Two scripted programs talked past each other on ARPANET

The conversation

On 18 September 1972, two early conversational programs were connected over the ARPANET rather than to a human: PARRY, Kenneth Colby's simulation of paranoid speech patterns, running at Stanford's SAIL, and DOCTOR, an implementation of ELIZA, running at BBN's Tenex system, both reached from UCLA. Vint Cerf recorded the session in RFC 439, "PARRY Encounters the DOCTOR", published four months later. The transcript opens with DOCTOR's standard prompt, "Tell me your problems. Please terminate input with a period or a question mark," and PARRY answering in character, "People get on my nerves sometimes," before steering the exchange toward horse racing and suspicion, the themes Colby's script was built to produce.

What the documents show

RFC 439 documents only the mechanics of the exchange and Cerf's own bracketed asides — at one point noting, of a reply that reads as unusually apt, "must have been an accident" — rather than any claim that either program understood the other. Colby's own paper, "Ten Criticisms of PARRY", published in ACM SIGART Bulletin in October 1974, describes PARRY as "a computer simulation of paranoid processes" that had "been available for interviewing on the ARPA network," and states "thousands of interviews have been conducted with several versions of the model" — a figure Colby attributes to his own research group, not to an independent audit. The same paper defends the model's design against what Colby calls "serious, reasoned, and well-founded" objections from colleagues.

The system boundary

Both documents describe PARRY as a research instrument, not a clinical system: Colby's group built it to test whether a program's output could be distinguished from a person's, using paranoid speech as a well-documented, rule-governed pattern to imitate. Neither program in the RFC 439 exchange escalates to a person or flags the conversation as unproductive; DOCTOR keeps reflecting PARRY's statements back as questions, and PARRY keeps redirecting toward its scripted preoccupations, because neither system was built to recognize the other as anything but text to react to.

Where it fails

The ARPANET session became a well-known artifact precisely because two non-understanding systems produced a transcript that reads, in places, like pointed conversation — the effect Cerf's own annotation flags as likely coincidence. That gap between apparent coherence and actual mechanism is the risk an operator inherits whenever two automated systems are allowed to converse with no person reviewing the result.

• If two automated systems interact directly, who reviews the transcript for a false impression of mutual understanding?
• Does a coherent-looking exchange get treated as validation, when neither system was evaluated for that purpose?
• Is a scale figure like "thousands of interviews" the operator's own count or an independent measurement?

RFC 439 endures less as evidence that either program understood the other than as an early record of what happens when scripted systems are pointed at each other with no human boundary in between.