This 1971 paper in the journal Artificial Intelligence, by Kenneth Mark Colby, Sylvia Weber, and Franklin Dennis Hilf of Stanford’s Computer Science Department, lays out the theory behind the program that became known as PARRY. The authors write that “a case of artificial paranoia has been synthesized in the form of a computer simulation model,” and that the program reproduces the input-output behavior of a paranoid patient. Where Joseph Weizenbaum’s ELIZA imitated a therapist, Colby’s program imitated the patient: a simulated paranoid who interpreted questions with suspicion, took offense, and steered conversations toward fixed delusional themes.
Colby was a psychiatrist, and his goal was scientific rather than playful. He argued that two information-processing systems should be considered equivalent when their input-output pairs are indistinguishable to an observer, and he proposed exactly that kind of indistinguishability test as the way to evaluate the model. In practice this meant having psychiatrists read transcripts and try to tell PARRY’s responses from those of real paranoid patients, an early and concrete application of the Turing-test idea to a clinical question.
Under the hood the model tracked internal “affect” variables for fear, anger, and mistrust, and adjusted them in response to what the interviewer said, so that pressing on sensitive topics would make the simulated patient defensive or hostile. This made PARRY noticeably more robust than ELIZA’s pure pattern-matching, and it became a fixture of early AI demonstrations, including a famous 1972 conversation between PARRY and ELIZA carried over the ARPANET.
For a general reader, PARRY is a reminder that the conversational-AI lineage runs through medicine and psychology, not just computer science, and that the question of how you would even know whether a program “understands” was being argued, and tested, more than half a century ago.