The two conventional winters re-examined with three series: the share of paper titles containing a term, conference attendance, and public-funding records.
Key point
The second winter around 1987 is visible in attendance: AAAI attendance fell from 5,334 in 1986 to 1,685 in 1992 (AI Index 2019 data).
Title counts lag by several years. OpenAlex records with “expert system” in the title peak at 1,857 in 1990 and fall to 490 in 1998; over the same years “neural network” rises from 1,745 to 4,243.
The first winter around 1974 could not be confirmed with numbers: no attendance or year-by-year funding series for that period was found, and OpenAlex title counts are too small.
An "AI winter" is a period in which funding and attention cooled. By convention there were two: the first around the 1973 Lighthill report in Britain and the contraction of DARPA support in the United States, the second after the expert-system and Lisp-machine markets collapsed around 1987. This page re-examines that convention with numbers. We write "confirmed" only where the series point the same way, and we say so where there is no data.
OpenAlex records whose title contains the phrase, divided by all OpenAlex records of that year, per 10,000 (read 2026-11-03). The 1970 value is omitted because undated records appear to have been assigned to that year.
The three phrases tell different stories. “Expert system” rises steeply from 1984, peaks in 1988–1990 at about 9 per 10,000, and falls through the 1990s to 1.5 in 2000. “Artificial intelligence” also declines after a 1987 peak (3.4) to 0.84 in 2000. “Neural network”, by contrast, surges from 1988 to 15.7 in 1994 and does not fall much afterwards. The 1990s "winter" was therefore not a winter for the whole field but a winter for the words symbolic AI and expert system; neural-network research grew in the same years.
The decline in title counts begins several years after the decline in attendance. Publication lag and index coverage may both contribute, so this series alone does not date the winter.
The years around 1974 cannot be judged from this series. In 1973–1975 there are only 34–46 records a year with “artificial intelligence” in the title, and single digits for “perceptron”; the counts are too small to separate change from noise.
Conference attendance
AAAIIJCAINeurIPS
Attendance at AAAI, IJCAI and NeurIPS. 1984–2019 from the AI Index 2019 report, chapter 2; 2020 onward from Our World in Data (AI Index 2026). IJCAI met only in odd years until 2014; the even-year values the report filled in by averaging are excluded.
AAAI attendance falls year after year from 5,334 in 1986 to 1,685 in 1992 and, apart from a rise in 1994 (1,964), on to 1,262 in 1997. IJCAI falls from 4,995 in 1985 to 1,325 in 1997. While the two conferences lose about three quarters of their mid-1980s size (AAAI 76%, IJCAI 73%), NeurIPS grows slowly from a small base (500 in 1995, 926 in 2005). Measured by attendance, the second winter is a long decline that begins in 1987 and lasts to the end of the 1990s, and the recovery of the 2000s came first on the neural-network side. The late-2010s rise is steep at all three (NeurIPS 13,500 in 2019).
Two cautions. First, the NeurIPS values for 1995–2001 increase by exactly 30 a year (500, 530, 560, …); they look like estimates rather than counts, so we read that stretch only as a trend. Second, no attendance data before 1984 was found, which is why the first winter cannot be seen in this series.
Papers accepted at NeurIPS
NeurIPS
Papers listed per year on proceedings.neurips.cc, counted on 2026-11-03. The 2025 listing was incomplete and is omitted.
Accepted papers grow gently from 90 in 1987 to 152 in 2000, then steeply: 370 in 2012, 679 in 2017, 1,898 in 2020, 4,493 in 2024. The neural-network conference grew overall during the 1990s "winter" (not every year: 1992, 1994 and 1997 were below the year before).
Public funding: records exist, year-by-year tables do not
The funding series could not be drawn. What we found are event records.
The Mansfield Amendment (1969). Section 203 of U.S. Public Law 91-121 provides that none of the funds authorized "may be used to carry out any research project or study unless such project or study has a direct and apparent relationship to a specific military function or operation" (83 Stat. 206). It is often cited as redirecting basic research money, but no year-by-year table of AI funding in the following years was found.
The Lighthill report (1973). "Artificial Intelligence: A General Survey", written for the UK Science Research Council, together with the published replies and commentaries, is available in full (Chilton Computing). The statement that it cut British AI support is widely accepted; the source does not give yearly amounts.
DARPA's Strategic Computing Initiative (1983–1993). Usually given as about one billion dollars (Roland & Shiman, Strategic Computing, 2002). Yearly outlays are in the book's appendices or budget documents and were not found as an online table.
Japan's Fifth Generation Computer project (1982–1992). The IPSJ Computer Museum states "approximately ¥54 billion was invested in the project over 11 years". No yearly table.
The conclusion of this page is therefore: the second winter is confirmed by attendance and by title counts. The first winter has event records but no numerical series to confirm it. That does not mean there was no winter; it means we have not found the data to say so with numbers.