An ARPANET novelty linked two chatbots, not two minds

The design

On 18 September 1972, two rule-based programs exchanged messages across the ARPANET: PARRY, Kenneth Colby's simulation of a paranoid patient, running at Stanford's SAIL laboratory, and a version of Joseph Weizenbaum's ELIZA called DOCTOR, running on a BBN Tenex machine. Vint Cerf wrote up the session as RFC 439, dated 21 January 1973, describing it as reproduced 'for your entertainment.' Both machines were reached remotely from a terminal Cerf's document identifies as UCLA/SEX. The document itself never claims the exchange was a scientific test.

What the evidence says

RFC 439's own framing undercuts the idea that this was evidence of anything beyond two scripts talking past each other. The transcript records PARRY's stock paranoid themes, horse racing and suspected Mafia surveillance, met by DOCTOR's stock Rogerian reflections, and Cerf's own margin comment on one exchange, 'talk about tricky operators,' treats a nonsensical DOCTOR reply as a joke rather than a result. The document does not describe a conference demonstration; the popular story that this exchange was staged at a computing conference is not established by the RFC itself, which records only a network session and names no venue beyond the two labs and their remote access point. A separate 1976 Stanford AI Lab memo collecting reviews of Weizenbaum's book adds independent skepticism from John McCarthy, who argues that Colby's later, more controlled blind-judge test of PARRY showed only that people are easily fooled by surface pattern-matching, not that either program understood the other.

What it asks of people

No real person's private information passed through this exchange; both interlocutors were programs. What it asks of a reader is more modest and more important: to resist reading a machine-to-machine novelty as proof of comprehension, and to notice how quickly a transcript authored for entertainment can be retold decades later as a milestone of dialogue.

Privacy and safeguards

Because DOCTOR and PARRY were both simulations, there was no personal disclosure to protect here. The absence of stakes is itself a useful contrast with later companion products, which route genuine private conversation through comparable rule-based and, later, statistical systems. Neither this record nor the reviews of Colby's separate human-subject test in the Stanford memo describe what safeguards, if any, applied to the real psychiatric patients used for comparison in that later work.

• Does the RFC establish anything beyond a network engineer's curiosity about two existing programs?
• What distinguishes a scripted novelty exchange from a genuine test of a system's reasoning?
• Where a popular history claims a conference demonstration, what does the primary record actually say happened?

RFC 439 is best read as what its author called it: an artifact of ARPANET's early culture, not a scientific finding about either program.