly robot that can "read" b l ueprints, choose parts, and assemble them, Ejiri , et a l. [ 1972]: "wa l king machine , " Waseda University; mobi l e de l i ve r y robot, Toshiba Research Laboratory. ## 7 . Game P l aying Chess program, Greenblatt [1969]: Pl ays Cl ass B chess, master checker player , Samue l [ 1959], master Ka l ah S l ag l e [ 1971 ], trivia l GO, Ryder [1971 ] . ~----- TECHNICAL -----l I I I I J I I I I I I I I I I I ![Image](https://ijcai.org/proceedings/1973/fulltext/15/image_000000_5443f18f1687abdd5587a9389d701cc0d557714b87b348da0160c116bc6f7b6c.png) ## Table 4 ## PRODUCTS USED IN THE STUDY 1. Automated Inquiry System: An automatic information-retrieval system, using a man/machine dialog to determine user needs, which can search its data base to present the user with specific information or "facts," rather than references to other sources . 2. 2 . Automated Intelligence System: General augmenter of human intelligence, capable of automatically monitoring ongoing streams of input data, coordinating facts, and making logi~al inferences to obtain insights and alert the human as appropriate . 3. 3, Talking Typewriter: Voice typewriter, capab l e of converting spoken natural language into typewritten form in essentially rea l time, with an error rate equal to a human. 4. 4 . Automatic Language Translator: Language trans l ating device capable of high-quality translation of text in one foreign language to another . (Both technical and commercial material.) 5. 5 . Automatic Identification System: System for automatical l y determining a person's identity by recognizing his voice, fingerprints, face, etc . 6. 6 . Mobile Robot: A military or police robot that can search for, locate, and deal with a prespecified subject or class of behavior patterns. 7. Machine/Animal Symbiont : Tapping the brain of a living animal such as a bird or rat, to obtain preprocessed sensory inputs (visual, auditory, olfactory, etc .) to augment the capabilities of a mechanical device. 8. Automatic Diagnostician : System capable of interactive and/or automatic medical diagnosis based on querying the patient, examination of biological tests, etc. 9. 9 . Personal Biological Model : System periodically monitors patient ' s blood circulati.on, lung and heart func ­ tion, muscle action, etc., to provide inputs to a personal biological model so that current status can be evaluated and the effects of medication and treatment can be simulated. 10. Computer Psychiatrist: A system in which the patient's written or verbal input is sufficiently understood that the system can legitimately counsel and advise the user for commonly encountered problems. 11. · Computer-Controlled Artificial Organs: Capable of replacing natural organs of the body (such as arms, legs, eyes, kidneys, heart, lungs) while still preserving homeostatic balance . 12. Computer Arbiter: A system in which adversaries enter their mutual complaints and the interactive system uses a data base of precedents and value rules to deliver either advice or a verdict . 13. Industrial Robot: Autonomous industrial robot capable of product inspection and assembly in an automated factory, using both visual and manipulative skills. 14. Voice-Response Order-Taker: Capable of handling inquiries on order status, forming new orders, looking up data in a catalog, and indicating the result of its actions verbally. 15. Insightful Weather-Analysis System: A system which combines strict numerical analysis using measurements, with the type of insight used by a weather expert, e.g., analysis of two-dimensional weather maps, patterns, review of previous records, etc. 16. Insightful Economic Model : System which combines straightforward business statistics and input/output analysis with heuristic insights concerning the economy (unemployment, interest rates, monetary and fiscal policy, etc .) to predict the effects of economic policies. 17. Robot Chauffeur: Robot cars capable of operation on standard city streets and country highways, using visual sensors. 18. Robot Tutor: A system which can accept verbal and biosensed responses from the student and tailor the course of instruction to suit the pupil 's needs. This would not be a simple- minded CAl -programmed learning device, but on a level with a good human teacher. 19. Universal Game Player: A system capable of p l aying Chess, Checkers, Kalah, Go, Bridge , Scrabble, Monopoly, etc., at a controllable level of proficiency, from master level to novice . 20. General Factotum: A robot servant, capable of smooth verbal interchange and versatile perception and manipu­ lative activity in a household environment. 