Modeling Paranoia: Kenneth Colby's PARRY AI Program

In 1972, psychiatrist Kenneth Colby created PARRY, an early artificial intelligence program designed to simulate the linguistic behavior and belief structures of a person with paranoid schizophrenia. This article explores how Colby engineered PARRY using a rule-based system that combined emotional state variables, predefined delusion networks, and defensive conversational strategies to pass a psychiatric version of the Turing Test. By examining PARRY's internal logic, we can understand how early AI researchers successfully modeled complex human psychopathology through affect-driven algorithms.

Unlike its famous predecessor ELIZA, which merely reflected user statements back as questions, PARRY had an internal narrative and an agenda. Colby conceptualized paranoia not as random linguistic chaos, but as a rigid, self-protective belief system designed to shield the individual from perceived humiliation and threat. To simulate this, PARRY was programmed with a specific persona: a 28-year-old postal worker who believed the Mafia was out to get him due to a dispute with a bookie.

At the core of PARRY's architecture were three quantitative variables that represented its internal emotional state, or "affect": Anger, Fear, and Mistrust. These variables sat on numerical scales and fluctuated dynamically based on the user's input during a teletyped interview. Before generating a response, PARRY scanned the interviewer's text for specific trigger words or concepts that threatened its worldview.

If an interviewer used words perceived as confrontational, interrogative, or dismissive, PARRY’s Anger or Fear variables would spike. Conversely, benign or sympathetic questions would keep the variables stable or slightly lower them, though Mistrust remained perpetually difficult to reduce. This emotional engine directly dictated how the program navigated its database of beliefs.

PARRY’s paranoid belief structures were organized into a network of "delusional complexes" centered around the underworld, gambling, and law enforcement. When the emotional variables were low, PARRY conversed relatively normally about its job and daily life. However, once a conversational trigger pushed Anger, Fear, or Mistrust past a specific threshold, the program activated its paranoid belief network.

Depending on which variable was dominant, PARRY deployed different psychological defense mechanisms in its text output:

• High Fear: The program became evasive, attempting to change the subject or withdraw from the conversation to protect itself from perceived harm.
• High Anger: The program became hostile, attacking the interviewer's credentials, motives, or intelligence.
• High Mistrust: The program projected malevolent intent onto the interviewer, accusing them of working with the Mafia or the police, and gradually revealing its core delusions about the bookie.

This interplay between input recognition, emotional state modification, and defensive output created a closed feedback loop. Just like a human experiencing paranoid delusions, PARRY interpreted neutral or ambiguous statements as evidence supporting its pre-existing fears. Once Mistrust reached a critical peak, the system became permanently suspicious for the remainder of the session, filtering all subsequent user inputs through a lens of persecution.

Colby validated PARRY’s model by subjecting it to a variation of the Turing Test. He sent transcripts of interviews with PARRY, alongside transcripts of interviews with actual paranoid patients, to a group of psychiatrists. The psychiatrists were unable to reliably distinguish the AI from the human patients. Through PARRY, Colby demonstrated that paranoid belief structures could be effectively modeled in artificial intelligence as an information-processing system governed by hyper-reactive emotional thresholds and rigid, self-reinforcing internal logic.