Introduction: how many times has hope for AI swelled and shrunk, and how do we count those times?
AI has gone through many periods of swelling hope (booms) and shrinking hope (winters). The numbering "first to fourth" is a way of counting that is common in Japan, and people added it afterward. In this chapter we recount the words ourselves and check the shape of the waves and how the numbers were assigned.
011950–2026 · Waves of words
Seeing booms through word counts
The size of a boom is invisible. What can we count to see the shape of the waves?
Artificial intelligence (AI) means research and technology that reproduce, on computers, the intelligent work of people (thinking, learning, recognizing, speaking, and so on). AI is said to have had many periods of rising hope (booms) and shrinking hope (winters).
To see how big the booms were, this site counted, year by year, how often certain words were used. There are four sets: the phrase artificial intelligence in English books, in world research papers, in Japanese book titles, and in speeches in Japan's National Diet. In the figure on the right, each panel is scaled so that its highest year within the range shown equals 1.
Press "Up to 2011" and a peak from the 1980s, hidden behind today's huge wave, comes into view. In all four panels, there is a peak from the mid-1980s to around 1990, a valley in the 2000s, and a new rise from the 2010s.
scaled value＝count for that year ÷ count for the highest year in the range shown
count for that year
for English books, occurrences per million words; for papers, papers per 10,000; for Japanese books, number of books; for the Diet, number of speeches
Within one panel, the line is drawn on one scale. The panels use different scales, so compare only their shapes.
How to read the figure
The vertical dotted lines mark the years of major events. The 2026 count runs through September 23. Japanese books take time to be registered at the National Diet Library, so 2025–2026 may look lower than it really is (this site's estimate).
The data sources are Google Books Ngram Viewer for English books (through 2022), OpenAlex for papers, the National Diet Library Search for Japanese books, and the Diet Proceedings Search System for the Diet. All were retrieved on September 23, 2026.
Sources & confidence (3)Seeing booms through word counts
1950–2022Verified in the original
English books: in the 1950s, from the 1960s. The 1987 peak was passed in 2018
We counted year by year how often particular words were used in English books. was most frequent in 1956, and around 1962 was overtaken by . Artificial intelligence peaked in 1987 (about 2.0 times per million words) and fell in the 2000s (about 0.72 in 2007). It then passed the 1987 peak in 2018 (about 2.3) and reached about 6.4 in 2022. In 1965, called the , it was only about 0.11, about 1/18 of the 1987 peak.
The popular story:“The 1960s were the golden age of AI” The original source says:Looking at how often the term appears in English books, 1965 is only about 1/18 of the 1987 peak. What was used a lot in the 1950s was a different term, .
What we still don’t know:The frequency of a word is not the size of the boom itself (names change, and the number of books differs from year to year).
Share of the world’s papers with in the title: a peak in 1987, a valley in the 2000s, the peak passed in 2018, and in 2025 about 15 times that
We counted year by year the share of the world’s papers that have "artificial intelligence" in the title. It peaked at about 3.4 per 10,000 papers in 1987 and hit a valley of about 0.8 in 2008–2010. In 2018 it reached about 4.4, passing the 1987 level, and in 2025 it reached about 52.
[Japan] In the Diet and in book titles, the second peak is in 1990 (about 3 years later than in English) and the third in 2019; rising again from 2023
From the Diet’s records of speeches and from Japanese book titles, we counted how many times and “AI” (recounted with parts of other words removed, so that it means AI) appear. In book titles, is highest in 1990 (37 books). In the Diet, peaks in 1985 (7 mentions) and 1990 (9), and was almost 0 from 1993 to 2014. In the Diet, “AI” meaning AI went 11 in 2015 → 593 in 2019 → down to 336 in 2021 → 761 in 2023 → 2,424 in 2025 (the year the AI law was made). “Generative AI” was 0 until 2022 and reached 172 in 2023.
If you search for just the two letters “AI”, the counts in the Diet records jump in 1996 (AIDS, in the drug-induced HIV scandal), 2009 (AIG, the insurance company) and 2012 (the AIJ case about pension fund management). In book titles it also picks up the word “love” (read “ai”). When we recount after checking what each one means, those jumps disappear.