21. Creation and Valuation System: Capable of creative work in such areas as music, art (painting, sculpture, architecture), literature (essays, novels, poetry), and mathematics, and able to evaluate the work of humans. The oroducts postulated in this paper differ most strikingly from conventional computer applications in three ways. First, is the requirement for natural com­ munication with humans via spoken or written natural · language. This capability is prerequisite to virtually any application intended to provide mass personal access to computing power. Second is a problem-solving capa­ bility so that humans can specify a desired effect rather than having to supply a detailed procedure for obtaining that effect. Thirdly, many of the products require an ability to interact directly with their real-world environment through sensors that acquire in­ formation, and effectors that implement a desired change. To establish the technical feasibility of the pro·­ jected products, we compared the levels of competence that could be achieved with current AI capabilities (Table 5) with those required to achieve the stated performance goals. One might conclude from Table 5 that a reasonable start has been made for many of the products. It must be pointed out, however, that in the past it has proven very difficult to extrapolate the results of AI research. Most of the work cited in this table has involved limited domains of knowledge, small data bases, and ad hoc techniques. It is unknown at this time how many of the techniques are extendable to "real-life" data bases. In predicting future capabilities in the artificial intelligence field, we were faced with the problem "How do you know that a development in a subfield of AI rep­ resents significant progress?" Critics of the AI field such as Dreyfus (1972), and Fein (1964), have doubted the significance of much of the alleged progress in the field. In particular, Dreyfus has used the analogy of a man who, having succeeded in climbing a tree, claims that he has solved part of the problem of going to the moon. There is no simple resolution to the problem, and indeed we cannot hope to settle the question here. (It may be that it is not possible to evaluate a de­ velopment until some years have passed, and one sees the effects that the development has produced.) The following indicators of progress were useful in our evaluation: Attainment. A project succeeds in doing something that was thought to be impossible or extremely difficult. aspire to great generality in a problem domain have most Theory-building. A project blends several appar­ ently disjoint facts within a common framework that promises a productive line of research. Unusual improvement. A significant improvement is made as far as speed, cost, memory size, etc. for an existing important accomplishment. Impossibility proof. A demonstration that some­ thing can't work, or can't be performed in a certain manner closes a research path in which progress had been stymied. Significant terminology or concept formation. The development of a language or representation upon which further work in the field can be founded. There are two types of difficulty in achieving the AI products, the first concerning present limitations of the AI field, and the other concerning the problem of formalizing areas of human endeavor which do not appear to be amenable to formalization. These two prob­ lem areas are discussed below. Fundamental Problems in AI Technology--In examining current capabilities in the AI product areas, Table 5, three main problems appear to be: - (1) Acquiring, representing, structuring, and retrieving world knowledge, - (2) Attaining generality from ad hoc programs, and - (3) Problems in perception. Acquiring, Representing, Structuring, and Retrieving World Knowledge--A crucial part of many of the AI products is the ability to acquire, represent, structure, and retrieve large amounts of general knowl­ edge about the world, ad hoc knowledge about specific problem domains, and self-knowledge regarding the system's own capabilities . The fundamental problem associated with large amounts of information is deter­ mining which subset is relevant to the task at hand. This problem includes the choice of a representation that presents the relevant data at an appropriate level of detail. The system must also be able to restructure the relevant knowledge so that it can be efficiently applied to a specific problem. Knowledge is typically acquired in an incremental manner; as new facts are encountered in the performance of tasks, they are inte­ grated- - whenever possible as gen~ralization& of exist1ng information. The system must be able to determine when necessary knowledge isn't available, and how it can best be acquired. Some other outstanding issues include the retrieval of knowledge stored in procedural form, the representation of "common sense," and the utilization of plausible in­ ference . Generality from Ad Hoc Programs--A! programs which often exhibited unacceptably poor performance. Conse­ quently, many AI workers have adopted the view that intelligence is best embodied in the collective effect of many ad hoc procedures, each of which solves a particular problem in a limited context. Generality is then attained by accumulating a library of such pro­ cedures and selecting the most appropriate ones for any given situation. This approach is reasonabl e given the limited number of situations likely to arise in the restricted domains of present systems. When such systems fa il, a human programmer can use his knowledge of