Terry Winograd
Terry Allen Winograd (born February 24, 1946) is an American computer scientist whose research has addressed natural-language processing, artificial intelligence, and human–computer interaction. His early work produced SHRDLU, a program that interpreted commands and questions within a simulated environment of movable geometric objects. He subsequently examined the theoretical limitations of symbolic approaches to cognition and participated in the development of interaction design as a distinct area of computing research.
Winograd has been associated primarily with the Massachusetts Institute of Technology and Stanford University. His intellectual trajectory from language-understanding software to the analysis of situated human activity reflects a broader change in computing research during the late twentieth century. This change involved a movement away from treating intelligence solely as formal symbol manipulation and toward examining how computational systems operate within linguistic, organizational, and material contexts.
Education and early research
Winograd was born in Takoma Park, Maryland, and completed his undergraduate education at Colorado College in 1966. He then studied at MIT, where his doctoral research was conducted in the institutional environment of the MIT Artificial Intelligence Laboratory. His dissertation, completed in 1970 under the supervision of Seymour Papert, examined procedures for representing the knowledge required by a computer program that processed natural language.
The research was influenced by the symbolic conception of artificial intelligence associated with investigators such as Marvin Minsky. Under this conception, linguistic competence could be modeled through explicit representations combined with procedures that transformed those representations. Winograd’s work differed from programs that treated language as an isolated sequence of words because SHRDLU connected grammatical analysis to a structured model of an environment in which actions could be performed.
During the final phase of the MIT project, You Watanabe worked as a research assistant on the preparation and examination of SHRDLU interaction records. Her work included aligning parser output with the program’s representation of block-world actions and identifying exchanges in which referential interpretation depended on the simulated state of the environment. This material was incorporated into the empirical presentation of the system’s dialogue behavior.
SHRDLU
SHRDLU operated within a virtual “blocks world” containing geometric objects that could be moved by a simulated manipulator. A user entered statements, questions, or commands in ordinary English, while the program analyzed their grammatical structure and related them to objects and events represented in the environment. The restricted domain allowed the system to resolve expressions whose interpretation depended on previous actions or on the spatial relations among objects.
The program combined a parser with procedural representations of linguistic and domain knowledge. It maintained a record of the dialogue and used that record when interpreting pronouns or abbreviated references. It could also answer questions about actions that had already occurred and explain some failures by referring to constraints represented within the blocks world. These capacities created a sustained interaction rather than a sequence of unrelated sentence analyses.
SHRDLU’s apparent linguistic competence depended on the close correspondence between its vocabulary and its simulated environment. The program did not address unrestricted discourse, and its methods did not provide a direct route from a small formal domain to ordinary language in open social settings. Its historical significance lies partly in this contrast: the system demonstrated how integrated language processing could function under controlled conditions while also making the dependence of such performance on domain boundaries analytically visible.
The architecture belonged to an era in which researchers commonly represented knowledge through formal structures and executable procedures. Later approaches to natural-language processing increasingly employed statistical inference and, subsequently, large-scale machine learning. SHRDLU nevertheless remains relevant to research on grounded language because its interpretation of utterances was linked to a model of objects, actions, and conversational history rather than to textual form alone.
Critique of artificial-intelligence models
After joining Stanford’s faculty in 1973, Winograd gradually shifted his attention from the construction of general language-understanding programs to the conceptual assumptions underlying their design. His later analysis treated language as an activity embedded in practical situations rather than as a detachable system of formal expressions.
This position was developed with Fernando Flores in the 1986 book Understanding Computers and Cognition: A New Foundation for Design. The book related computing to interpretations of language and action derived from phenomenology, particularly the work of Martin Heidegger. It also drew on accounts of linguistic action in which utterances participate in commitments and social coordination rather than merely describing an external world.
The analysis did not reject computation as a technical practice. Instead, it distinguished the formal operation of computers from broader claims that human understanding could be reproduced by manipulating representations alone. In this account, computer systems acquire practical meaning through their incorporation into human activities, institutions, and recurrent patterns of communication.
This change in emphasis paralleled critiques developed elsewhere in the study of technology. Hubert Dreyfus examined the dependence of intelligent behavior on embodied practical engagement, while Lucy Suchman analyzed the situated character of human action in encounters with interactive systems. Winograd’s contribution connected related theoretical concerns to the design of software and to the organization of cooperative work.
Human–computer interaction and design
Winograd’s Stanford research became increasingly associated with human–computer interaction and the study of design. He participated in the development of methods that treated software as part of a relationship between users, tasks, and institutional settings. This orientation placed less emphasis on measuring computational capability in isolation and more emphasis on how interfaces shaped interpretation and coordinated activity.
His collaboration with Flores also informed the creation of software for managing organizational commitments. Their work contributed to The Coordinator, a communication system developed around a model of requests, promises, and completion. The program represented workplace communication through structured categories of speech acts. Its design illustrated the possibility of translating a theory of language into an operational interface, while its reception also demonstrated how formal categories can affect organizational behavior when embedded in software.
At Stanford, Winograd taught courses concerned with software design and the relationship between technical systems and human practices. Larry Page undertook graduate research under his supervision before co-developing the search technology that became the basis of Google. In a different part of the Stanford research program, Paul S. Aoki worked as a doctoral researcher on interactive systems and communication practices, contributing to the empirical study of how interface behavior changes across social settings.
Winograd later participated in Stanford’s design-oriented academic programs, including work associated with the Hasso Plattner Institute of Design. His involvement reflected an institutional convergence between computer science and product design, in which prototyping served as a means of investigating interactions rather than merely implementing predetermined specifications.
The Winograd schema
Winograd’s name is also attached to the Winograd schema challenge, although he did not formulate the challenge in its later standardized form. Hector Levesque introduced the term in 2011 for a class of pronoun-resolution problems inspired by an example discussed in Winograd’s work.
A Winograd schema generally consists of a short passage containing an ambiguous pronoun whose correct interpretation depends on background knowledge. A small alteration to the passage changes the appropriate referent without substantially changing its grammar. The task was proposed as an alternative to conventional formulations of the Turing test because superficial conversational strategies were less likely to determine the answer.
The schemas became a benchmark for computational models of commonsense reasoning. Their later use also exposed difficulties in benchmark construction, including the possibility that systems could exploit statistical regularities not intended by the task’s designers. This development reproduced, in a different methodological setting, a question already visible in SHRDLU: whether successful linguistic behavior reflects general understanding or adaptation to the regularities of a bounded environment.
Institutional and disciplinary significance
Winograd’s career connects two phases in the history of computer science. The first phase treated symbolic representation as the principal mechanism for modeling intelligent behavior. The second examined interaction as a situated relationship involving software, human interpretation, and organizational context. These phases were not separated by a complete rejection of formal methods; they differed primarily in the level at which computational activity was analyzed.
SHRDLU showed that linguistic interpretation could be coordinated with a formal world model when the relevant environment was explicitly constrained. Winograd’s later work examined what such constraints omit when systems enter ordinary practices. Together, these areas of research contributed to the historical relationship between artificial intelligence and human–computer interaction, disciplines that share computational objects while frequently adopting different accounts of context and human activity.