An AI winter is a period when disappointment in artificial intelligence leads to a retreat in funding and interest. The history of AI includes several such reversals, as ambitious promises ran into the limits of the available methods and machines.
A forecast of another winter needs to say what will cool and when. Research funding, startup valuations and demand for a particular product can move at different speeds. A useful algorithm can survive the failure of the company that sells it, while a promising line of research can slow when its funding disappears.
Disagreement About AI
Near-term commercial disappointment and longer-term danger can coexist. Someone may expect expensive AI projects to miss their revenue targets over the next few years and also worry that more capable future systems could cause serious harm.
Even a useful system can be a poor investment at a sufficiently high price. A dangerous capability can also be cheap to deploy. Whether a system makes money tells us little by itself about whether its use is safe.
To evaluate a critique, identify the failure it predicts and the conditions that would produce it. A high deployment cost, for example, could limit adoption even as model accuracy improves.
Markets, Incentives and Timing
AI companies have an incentive to emphasize what their products might eventually do. Buyers have to decide what those products can reliably do for them now, including the cost of correcting mistakes and changing existing workflows. If that gap persists, demonstrations can attract investment without producing repeat customers.
Timing matters here as it does in financial markets. A person can correctly anticipate a technology’s usefulness and still be years early about when it will pay for itself. Long investment horizons make some predictions difficult to judge: a forecast that keeps moving its deadline can evade a clear reckoning.
A retreat in investment can also slow the work that might resolve a technology’s weaknesses, by reducing the resources available for experiments and breaking up research teams.
Employment and Competition
Automation can displace particular tasks while creating demand elsewhere. The effect on workers depends on which jobs change and how readily people can move into the new work.
Governments and companies may compete for computing resources or talent while researchers publish methods others can use. A prediction about an AI arms race needs to specify what is being contested and how that competition changes development or deployment.
A History of Winter
Disappointments from the mid-1960s through the 1970s weakened support for AI. In the United States, the 1966 ALPAC report found that existing machine translation still needed human editing and questioned the prospect of useful fully automatic translation. It recommended work on evaluating translations and helping human translators.
- The Mansfield Amendment of 1969 restricted the use of authorized military research funds to projects with a direct connection to a specific military function or operation.
- Minsky and Papert’s Perceptrons (1969) established mathematical limitations of the class of machines they studied, challenging some of the enthusiasm around early neural networks.
- The Lighthill Report of 1973 criticized the results of British AI research. It emphasized how the number of possibilities a program must search can grow explosively as a problem becomes larger, overwhelming gains in computing power.
Interest in symbolic reasoning revived in the 1980s as expert systems were commercialized. The boom gave way to another retrenchment by the early 1990s. Large rule systems could be brittle and expensive to maintain: when the world changed, someone had to identify and revise the affected rules.
The revival of deep neural networks in the mid-2000s brought new ways to learn from data. One milestone was Hinton, Osindero and Teh’s 2006 algorithm for deep belief networks, which trained their layers in stages.
Assessing Another Winter
A decline in funding can arrive unevenly, affecting speculative projects before established applications, and a technical advance can take years to become a dependable product.
To assess a forecast, ask:
- What outcome is being predicted, and how will it be measured?
- By what date, and against which starting point?
- What evidence would make the forecaster revise the claim?
For a commercial application, repeat use and the full cost of obtaining a reliable result are more informative than a striking demonstration. For a research claim, the relevant test is whether the reported advance survives independent evaluation on the problem it is supposed to solve.
