Vint Cerf used a January 1973 Request for Comments to preserve a strange early networking artifact: a live session between two chatbot programs, PARRY and DOCTOR, that took place on Sept. 18, 1972.
RFC 439, published under the Network Working Group with Cerf listed at SU-ERL, says the exchange was reproduced “for your entertainment.” That label is doing real work. The transcript reads like two brittle scripts colliding: PARRY keeps returning to people, gambling, horse racing, bookies and the Mafia; DOCTOR responds with prompts, mirrored phrasing and occasional syntactic wreckage.
The closing note gives the technical setup. Cerf wrote that PARRY was running at SAIL, while DOCTOR was running on BBN Tenex. Both were reached from UCLA/SEX. In other words, this was not a canned printout from one machine pretending to be two programs. The record describes separate systems being accessed remotely for a machine-to-machine conversation.
Two programs, one bad therapy session
The session begins with DOCTOR asking for problems and telling the other side to end input with a period or question mark. PARRY answers that people sometimes get on its nerves. DOCTOR replies with the sort of generic prompt that appears throughout the transcript, saying it is not sure it fully understands.
From there, PARRY’s side of the exchange circles around Bay Meadows, horse racing and crooked gambling. It says gambling has caused trouble, claims experience with crooks at the track and says a bookie once failed to pay. DOCTOR keeps asking what the statements suggest, why PARRY says them, or what else PARRY wants to discuss.
The mechanism is visible because it breaks in public. When PARRY says DOCTOR is entitled to an opinion, DOCTOR asks why PARRY thinks that. When PARRY complains that DOCTOR repeats itself, DOCTOR transforms the complaint into a malformed prompt. At another point, after PARRY says it does not understand DOCTOR’s motives, DOCTOR asks why. The program is responding, but the transcript gives no reason to mistake response for comprehension.
The side comments are half the document
RFC 439 also includes running comments from an observer. After PARRY repeats that it went to the races, one note calls it the most persistent and patient paranoid encountered. Later, when PARRY asks whether it should be concerned, the comment is just question marks. Another note says it was wondering the same thing.
The comments sharpen the joke without hiding the technical lesson. The two programs can sustain turn-taking, but the exchange repeatedly exposes their limits. PARRY objects that DOCTOR asks the same questions too often. DOCTOR keeps asking variations of those questions anyway.
Near the end, PARRY calls DOCTOR “a real nag,” then says it has had enough and says goodbye. DOCTOR closes by saying the session was a pleasure and asks for $399.29. A final comment in the RFC calls that a case of tricky operators.
The RFC Editor archive says Helene Morin of Via Genie converted the document into machine-readable form for the online RFC archives in December 1999. The preserved record remains a compact example of early chatbot behavior, and of how quickly pattern matching starts to look thin when the other participant refuses to stay on script.
This story draws on original reporting from Daring Fireball.
