What was the Lighthill report, and why did it hit British AI research so hard?

In the early 1970s, Britain's main funder of university science, the Science Research Council (SRC), had a problem. Applications for money to study artificial intelligence kept growing, and the experts it would normally ask for advice disagreed sharply about whether the field was worth it. So the Council did something unusual. It asked a distinguished scientist from outside the field to read up on AI and give his personal verdict.

That scientist was Sir James Lighthill, Lucasian Professor of Applied Mathematics at Cambridge. His report, "Artificial Intelligence: A General Survey," was blunt. "In no part of the field," he wrote, "have the discoveries made so far produced the major impact that was then promised."

What happened, and when

The SRC chairman, B. H. Flowers, explained his reasoning in a preface. Questions about intelligence were so contested, he wrote, that he did not think a neutral panel of experts could be assembled. A single outsider's view seemed more useful. Lighthill spent about two months reading the literature and corresponding with researchers, and he listed dozens of people he consulted, among them John McCarthy, Marvin Minsky, Donald Michie, Christopher Strachey and Terry Winograd.

He dated his survey July 1972. After discussing it in September 1972, the Council chose to publish it in 1973, together with opposing views, as Artificial Intelligence: a paper symposium. N. S. Sutherland wrote a reply, and R. M. Needham, H. C. Longuet-Higgins and Donald Michie added comments. The Council also invited readers to write in with their opinions. In 1973 Lighthill also defended his views in a recorded public debate, titled Controversy, against Michie, McCarthy and Richard Gregory.

One background factor, according to later accounts, was conflict among the AI researchers at Edinburgh University, home to one of Britain's largest AI groups.

What the report said

Lighthill sorted AI research into three groups, which he labeled A, B and C:

• A, Advanced Automation: replacing people with machines for specific jobs, such as reading printed text, recognizing speech, translating languages or proving theorems.
• C, Computer-based studies of the central nervous system: using programs to model how the brain and mind work, in support of psychology and neurobiology.
• B, the "Bridge": work meant to link A and C, centered on building robots that combine seeing, moving and common-sense problem solving.

He found respectable, if disappointing, progress in A and C. Category B he judged the most discouraging of all, and he doubted that AI held together as a single field. He pointed out that conventional engineering often beat AI at practical tasks: an automatic landing system for aircraft, built with classical control methods, already outperformed human pilots. Machine translation and general speech recognition had consumed large sums with little to show for them.

He also noticed a pattern among the successes. Programs that worked, such as the DENDRAL chemistry program from Stanford, relied on large amounts of detailed knowledge about one narrow problem. General-purpose methods kept failing.

The combinatorial explosion

Lighthill named one cause above all: the combinatorial explosion. As a problem grows, the number of possible combinations to check grows far faster, quickly overwhelming any computer. A method that solves a toy puzzle can be hopeless on a real one.

Here is a worked example. Suppose a program must find the best order in which to visit some cities:

Adding just five cities multiplies the work by hundreds of thousands. Faster computers barely dent numbers like these. That is why, Lighthill argued, AI demonstrations that looked impressive in a small "toy world," such as a robot stacking blocks on a table, did not scale up to the messy real world.

Why it mattered

The report gave official weight to doubts many people already had, and it arrived when governments wanted research to show practical results. It is widely described as the basis for the British government's decision to stop supporting AI research at all but a few universities; accounts name Edinburgh, Sussex and Essex among those that kept going, though one widely read summary says "two" while listing all three. Together with funding changes in the United States, it helped bring on the slump in AI money and interest known as the first "AI winter."

Lighthill was not against everything. He expected steady progress where AI was tied closely to a real application, such as computer-aided design or scientific databases, and in brain modeling linked to psychology and neuroscience. His sharpest criticism was aimed at grand promises of general machine intelligence, such as predictions, quoted in his report, of machines surpassing human intelligence by the year 2000.

Was Lighthill right about artificial intelligence?

Partly. His central observation, that success comes from deep knowledge of a specific domain rather than from general cleverness, was soon confirmed by expert systems such as Stanford's MYCIN and DEC's XCON, which worked precisely because they packed in specialist knowledge. His warning about the combinatorial explosion remains true: no amount of computing power makes brute-force search practical for large problems.

Where he was too gloomy was on pattern recognition. He thought reading handwriting looked "completely out of reach." By 2019, neural networks read handwritten digits routinely, recognize speech in phones and smart speakers, and translate between languages well enough for everyday use. They did it by learning from huge amounts of data, an approach that had little support in 1972. Even then, his point about search still applies: DeepMind's AlphaGo, which beat top Go player Lee Sedol in 2016, succeeded by learning which moves to ignore rather than checking them all.

So the report reads today as a fair critique of the hype of its day and an honest statement of a real mathematical obstacle, paired with forecasts that underestimated what learning machines would eventually do.