The ALPAC report of 1966 is the single most consequential document in the history of automatic translation, and its impact is still the reason the field describes a "winter". Between 1954 and 1966 the United States government had spent roughly twenty million dollars on machine translation, most of it on Russian-to-English systems for scientific and military material. The committee was asked whether the money was producing anything. It concluded that it was not, and the research programme it had been asked to evaluate did not recover for about two decades.
The findings of the ALPAC report of 1966
The findings are narrower and more practical than their reputation suggests. The committee did not argue that machine translation was impossible. It argued three things that were, at the time, difficult to dispute.
First, that there was no shortage of human translators. The supply of qualified Russian translators in the United States exceeded demand, and translation was not a bottleneck in scientific work. Second, that machine output required so much post-editing that the total cost per page exceeded human translation, and was slower end to end. Third, that the research had drifted: money allocated to translation was producing linguistics, and the committee thought the linguistics was the more valuable product.
The committee's recommendation was not to stop the science. It was to stop paying for the application and to fund computational linguistics directly instead.
The 1966 committee and how it reached its conclusions
The committee was chaired by John R. Pierce of Bell Laboratories, an engineer with no particular stake in the translation programme, and included Bar-Hillel among its consultants. Its method was comparative rather than theoretical: it commissioned parallel translations, measured post-editing time, and compared costs. That is why the findings of the ALPAC report proved so hard to argue with. The objection was not philosophical but accounting.
The weakness of the method is visible in hindsight. The committee measured the systems that existed in 1966 and generalised from them. It had no way to price a technology whose cost curve would later be driven by data availability and compute rather than by rule-writing labour.
The impact on funding, and what survived it
The immediate impact was financial, and the wider impact was geographic. Federal support in the United States collapsed, university groups dispersed, and the field lost most of a research generation. What is less often noted is that the ALPAC report of 1966 applied to one country. Work continued in Canada, where the TAUM group produced the Meteo system for weather bulletins, and in Europe, where the Commission of the European Communities acquired Systran in 1976 and later funded Eurotra.
Commercial development also continued, quietly, for buyers who did not need publishable quality. Institutional users wanted gist translation of large volumes at low cost, which is precisely the use the committee had judged uninteresting.
Why the funding argument returned in the 1980s
The revival did not come from a rebuttal. It came from a change in what was cheap. By the late nineteen-eighties, machine-readable parallel text existed in quantity, and IBM's Candide group showed that translation could be estimated from it statistically rather than written by hand. The economics the committee had measured no longer held, because the expensive input had changed from linguists to data.
What the committee could not have priced
One number explains the verdict better than any argument in the report. Post-editing a page of machine output took longer than translating it from scratch, and the committee could see no reason for that to change, because the expensive input was linguists writing rules and there were only so many linguists. Every improvement meant more rules, and more rules meant more people. The cost curve was linear in labour and nobody had a way to bend it.
What actually bent it was a change of input. Once translation could be estimated from text that already existed in two languages, the expensive ingredient stopped being expert time and became data and compute, both of which fall in price on their own. That is why the report reads as sound analysis and poor prophecy at once: it measured a cost structure accurately and then assumed the structure itself was permanent.
What did the ALPAC report recommend?
That funding for machine translation as an operational goal be discontinued, and that support be redirected to computational linguistics and to tools that assist human translators rather than replace them.
Did ALPAC say machine translation was impossible?
No. That argument belongs to Bar-Hillel, who had written in 1960 that fully automatic high-quality translation was unreachable without world knowledge. The committee's case was economic: too slow, too expensive, and solving a shortage that did not exist.
How long did the funding cut last?
Roughly twenty years in the United States. Substantial federal interest returned in the late nineteen-eighties, alongside the statistical methods that made the earlier cost objection obsolete.
Was the committee wrong?
The findings of the ALPAC report describe the systems of 1966 accurately. As a forecast they fail: the committee treated the cost structure it measured as permanent, and that structure changed completely once parallel text became the expensive input rather than linguists.
See where 1966 sits in the full timeline or read the 1954 demonstration that started the spending.