The CIA's 1964 report on Soviet artificial intelligence research

Artificial-Intelligence Research in the USSR is a Scientific Intelligence Report produced by the CIA's Office of Scientific Intelligence, numbered OSI-SR/64-37 and dated September 8, 1964. It was originally marked For Official Use Only, meaning its distribution was restricted but it did not carry a higher classification such as Secret or Top Secret. The report was compiled from Soviet open literature, meaning published scientific journals, conference proceedings, and books available to Western analysts, rather than from clandestine sources.

The document runs to roughly twenty pages of analysis plus a detailed appendix listing the Soviet institutes and individual scientists the CIA had identified as active in the field. It was released to the public through the CIA's CREST database in March 2018.

The report's central finding was blunt. Soviet research in the theoretical and hardware aspects of artificial intelligence was, in the CIA's assessment, about as advanced as US work, and could be expected to continue at a rate equal to or greater than that observed in the West. The analysts attributed this not to any Soviet technical advantage but to the scale of government and Communist Party support the field was receiving in the USSR, support the report says was actually greater, in relative terms, than what the field enjoyed in the United States at the time.

The report also flagged something its authors seemed to find notable: the degree of open scientific debate Soviet researchers were permitted on the subject. Questions about the nature of intelligence, of consciousness, and of what separates a human mind from a machine were being argued out in Soviet scientific journals with what the report calls unusual freedom of open discussion, a phrase that stands out coming from a Cold War intelligence assessment of a closed society.

The report's most striking section describes a June 1962 conference in Moscow on The Philosophical Problems of Cybernetics, co-sponsored by three bodies under the Presidium of the Academy of Sciences, USSR. Roughly 1,000 specialists attended: mathematicians, philosophers, physicists, engineers, biologists, linguists, and economists. Six of the ten papers presented dealt directly with what the CIA report calls the problem most animatedly discussed, and most disturbing of all: whether a machine can think.

Among the participants was A.N. Kolmogorov, a mathematician of major international standing. According to the CIA's summary of his position, Kolmogorov argued that a sufficiently complete model of a living being could, in all justice, be called a living being, and a model of a thinking being, a thinking being. He is described as calling for the definitions of life and thought to be freed from what he considered arbitrary premises and redefined along purely functional lines.

He was not alone. V.M. Glushkov, described in the report as one of the most politically powerful figures among Soviet cyberneticists, is recorded arguing that any form of human thought could, in principle, be reproduced in an artificially created cybernetic system, while maintaining that historical priority would keep human judgment more important than machine judgment in the actual processes of thought. Other participants pushed back, arguing that a machine's capabilities are permanently limited to solving problems assigned to it by the humans who built it. The debate was not resolved. The report simply records that it happened, in detail, in 1962, inside the Soviet Academy of Sciences.

Beyond the philosophy, the report catalogs specific technical work. Soviet scientists had, by 1964, built working pattern recognition systems, including machines that could read typed and handwritten Cyrillic characters, in one case failing to recognize only 2 out of 35,000 test numbers. Researchers at the Institute of Automation and Telemechanics in Moscow had programmed a machine to distinguish handwriting styles. At the Institute of Surgery, a learning system was being trained to assess burn severity from patient case histories, an early and specific instance of what would now be called a diagnostic machine learning system.

The report also documents extensive Soviet work on what it calls brain modeling: laboratory attempts, dating to 1955, to simulate neural processes using conditioned-reflex models drawn from Pavlovian physiology, married to computer hardware built at institutions like the Moscow Power Engineering Institute. The CIA's analysts note that Soviet research on machine learning and pattern recognition compared favorably with US work, while research on planning and induction, the ability of a machine to build an internal model of its environment, was, in both countries, at a very early stage.

The report establishes, in the CIA's own words, that Soviet AI research as of 1964 was receiving substantial and growing state support, that Soviet scientists were debating the philosophical implications of machine intelligence in public and technical literature, and that Soviet analysts assessed their own theoretical and hardware progress as roughly on par with the United States, a judgment the CIA's own analysts appear to have shared.

What the report does not do is predict where any of this would lead. Its authors explicitly declined to evaluate the potential of the field, writing that its future accomplishments could be only dimly seen and not evaluated at present. The report is a snapshot of a single year's open literature, filtered through a single intelligence office's judgment. It says nothing about what Soviet AI research became after 1964, and it makes no claims about classified Soviet military applications, since it draws exclusively on published sources.