The Reports

There is a tidy story people like to tell about the first time the world lost faith in thinking machines. In it, a pair of devastating reports landed, the money dried up overnight, and an entire field went dark for a decade. It is a good story, and like most good stories about science it is mostly wrong. There were two reports, not one, and they were separated by an ocean and by seven years. They asked different questions, they were written by different kinds of men, and neither of them killed the thing it is now blamed for killing. What they did do was harder to dramatize and far more interesting: they applied a cold, outsider's arithmetic to a field intoxicated by its own promises, and they were largely right to.

The first of the two came out of an institution that had, only a few years earlier, helped invent the very idea that machines might handle human language at all. In October 1963, the director of the National Science Foundation wrote to the National Academy of Sciences on behalf of three agencies — the Defense Department, the CIA, and the NSF — which together had been pouring money into automatic translation of Russian for roughly a decade. They wanted to know whether the investment was working. The committee assembled to answer them was chaired by John R. Pierce of Bell Labs, and his presence is the thread that ties this whole episode back to the beginning of our story. Pierce was the man who had coined the word "transistor," who had championed the first communications satellites, and who had spent years across the bench from Claude Shannon, sparring over the limits of what information theory could and could not do. He was, in other words, exactly the sort of hard-nosed engineer who measured a field by its near-term payoff and found most glamorous research wanting. He would later turn the same withering eye on speech recognition. When his committee — which included two former machine-translation researchers who had grown disillusioned with the field — sat down to judge the translation effort, it was, in effect, information theory's pragmatism passing sentence on one of its own offspring.

The verdict, delivered in November 1966 in a report titled Language and Machines, was blunt. There had been, it found, no machine translation of general scientific text, and none was in immediate prospect. The machines were slower, less accurate, and more expensive than the human translators they were meant to replace; by the committee's reckoning, once a document was destined for more than a couple dozen readers, a human did the job more cheaply. The headline figure — some twenty million dollars spent over the previous decade — became the number everyone repeated, though the real sum spent on translation proper was probably closer to twelve or thirteen million once you stripped out the unrelated research it had quietly absorbed. The committee did not recommend abandoning language and machines altogether. It urged a pivot toward the basic science of computational linguistics, a term the report itself helped enshrine, and toward practical tools to assist human translators rather than supplant them.

What makes the moment poignant is how thoroughly the committee was confirming a verdict already reached on first principles. The whole American effort had been launched by Warren Weaver, Shannon's collaborator, who in a 1949 memorandum had floated the seductive idea that a book in Chinese was merely a book in English written in a strange code, and that the cryptographic methods of the war might therefore crack translation too. A decade later the philosopher Yehoshua Bar-Hillel had demolished the dream, arguing that no dictionary could ever resolve a sentence like "the box was in the pen" without the kind of worldly knowledge a machine simply did not possess. Pierce's committee reached the same destination by a different road, through cost and benefit rather than logic, and the effect was the same: the substantial federal funding for translation in the United States dried up for the better part of twenty years, and an interest in the subject became something a serious researcher learned to keep quiet about. Yet the work itself never stopped. Systran was founded two years later and deployed by the Air Force by 1970; Canada's Météo system was translating public weather bulletins by the late seventies; projects continued in France, Germany, the Soviet Union, and at American universities well past 1966. The report contracted a field. It did not bury it.

The second report belongs to a different country and a different man entirely. In 1971 the chairman of Britain's Science Research Council asked Sir James Lighthill to survey the whole of artificial intelligence — not merely translation, but the entire ambition. Lighthill was a spectacular choice and a strange one: a Cambridge applied mathematician, the founder of the science of aeroacoustics, the man who held Newton's old Lucasian chair before Hawking did, and someone who freely admitted he had no previous acquaintance with the field he was being asked to judge. He had a record of writing sharp minority reports that ended things he disapproved of, and his sponsor, it later emerged from his papers, had quietly nudged him toward skepticism from the start, fretting about a discipline that had become the preserve of experts and "perhaps sometimes of plausible charlatans." Lighthill obliged, and because he wrote alone, with no committee to soften him, the report carries the unmistakable voice of a single confident mind.

His method was to divide the field into three. Category A, advanced automation, he approved of, because it served industry and defense. Category C, the use of computers to model the brain and nervous system, he approved of too, because it served biology and medicine. His scorn fell entirely on what lay between them — Category B, the attempt to build general-purpose robots, the supposed bridge that was meant to justify the whole enterprise as a single coherent science. Here, he wrote, the disappointment was most acute, and he traced it to a single mathematician's diagnosis: the failure to reckon with the combinatorial explosion, the way the number of possibilities a thinking machine must consider balloons beyond all hope as a problem grows. His most quotable line cited the Americans directly — the most notorious disappointments, he noted, had come in machine translation, where enormous sums had bought very little, just as the National Academy of Sciences had concluded in 1966. With that sentence the two reports became one chain. Lighthill reached across the Atlantic and seven years of history to make ALPAC his own evidence.

The drama played out in public. In May 1973 the BBC staged a televised debate at the Royal Institution on the motion that the general-purpose robot was a mirage. Lighthill argued from a raised lectern that every existing robot lived in an "extremely restricted world, a sort of playpen"; below him, the Edinburgh roboticist Donald Michie ran film of his Freddy II machine sorting jumbled parts and assembling a model car, while John McCarthy insisted that artificial intelligence was a basic science like physics, and reminded the room that the very first work in the field, Turing's own, had already wrestled with the explosion of possibilities Lighthill thought fatal. No winner was declared. The footage survived, and it remains the most vivid record we have of a discipline defending its right to exist.

The consequences, again, were real but partial. The Council split its funding to favor Lighthill's approved categories and left the bridge exposed. Michie's robotics work was attacked, and his department was reorganized in 1974 under someone else, with Michie himself shunted into a small unit of his own. But the rest of Edinburgh did not collapse. The Prolog work, the automated-reasoning research, the lineage that would eventually power Britain's Alvey programme in the 1980s — all of it survived and in time flourished, and Lighthill himself would later claim, with some self-interest, that cutting the field loose from the robot-builders had done it good.

It is tempting to call all of this the first AI winter, and almost everyone does. But the phrase did not exist yet. It was coined in 1984, at a meeting of American AI researchers, by Roger Schank and Marvin Minsky, who borrowed the imagery of nuclear winter to warn of a bust they saw coming — pessimism in the community, then in the press, then in the budgets, then the end of serious work. They were forecasting the late-eighties collapse of the expert-systems boom, not describing 1966 or 1973. To call the two reports a winter is to read a later anxiety backward onto two men who thought they were simply telling the truth about what their fields had failed to do. And in a sense they were. Pierce checked the statistical lineage that traced back to Weaver and his noisy channel; Lighthill checked the symbolic tradition that ran through Edinburgh. Both ideas would lie dormant for a generation. Both would return, transformed, when machines finally learned to handle language not by following rules but by counting — vindicating, in the end, the very skepticism that had once shut them down.

=> The Statisticians