A Stanford program simulated paranoia before anyone could test it

The design

Kenneth Colby, a psychiatrist at Stanford, built a program that simulated the conversational style of a person experiencing paranoid delusions. Colby's own extended account of the program, catalogued by the Internet Archive as a 1975 Pergamon Press monograph titled Artificial Paranoia: A Computer Simulation of Paranoid Processes, establishes the bibliographic facts of that record: Colby wrote it, it runs 138 pages, and it documents the program later nicknamed PARRY. The publisher's catalogue record does not itself narrate the program's inner logic. What the surviving evidence does show directly is PARRY's conversational behavior: a January 1973 internet standards document, RFC 439, reproduces a verbatim session recorded on 18 September 1972 in which PARRY, running on Stanford's SAIL computer, discussed horse racing at Bay Meadows, crooked gambling, and suspicion of Mafia surveillance.

What the evidence says

RFC 439 is a network curiosity, not a validation study; it says so itself, introducing the transcript as reproduced 'for your entertainment.' Colby's actual evaluation is described only secondhand here: a 1976 Stanford Artificial Intelligence Laboratory memo collecting three reviews of Joseph Weizenbaum's book includes John McCarthy's account of Colby's blind test, in which psychiatrists reading teletype transcripts of real patients and of PARRY judged correctly no better than chance. McCarthy adds his own qualification: PARRY, he writes, is 'incapable of the simplest causal reasoning,' so a question like how someone knew strangers were Mafia, answered only with an observation about looking Italian, should have been a giveaway. The 1971 journal paper describing PARRY's original design sits behind a publisher access wall this record could not get past; the design description here relies instead on the monograph's catalogue record and McCarthy's independent description of the later test, not the original paper itself.

What it asks of people

The 1973 evaluation asked real psychiatric patients to be interviewed alongside a simulation of one, so their transcripts could be compared. It asked the judging psychiatrists to trust clinical instinct against a program McCarthy calls reasoning-free pattern matching. It asks a modern reader to resist the shorthand that fooling a psychiatrist means a program understood the person it imitated.

Privacy and safeguards

Neither record available here addresses how the real patients' teletype transcripts were anonymized, stored, or consented to for comparison against a machine. RFC 439 raises no such question because its subject, PARRY, was never a real person. The absence of any documented consent or confidentiality framework for those patients is itself notable, though this record does not establish what protections, if any, existed at the time.

• What would a blind evaluation like Colby's need to show today to count as evidence of understanding rather than mimicry?
• How were the real patients whose transcripts were used for comparison protected, and does any surviving document say?
• Does a program's ability to pass as human in a short exchange tell a reader anything about its reasoning at all?

PARRY's fame rests on a test result described secondhand and a network transcript offered as entertainment; both are real documents, but neither supports more than the narrow claims each actually makes.