the task requirements, the problem domain, and the machine's capabilities to add a routine appropriate for the new situation. Thus, one approach to generality is to pro­ vide a future system with similar knowledge so that it can emulate this problem-solving function of a human programmer. The user of such a system might be able to guide the machine in constructing its own programs, ## l. Automated Inquiry System Automatic inquiry systems with inferential reasoning are currently practical for small data bases and for large well structured ones. Effective searches may require the human user to refine his query interactively. ## 2. Automated Intelligence System Systems have been built that augment intelligence by clustering data, identifying critical factors in multi­ variant analysis (factor analysis), and by searching for inferential chains that relate two arbitrary facts in a symbolic data base. However, such systems cannot acquire their own knowledge. Inputs must be manually filtered for relevance and supplied in an appropriate internal representation. ## 3 . Talking Typewriter A voice typewriter could today provide reliable phonetic transcription for about 100 to 500 arbitrary words if spoken one at a time . Transcription of continuously flowing speech could be accomplished for the same number of words if the discourse were confined to a limited domain for which the machine had adequate seman­ tic knowledge to resolve segmentation ambiguities. Present pattern classification techniques have proven unable to discriminate more than a few hundred categories . (This word limitation could be circumvented by the crude expedient of using spelled speech.) ## 4. Automatic Language Translator Low quality automatic language translation is already commercially used to obtain crude but readable ab­ stracts of foreign technical papers. High quality translation that captures all intended subtleties requires that the input utterance be first understood and then reconstituted in the second language. Such translation is thus feasible only in the limited semantic domains which have been successfully handled by existing lan­ guage understanding programs. ## 5 . Automatic Identification Systems These systems now suffer from the same pattern classification problems that limit the voice typewriter. It is difficult to build systems that discriminate more than 50 to 100 categories (e.g., faces, fingerprints, voices). An interactive system in which humans provide descriptions of important features for a mechanized decision maker could perhaps be made practical in the near future . ## 6. Mobile Robot Autonomous real-world robots are not yet feasible, although at least two laboratories are actively pursuing the following goals : (a) a mobile robot that plans and executes tasks requiring polyhedral solids to be pushed between otherwise empty rooms (future operation will include a real office environment); (b) a robot vehicle that can circumnavigate a closed roadway system. ## 7. Machine/Animal Symbiont Several research projects are currently attempting to do pattern classifications using inputs obtained from the neurons of animals and even from human EEG signals. A small number of gross patterns can be discrimi­ nated, but the results at present are far too crude for practical use. ## 8. Automatic Diagnostician Many diagnostic techniques have already been successfully automated, including chemical laboratory analyses, cell counting, x-ray and electrocardiogram interpretation, etc . Several interactive systems have been built which accept symptoms, test results, and patient history, and return diagnoses or specific requests for more data. Progress in this field is limited because a machine cannot yet observe a patient directly and because each application must be programmed from scratch. More fundamental problems exist due to the crude state of current medical knowledge. ## 9. Personal Biological Model Many critical biological functions are already automatically monitored in hospital intensive care wards. The problem of making such capabilities available to eve ryone on a continuing basis is mostly one of econom­ ics . Further work by biologists in modeling the human system is needed to adequately simulate the effects of medication . ## 10. Computer Psychiatrist Computerized psychiatry is not yet feasible . After the problem of natural language understanding is solved, attention can then focus on the equally challenging task of representing the semantics and pragmatics of human interpersonal relations. ## CAPABILITIES IN DERIVED PRODUCT AREAS ACHIEVED (ACHIEVABLE) WITH CURRENT A.I. EXPERTISE ## 11. Artificial Organs Although the problem of controlling artificial organs with feedback from the body is currently more in the domain of control systems than A.I., future implementations may involve A.I. concepts. Present experimenta­ tion with feedback controlled pacemakers and, also, the life support systems in satellites can be viewed as first examples of this promising technique. ## 12. Computer Arbiter A system has been built that simulates decisions of the supreme court