What we still don’t know:The reason the counts fell in 2020–2021 cannot be pinned down to one cause (it overlaps with the period of the infectious disease epidemic).
What the figure shows:All 4 different sources show a peak around the mid-1980s to 1990, a low in the 2000s, and renewed growth from the 2010s. Narrowing the view reveals older peaks hidden by today's large wave.
021940–1995 · Names
1950s: the word was
Before became the common term, what were thinking machines called?
Let us look at how often words appear in English books, measured in occurrences per million words. In the 1950s the most used term was electronic brain, which peaked in 1956 (about 0.21 occurrences). 1956 is also the year the name was born. Artificial intelligence overtook electronic brain in 1962.
Even in 1965, which is called a , artificial intelligence appears only about 0.11 times. That is about 1/18 of the 1987 peak (about 2.0 times).
In the figure on the right, switch the vertical scale to "10× steps" and the movement of small numbers becomes visible. On the "Normal" scale, the 1960s look almost like 0. When the name changes, a wave followed through a single word disappears from view partway through.
occurrences per million words＝number of times the word appears ÷ number of words in that year's books × 1,000,000
10× steps
a scale where each tick multiplies the value by 10 (a logarithmic scale)
What you think you knowThe popular story:The 1960s were AI's golden age The original source says:By the number of times appears in English books, 1965 is about 1/18 of the 1987 peak. What was common in the 1950s was a different word,
Sources & confidence (1)1950s: the word was
1950–2022Verified in the original
English books: in the 1950s, from the 1960s. The 1987 peak was passed in 2018
We counted year by year how often particular words were used in English books. was most frequent in 1956, and around 1962 was overtaken by . Artificial intelligence peaked in 1987 (about 2.0 times per million words) and fell in the 2000s (about 0.72 in 2007). It then passed the 1987 peak in 2018 (about 2.3) and reached about 6.4 in 2022. In 1965, called the , it was only about 0.11, about 1/18 of the 1987 peak.
The popular story:“The 1960s were the golden age of AI” The original source says:Looking at how often the term appears in English books, 1965 is only about 1/18 of the 1987 peak. What was used a lot in the 1950s was a different term, .
What we still don’t know:The frequency of a word is not the size of the boom itself (names change, and the number of books differs from year to year).
What the figure shows:The term “electronic brain” was widely used in the 1950s, and “Artificial Intelligence” overtook it in 1962. Tracking only 1 term makes the wave disappear as soon as the name changes.
031950–2025 · Shares
The paper count itself keeps growing
If AI papers increase eightfold, can we say interest in AI also grew eightfold?
Among the world's research papers, those with "artificial intelligence" in the title numbered 585 in 1987 and 4,640 in 2018. Looking only at counts, that is about eight times as many. But over the same period, the total number of papers in the world also grew, from about 1.73 million to about 10.44 million.
So let us convert to papers per 10,000 papers (a share). It is about 3.4 in 1987 and about 4.4 in 2018, only about 1.3 times as much.
Try switching in the figure on the right. In raw counts, 2005 (565 papers) looks like a return to near the 1987 peak. As a share, about 0.8 per 10,000 in 2008–2010 is a deep valley. Change how you count, and a winter appears or vanishes.
per 10,000＝number of papers with AI in the title ÷ number of all papers that year × 10,000
graph below
multiples with 1987 set to 1. The wider the gap between the two lines, the bigger the difference between raw counts and shares
Even as a share, trends in how titles are written and the range of papers registered in OpenAlex differ from year to year.
Sources & confidence (1)The paper count itself keeps growing
1950–2025Verified in the original
Share of the world’s papers with in the title: a peak in 1987, a valley in the 2000s, the peak passed in 2018, and in 2025 about 15 times that
We counted year by year the share of the world’s papers that have "artificial intelligence" in the title. It peaked at about 3.4 per 10,000 papers in 1987 and hit a valley of about 0.8 in 2008–2010. In 2018 it reached about 4.4, passing the 1987 level, and in 2025 it reached about 52.
