The first major public reviews of artificial intelligence research were critical ones. A 1966 report from the U.S. National Research Council judged machine translation a poor return on spending, and a July 1972 survey by Sir James Lighthill for Britain's Science Research Council blamed a "combinatorial explosion" for the field's disappointments. Both documents are often cited as the start of the "AI winters," the periods when funding and enthusiasm fell, yet what they actually argued was narrower than the legend.
Key takeaways
- The 1966 National Research Council report was titled Language and Machines and dealt with machine translation and linguistics, not AI in general.
- Lighthill wrote his survey after two months of reading, at the request of the Science Research Council's chairman, and called it a personal view from outside the field.
- He sorted AI work into three categories: advanced automation, a bridge category centered on building robots, and computer-based study of the central nervous system. The bridge category was judged the most disappointing.
- His forecast for the following 25 years was that the field would split apart, and that applied work would succeed where it drew on detailed knowledge of a specific problem.
- Claims about how much funding changed after each report rest on secondary accounts and should be read with care.
The 1966 review of machine translation
The National Academies Press lists Language and Machines: Computers in Translation and Linguistics as a 1966 National Research Council publication. Its table of contents includes a chapter on the present state of machine translation, a section on avenues to improvement, a short recommendations chapter, and an appendix on government support of machine-translation research. The title shows the scope: translation and linguistics.
The full text could not be retrieved in readable form for this article, so its wording is not quoted here. Lighthill's survey describes the review as having concluded that large sums had been spent on machine translation with very little useful result, and says that later developments had not shaken that conclusion. That is a second-hand but contemporary description of what the review found.
What Lighthill was asked to do, and how he did it
In his introduction, Lighthill explains that the Science Research Council had been receiving a growing number of applications for support in a broad area that went under the name artificial intelligence. The Council's chairman decided to commission an independent report from someone outside the field. Lighthill, then Lucasian Professor of Applied Mathematics at Cambridge, undertook it on the understanding that it would describe how the subject looked to a layperson after two months of reading and conversation. He warned that it must not be viewed as more than a highly personal view.
That framing matters. The report was an adviser's briefing for funders, written by an applied mathematician, and it was never a formal peer review of the field. The Edinburgh AI Applications Institute, which hosts the report, dates it to July 1972.
Three categories, one general complaint
Lighthill divided the field into category A, advanced automation, meaning work meant to replace human beings with machines for specific industrial, military, mathematical, or scientific purposes. Category C was computer-based study of the central nervous system in humans and animals. Category B was a bridge between them, whose basic component he described as building robots.
His judgment was uneven across these. He wrote that progress in the bridge category had been even more disappointing than elsewhere, both in work done and in the reasons offered for doing it. His single general explanation for the disappointments was a failure to recognize the implications of the combinatorial explosion: when the number of possible ways of grouping elements of a knowledge base grows explosively as the base gets larger, methods that work on small problems cannot simply be scaled. This is the same barrier that makes exhaustive search impractical in optimization.
The June 1973 debate
The BBC broadcast a televised debate on the report in June 1973 from the Royal Institution. The Edinburgh archive lists the participants as Lighthill, Donald Michie, Richard Gregory, and John McCarthy. The existence of a televised debate shows that the report was taken as a significant challenge by working researchers, which is separate from the question of what funding bodies did afterward.
What the survey predicted
The final section looks 25 years ahead. Lighthill expected research on advanced automation to become more specialized, noting that AI research was successful to a degree closely correlated with the use made of detailed knowledge of the problem domain. He named the DENDRAL program for inferring chemical structure as a good example of success in category A. For category C he expected substantial advances tied to psychology and neurobiology. For building general robots he predicted continued disappointment, and he expected the unity of the AI field to fission into those linked areas.
Whether the forecast held is a judgment for histories of the field, and this article does not score it.
| Document | Date | Commissioned or issued by | Scope | Main stated finding |
|---|---|---|---|---|
| Language and Machines | 1966 | National Research Council (United States) | Machine translation and linguistics | Spending produced little useful result, as summarized by Lighthill |
| Artificial Intelligence: A General Survey | July 1972 | Science Research Council (United Kingdom) | The AI field in three categories | Combinatorial explosion as the general cause of disappointment; bridge category judged weakest |
Questions for reading any "AI winter" story
- Which document or decision is actually being cited, and who wrote it?
- Does the document cover all of AI or only one area, such as translation?
- Is the effect on funding documented with budgets, or only asserted?
- Did the criticized methods fail because of a technical barrier or because of limits on computing power at the time?
FAQ
Did the Lighthill report end AI research in Britain?
The report's closing section expected some areas to grow and others to disappoint. It is widely linked to reduced support for parts of the field, but the funding figures were not verified for this article, so the size of the effect is left open.
What is a combinatorial explosion?
It is the rapid growth in the number of possible combinations as a problem gets bigger. Searching every combination works for ten items and fails for a thousand, so a method must use knowledge about the problem to avoid trying most of them.
Was machine translation abandoned after 1966?
The 1966 review concerned the state of the technology and the case for government support. The sources checked do not show that all work stopped, and they do not describe later methods, so the question is left to histories of that field.
Why call it a winter at all?
The term is a later metaphor for falling funding and enthusiasm. The reports discussed here do not use it, which is one reason to read the originals instead of summaries.
Related reading
- A year-by-year timeline of AI history
- How AI developed and spread over the decades
- Why researchers set out to build AI
- The history of conversational interfaces
More in AI Basics.
Last verified: October 2026
Sources
- Language and Machines: Computers in Translation and Linguistics, National Research Council, National Academies Press, 1966
- Artificial Intelligence: A General Survey, Lighthill J, Science Research Council, dated July 1972, hosted by the AI Applications Institute, University of Edinburgh
- Lighthill Controversy Debate at the Royal Institution, BBC Television, June 1973, AI Applications Institute, University of Edinburgh