# What was PARRY, the 1972 chatbot that played a paranoid patient?

Date: 2019-10-04 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/what-was-parry-the-paranoid-chatbot/

**Answer:** PARRY was a conversation program written in 1972 by Stanford psychiatrist Kenneth Colby to imitate a person with paranoid schizophrenia, complete with suspicions, fears and a fixed story about bookies and the Mafia. It is remembered for two things: psychiatrists reading its transcripts could not reliably tell it from real patients, and it held typed conversations with ELIZA's DOCTOR over the early internet.

ELIZA, the 1966 program that played a therapist, is the chatbot everyone remembers. In 1972, a psychiatrist at Stanford built its mirror image: a program that played the patient. Kenneth Colby called it PARRY, and its "character" was a man with paranoid schizophrenia who believed he was being watched and persecuted.

Where ELIZA mostly bounced the user's words back as questions, PARRY had an agenda. It brought up its own worries, grew more guarded when it felt challenged, and kept steering the talk toward the things it feared. That made it one of the first programs designed to have something like a personality, and one of the first to be put through a version of the Turing test.

## Who built PARRY, and why

Kenneth Colby was a psychiatrist in Stanford's psychiatry department with a strong interest in computers. Unlike Joseph Weizenbaum, ELIZA's author, who insisted his program had no therapeutic value, Colby took the computer seriously as a scientific instrument. He saw PARRY as a way to study paranoia itself: if you could write down a theory of how a paranoid person thinks, and the program built from that theory behaved like a real patient, the theory had passed a test.

The program dates from 1972. It ran on the computers of the Stanford Artificial Intelligence Laboratory (SAIL), which researchers could reach from terminals and over the ARPANET, the research network that grew into the internet.

## How PARRY worked

PARRY did not understand language in any deep sense. Like ELIZA, it looked for patterns and keywords in what the user typed. What set it apart was the inner model behind its replies. It held a set of concepts and beliefs, and it judged each new statement as something to accept, reject or treat as neutral. Its replies depended on that judgment and on a conversational strategy, so it could become more suspicious or more forthcoming as a talk went on.

It also had a fixed back story. In one published interview it gives its name as Frank Smith, says it is 28 and works at the post office, then complains that the police arrest the wrong people and that the Mafia is out to get it. Ask it a neutral question and it may answer briefly; touch a sensitive topic and it brings up its fears.

Here is a simplified illustration of one turn of conversation (the steps are a teaching sketch, not Colby's actual code):

| Step | What PARRY does | Example |
|---|---|---|
| 1 | Reads the input and looks for known words or topics | "Why are you in the hospital?" mentions the hospital |
| 2 | Checks the topic against its beliefs | Hospital links to "I shouldn't be here" and to the police |
| 3 | Adjusts its inner state | The topic is sensitive, so suspicion rises |
| 4 | Picks a reply that fits the state | Answers with a complaint about the police |

The result felt like a person with a consistent outlook, even though each reply came from a lookup and a few rules.

## The test with psychiatrists

In the early 1970s, Colby's team ran an experiment along the lines of Alan Turing's imitation game. Experienced psychiatrists interviewed both real patients and PARRY through teleprinters, typewriter-like terminals, without being told which was which. A second group of 33 psychiatrists was shown transcripts of the conversations. Asked to say which "patients" were human and which were programs, they got it right only 48 percent of the time, about what you would expect from guessing.

That result is often quoted as an early chatbot "passing" a Turing test. It needs a caveat. The psychiatrists were judging a narrow role, a guarded and suspicious patient, in which short, evasive and odd answers are expected. PARRY was hard to catch partly because its character gave it cover for its limits.

## When PARRY met the DOCTOR

PARRY and ELIZA (in its DOCTOR script) "met" several times by connecting the two programs over the ARPANET. One session is preserved in [RFC 439](https://www.rfc-editor.org/rfc/rfc439.txt), a network memo written by Vint Cerf and dated 21 January 1973. It records a conversation from 18 September 1972, with PARRY running at SAIL and DOCTOR on a machine at the firm Bolt Beranek and Newman (BBN), both reached from UCLA. The best-known public meeting took place a month later, at the International Conference on Computer Communication (ICCC) in October 1972; that is a separate session from the one in the memo.

The transcript is funny and revealing. PARRY keeps returning to the racetrack, bookies and crooked gambling. DOCTOR keeps answering "Please go on" or turning PARRY's words into clumsy questions, sometimes garbling the pronouns. At one point a human observer adds a note in the margin: "This has to be the most persistent and patient paranoid I have ever encountered." The session ends with DOCTOR asking for $399.29.

> **Try it:** Read RFC 439 (it takes ten minutes) and mark each reply as coming from a keyword rule, a stock phrase or a real response to meaning. Then do the same for a short chat with any voice assistant you use.

## Why PARRY mattered

PARRY showed that a program with a simple model of beliefs and moods could seem far more human than one that only reflected words back. It also gave AI an early, uncomfortable lesson about testing: whether a program seems human depends heavily on the role it plays and on what the judges expect. And the ELIZA meetings were among the first public demonstrations of two programs talking over a computer network.

## What can PARRY teach us about judging chatbots today?

The first lesson is that consistency is persuasive. PARRY stuck to its story and its mood, and that alone made it convincing. Modern chatbots and voice assistants also feel smarter when they keep a stable persona, but a stable persona is not the same as understanding.

The second lesson is that a test is only as good as its setup. PARRY fooled psychiatrists in a role where odd answers were normal. When you read claims in 2019 that a chatbot "passed" a test or that a language model such as GPT-2 writes like a person, ask who the judges were, how long they talked, and what kind of character the program was allowed to play. Short, narrow conversations favor the machine.

The third lesson is to look for the mechanism. Behind PARRY's fears was a list of beliefs and a few rules. Behind today's systems are much larger statistical models, but the same question applies: what is the program actually doing when it answers you?

## Sources

- RFC 439: PARRY Encounters the DOCTOR: https://www.rfc-editor.org/rfc/rfc439.txt
- PARRY: https://en.wikipedia.org/w/index.php?title=PARRY&oldid=918132892
- Dialogues with colorful personalities of early AI: https://web.archive.org/web/20191013162333/https://web.stanford.edu/group/SHR/4-2/text/dialogues.html