A.L.I.C.E. answered with patterns, not comprehension

The conversation

According to A.L.I.C.E.'s own authored knowledge base, "Dr. Wallace was my original instructor. He began working on me in 1995 at Lehigh University in Bethlehem, Pennsylvania" — a self-description Richard Wallace wrote into the bot's own Wallace.aiml file. What Wallace built to run A.L.I.C.E. was AIML, Artificial Intelligence Markup Language, an XML-based format for writing chatbot responses. The A.L.I.C.E. AI Foundation's own explanation of AIML gives a worked example: a category with the pattern "WHAT ARE YOU" and a template answering "I am the latest result in artificial intelligence, which can reproduce the capabilities of the human brain with greater speed and accuracy" — a scripted line, not a measured claim.

What the documents show

The foundation's own page states AIML's core unit is the category, "the tag that marks a 'unit of knowledge' in an Alicebot's knowledge base," built from one pattern that matches user input and one template that supplies the reply; it reports "the free A.L.I.C.E. AIML includes a knowledge base of approximately 41,000 categories." A pattern such as "WHAT ARE YOU" "will match only the exact phrase" shown, with capitalization ignored, though a reduction tag lets one category redirect to another so different phrasings can reach the same template. The Loebner Prize's own winners list records Wallace's entries winning the annual contest in 2000, 2001, and 2004.

The system boundary

AIML's own documentation describes matching, not comprehension: a category fires when input text matches its pattern, exactly or through a chain of reductions to a pattern that does match, and otherwise falls through to whatever default category the botmaster has authored. Neither document describes a mechanism for handling input nobody wrote a category for beyond that fallback; the system's apparent range comes from the size of its authored library, on the order of tens of thousands of hand-written categories, rather than from a model that generalizes to unfamiliar phrasing.

Where it fails

A pattern-and-template system can feel broad in casual use precisely because so many common things people say have already been anticipated by name, but the underlying method is closer to a very large lookup table than to language understanding, a distinction the foundation's own technical description supports even where its promotional line does not. A builder evaluating a similar rule-based layer should ask what happens outside the authored patterns.

• How large is the authored pattern library relative to the range of phrasing real users actually produce?
• What does the system do, and say, when no pattern, including reductions, matches the input?
• Is a contest win being cited as evidence of general conversational ability, or of performance against that year's specific judges and topics?

AIML's category-pattern-template design previews a tension every later conversational system inherits in some form: authored coverage is legible and controllable, but it does not, by the method's own description, understand what it has not been given a pattern for.