with some degree of success. The com­ puter bases its decisions on precedents of law a nd biases of the justices disclosed in past decisions. If those biases were replaced by a formalized statement of the principles of justice, one might have the basis for an a u tomatic arbiter . However, such formalization has not yet been attempted. ## 13 . Industrial Robot Robot manipulators are being used in increasing numbers on automobile assemb ly lines, to do repetitive tasks like spot welding, which can be preprogrammed and which operate without feedback. visual and tactile sensors would significantly broaden the range of application . Motors Research Lab successfully demonstrated a system that could mount wheels on niques to align the wheel with the studs . The addition of simple, For example, the General a hub, using visual tech- ## 14. Voice Order-Taker A voice order-taker, like the voice typewriter, can now be built to handle about 500 items . If the catalog is more extensive, items would have to be specified by giving code numbers. Limited voice response by such a system, using combinations of prestored words and phrases, is now available. ## 15. Insightful Weather; Insightful Economic Model 16. Predicting weather or economic activity is still as much an art as a science. The main problem with automat ­ ing such activities is that the experts themselves have yet to agree on the most important decision criteria or even on the most relevant input parameters. It is perhaps more feasible to simulate a particular expert, but even here the expert probably would be unable to express formally the subjective criteria he uses to reach decisions . The machine is further handicapped in areas like weather analysis, where important inputs are only avai lable in pictorial form (e.g . , cloud cover photographs). In these domains, an interactive sys­ tem could be used, where the machine ' s decision is based on manually interpreted input data . ## 17 . Robot Chauffer Sensors and systems are now available that can augment driving skills (e . g., collision avoidance radar, anti­ skid braking computer, etc . ). Moreover, laboratory vehicles have been made to follow a white line and pro­ grams have been written to detect the edge of a road and to discriminate planar shadows from real obstacles. ## 18. Robot Tutor The best existing CAl systems allow mixed initiative interactions, wherei n the student can alter the course of instruction by asking the machine unanticipated questions about the subject area to test his own under­ standing. Researchers .are currently planning systems which wil l ask probing questions to model a student's comprehension of a subject area, and then plan a customized tutorial strategy intended to transform the existing conceptual structure into a desired one. The success of such research is very dependent on progress in natural language understanding and psychological investigations of knowledge representations and learning. ## 19 . Universa l Game Player Many game playing programs have been written in the course of studying heuristic programming and learning. Machines can play champion level checkers and dominoes; reasonable chess (B-level), Kalah, and Scrabble; and poor Go. However, each game was individually programmed using heuristics deduced from careful human intro­ spection and representations painfully derived to a llow effective utilization of those heuristics. It is still premature to contemplate a general game player that can be taught to play a new game as one would teach a human opponent, or significantly improve its internal representations, and thus performance, as a result of playing. ## 20. General Factotum General-purpose humanoid-type robots remain, for the time being, in the realm of science fiction . However, until very recently, no A.I. research has been seriously directed toward this goa l. The mobile robot effort described under Product 6 could be viewed as a beginning. ## 21. Creation and Evaluation System The evaluation of human artistic endeavors is perhaps the most subjective of all human judgements. Conse­ quently, the criteria of judgement are among the most difficult to forma lize as computer algorithms, except for a few general notions like harmony or color balance. Perhaps the computer's inability to evaluate its own creative efforts in human terms explains the primitive state of computer-generated art and poetry. (See Gip and Stiny [ 1973] for a discussion of aesthetic systems.) for example by suggesting promising heuristics , pro­ viding necessary knowledge, and evaluating performance on test cases . Such an interactive problem solver should have a motive- guessing capability with which to attempt solutions to incomplet e ly formulated problem requirements. Indeed, a large part of the solution to many problems (like most of those arising in AI) lies in properly formulating the problem. Perception Problems--A number of difficult prob­ lems arise in real-world perception (e.g., speech, vision) that for the most part have been systematically eliminated from the limited domains of present systems. These problems