AI winter

"AI winter" names a mechanism, not a capability level: expectations rise past what the work supports, disappointment follows, funders withdraw, and the research stops for reasons that have nothing to do with whether it was working. The folklore compresses three separate documents into one story about overpromising. Each of the three says something narrower and more interesting than the compression.

ALPAC, 1966, asked whether the demand existed. The report is titled Language and Machines: Computers in Translation and Linguistics, issued by the National Research Council. Its table of contents is the tell: chapter 10 is called "Is there a Shortage of Translators or Translation?" and runs two pages; the chapter titled "Recommendations" is one page long. The committee's line of attack was not that machines could not translate but that the premise of urgent unmet demand — the thing the funding rested on — had not been established.

Lighthill's survey, dated July 1972 and published early in 1973, graded three categories and failed only the middle one. Both dates are correct and the report is cited under each, which is why the same document turns up in one account as a 1972 verdict and in another as a 1973 one; the later date is when the argument actually reached anyone. Sir James Lighthill's report for the Science Research Council divided the field into category A (advanced automation — machines built to do specific industrial, military or scientific jobs), category C (computer-based modelling of the central nervous system, checked against neurobiology and psychology), and category B, the "bridge": general-purpose robotics and the claim that a theory of intelligent machinery connects the other two. He rated progress in A and C as respectable. Progress in B he found "even slower and more discouraging," and the failure he named was structural: the "combinatorial explosion," which he defined as "the explosive growth of any combinatorial expression, representing numbers of possible ways of grouping elements of the knowledge base according to particular rules, as the base's size increases." The famous sentence is about promises rather than results: "In no part of the field have the discoveries made so far produced the major impact that was then promised." So is the one before it: "Most workers in AI research and in related fields confess to a pronounced feeling of disappointment in what has been achieved in the past twenty-five years." Lighthill did not need to dispute anyone's results. He disputed the existence of the field as a field.

The 1984 warning was published before the collapse it predicted. The panel that gave the phrase its currency ran at AAAI-84 in Austin and was written up as "The Dark Ages of AI: A Panel Discussion at AAAI-84" — Drew McDermott, M. Mitchell Waldrop, B. Chandrasekaran, John McDermott and Roger Schank — in AI Magazine volume 6, number 3, Fall 1985. The published abstract records "deep unease among AI researchers who have been around more than the last four years or so ... that perhaps expectations about AI are too high." The industry that unease was about had not yet turned. Dan Weinreb's account of the Symbolics history, drawing on a study of the company, has revenues rising steadily through 1986 and then falling, with negative earnings from 1987 to 1989. The warning was in print for roughly two years before the numbers moved.

The practical use of the term today is as a test of specificity. Someone forecasting a winter is forecasting one of three distinct things — that capability stalls, that expectations correct, or that capital leaves — and the record shows the three coming apart. Lighthill judged the work in two of his three categories respectable and still argued the field had no coherent middle to fund. The expectations correction was in print in 1985 while revenue was still rising. Anyone using the word without saying which of the three links they mean is not making a claim that can be wrong.