include sensory overload which conceals objects of interest in a torrent of irrelevant detail , the related problem of partitioning interesting "figures" from the general background, forming internal representations of complex or amorphous real-world objects, and finally matching such descriptions against a large data base of known objects. The foremost prob­ lem appears to be in partitioning the input into mean­ ingful entities in the context of the current problem. In vision, objects of interest must be isolated from the background. In speech, a continuous waveform must be divided into segments corresponding to words. The problem is well known to Gestalt psychologists who have observed that the interpretation of parts or features of an object is often determined by the interpretation of the whole. Thus machines, in proceeding from the parts to a whole can be overwhelmed by the combinatorics of examining every alternative interpretation for each part. Recently, one of the authors had some success in applying dynamic-programming-type techniques to reduce the combinatorics of a brute-force search for the best interpretation of a scene, according to a model of the desired object. It may be practical to use such tech­ niques in a general scene analysis system, to follow up hypotheses resulting .from a preliminary organization (Fischler, 1973). The Problem of Value Formalization--Several of the products postulated require that a formalization of aesthetics, mores, ethics, and other judgmental areas be available . For example, Product 12, the computer arbiter, requires not only the ability to understand the subtleties of human emotion,* but must also have available some codificat ion of the moral, ethical, and legal rules of society so as to be able to give advice, or render a verdict . The creation and valuation system, Product 21, requires a formalism which somehow expresses man's concepts of exce llence in the arts. There is little to report in the way of progress in these areas. Aside from some work in music by com­ puter, von Foerster and Beachamp (1969), some elemen­ tary attempts at computer poetry and prose , and some rough analysis of court decisions, there is little hope (or danger, depending on one's point of view) that the problem of formalizing such areas of human endeavor will be solved in the near term. Even if such formalization could be made, society would be faced with the embarrassing problem of revealing the inconsistencies inherent in any organization of of human beings. We know that there are many situations in which some people are considered "more equal" than others. For example, white collar fraud involving theft from a company is frequently dealt with less severely than blue collar theft f r om the same company. Appear­ ances and attitudes often affect judgments, as do family connecti ons or personal acquaintances . Although the adverse l y affected segment of the pbulic seems to be willing to tolerate such inequit ies, due to ignorance, lack of power, or indifference, one doubts whether the same passive behavior would occur if an attempt was made to incorporate explicitly the present defects of society into an automated arbitration system. A conflict would undoubtedly result with those having a vested interest in the status quo. * Either via natural language or by means of sensors, such as a "lie detector" sort of device. Decisions concerning the rate of technological development, time of appearance of a product, and the societal implications are judgmental in nature, and for this portion of the study we used the Delphi technique, a method for systematically soliciting and collating informed judgments on a particular topic.* Under this procedure, participants respond to a series of questionnaires interspersed with summaries of the group members to earlier questionnaires. differs from simple polling procedures in responses by The method that the feedback obtained from the respondents allows the ques­ tions to be modified (e . g ., a question may be too vague), and tends to prevent "snap-judgments" since the respon­ dent can see how his answer stands with respect to the other participants . As indicated in Section 2, we first partitioned the products into the societal areas which they most affect: human information processing, security, health, law, production and manufacturing, commerce, education, social interaction, and aesthetics . We then derived applications and impli cations for each product, where an application is a statement of how the product would be used by society, and an implication is the effect on society of these applications. Thus, for Product 13. Industrial Robot, one of the applications listed was "Robots for tedious assembly line tasks," while one of the implications was, "displacement of blue-collar workers." We used a format similar to that of De Brigard and Helmer (1971), as shown in Fig. 2. For each product, we indicated the social impact sector, the applications and implications, and the AI categories. The informati on solicited was: - (1) What revisions should be made in the product list? - (2) What revisions should be made in t he appli­ cations and implications for each product? - (3) What is the potential significance of each product? * See Lipinski et al (1972) for an interesting discussion of expert interrogation. - (4) What is the respondent's best estimate of the prototype and commerci~l date of the product?* - (5) What is the likelihood of each implication and applicat i on? - (6) What is t h e desirability of each implication and application: The questionnaire was sent to members of the International Joint Artificial Intelligence Council a nd to other experts in the AI field throughout the world. A total of sixty questionnaires was mailed ou t, and twenty- one responses were obtained. An additional mailing was sent to the San Francisco I EEE Systems , Man, and Cybernetics Society , and a total of twenty question­ naires were r eceived . The analys is of r esul ts given below is based on two rounds of responses from t h e first group; the complete resul ts will be presented in a forthcoming IEEE r eport. Where interesting differences between the two groups occurred, they are also briefly noted. ## Results of the Delphi Study The Delphi comments wer e reviewed manually, while the time estimates and likelihood and desirability of the products were tabulated by a simple computer pro­ gram. These results were presented directly on the questionnaire, as shown in Fig. 2 for review by the part ic ipants . The left side of the "house-like" figure r epresents the lower quartile, the roof peak is the median, and the right side is the uppe r quartile. General Observations on the Results--Due to lack of space, the complete 13 page Del phi Questionnaire will not be reproduced here. Instead an overal l sum­ mary of the Delphi responses is g iven in Table 6 . The results were obtained by weighing each r esponse with the self-ranking of expertise supplied by each respon­ den t for each product . t In the table, products were first grouped according to the median rating they re­ ceived for po tent i al significance (high, med ium, and l ow), and within these signi ficance groups, they were listed in order of prototype date . The overall de­ sirabili ty is a rough summary of the responses con ­ cerning desirability for each product. The following results are of interest: ## (l) Product Potential Significance Hal f of the AI products were thought to be of high potential significan ce and most of the rest fell in the medium potential signif i cance category . *Note that in asking for judgments concerning time of commercial appearance, we are asking not only whe t her a product can b e made, but also when the commercia l forces of need plus profitability wil l cause a product to appear on t h e market. t A weight of 1 thr ou gh 5 was appli ed to (1) unfamiliar, (2) casually acquainted, (3) familiar, (4) quite familiar, and (5) expert ratings. ## (2) Rate of Progress Respondents were generally very optimistic concerning technology deve l opment. As shown in Table 6, thir teen of the products have median dates for prototype product developme n t of 1985 or earlier. In the detailed responses, only one product had an upper quartile in the "never" range (P7). It is interesting to note, however, that in going from the first to the second round, there was a very slight, but noticeable, shift toward conservatism in prod­ uct dates as well as the expected observat i on of a reduction in the width of the houses in­ dicating a sharpening of concensus. ## (3) Prototype vs. Commercial Dates For the products having prototype dates of 1985 or earlier there is a fair l y systematic translation of 5 : 2 years as one moves f r om median prototype to median commerc ial date. ## ( 4 ) Likelihood and Desirability There is also a strong mood of optimism among the respondents regarding the likelihood and desirability of the various products. On l y five products had implications whose desir­ ability was deemed "very detrimental" (Pl, P2, P5, P6, and Pl8), and only one product (P7) had an overall detrimental rating. ## (5) Products Potentially Dangerous to Society Singled ou t for special attention were prod­ ucts that were simultaneously probable and detrimental for one of their applications or implications , i. e ., had a median estimate of likelihood in the range (possible, probable , highly probable) and a median desirability in the r ange (det rimental , very detrimental). These products and their associated problem area(s ) are indicated below: - Pl - Automated Inqui ry System (opportunity for censorship) - P2 - Automated Intelligence System (use by government to monitor actions of citizens) - P5 - Automatic Ide nti fication System (use by government to monitor the civilian population) - P6 - Mobile Robot (robot soldier; aggressive actions by some nations) - P7 - Mach ine/Anima l Symbiont (use for weapons systems sensors; alterat i on of human values) - PlO- Compu ter Psychiatrist (use of system by government to influence behavior ) Pl8- Robot Tutor (possible indoctrination of students by government ) P20- General Factotum (possible emotional impact on children)