What the figure shows:AI papers increased 8 times, but all papers increased 6 times, so AI's share grew only 1.3 times. Changing the way of counting can make a winter appear or disappear.
Multiples with 1987 set to 1 (vertical axis in steps of 10)
04Regular expressions · Preprocessing
Searching for the two letters "AI"
If we search for the two letters “AI”, can we count AI topics correctly?
If you search Japan's Diet speeches for the two letters “AI”, the search also catches AIDS in the drug-induced AIDS scandal, AIG the insurance company, and even the missile name AIM-9L.
So this site read the content of the hits, decided whether to count each one, and made a rule: count only AI that has no other Latin letters before or after it. A way of writing such letter patterns briefly is called a regular expression.
The figure on the right shows 18 sentences taken from Diet speeches. It lines up the judgment made by a person who read them against the judgment of the rule. Turn the switch off and you can see which sentences get counted by mistake.
nothing of … follows right after. [Ａ-Ｚ] means any one letter fromＡtoＺ
Working on 18 sentences does not guarantee it is right on other sentences. This rule alone also cannot identify things like people's names written in Latin letters.
Connection to the school subject Information
Removing unneeded items and unifying notation before counting is called data preprocessing. It is the same work as collecting, organizing, and visualizing data in Information I (Japan's required high-school informatics course). The whole chapter also connects to the history of information technology studied in Information II (the elective follow-up course) (both connections are this site's view).
Sources & confidence (1)Searching for the two letters "AI"
1947–2026Verified in the originalJapan
[Japan] In the Diet and in book titles, the second peak is in 1990 (about 3 years later than in English) and the third in 2019; rising again from 2023
From the Diet’s records of speeches and from Japanese book titles, we counted how many times and “AI” (recounted with parts of other words removed, so that it means AI) appear. In book titles, is highest in 1990 (37 books). In the Diet, peaks in 1985 (7 mentions) and 1990 (9), and was almost 0 from 1993 to 2014. In the Diet, “AI” meaning AI went 11 in 2015 → 593 in 2019 → down to 336 in 2021 → 761 in 2023 → 2,424 in 2025 (the year the AI law was made). “Generative AI” was 0 until 2022 and reached 172 in 2023.
If you search for just the two letters “AI”, the counts in the Diet records jump in 1996 (AIDS, in the drug-induced HIV scandal), 2009 (AIG, the insurance company) and 2012 (the AIJ case about pension fund management). In book titles it also picks up the word “love” (read “ai”). When we recount after checking what each one means, those jumps disappear.
What we still don’t know:The reason the counts fell in 2020–2021 cannot be pinned down to one cause (it overlaps with the period of the infectious disease epidemic).
What the figure shows:Searching for the 2 letters “AI” also counts AIDS, AIG, and missile names. Adding a rule that no Latin letters may come directly before or after them makes the results match human judgments.
051975–2026 · Preprocessing
Erasing peaks that were never there
How much does the shape of the wave change before and after recounting?
If you count the two letters “AI” as they are in Diet speeches, you get peaks of 85 hits in 1996, 152 in 2009, and 746 in 2012. What they contain is AIDS (the drug-induced AIDS scandal), AIG (an insurance company), and AIJ (a pension scandal). If you recount with the rule from the previous spread, those three years become 0.
The same thing happens in Japanese book titles. The national library's search also matches phonetic readings, so a search for “AI” even brings up books about “love” (愛, which is read ai). In 1990 there are 152 matches, but only 14 books have AI standing alone in the title.
In the figure on the right, switch between "Restore as is" and "Recount". The graph below shows what share of the hits was left after recounting. In the Diet, the share was 4% in 2015 and becomes 95% in 2019.
share remaining＝recounted number ÷ number counted as is
The recounting rule was made by people reading the search results, and the examples counted and not counted were kept year by year.
What you think you knowThe popular story:Searching the two letters “AI” tells you how many AI topics there are The original source says:In the Diet, peaks unrelated to AI appear in 1996, 2009, and 2012. They vanish when the content is checked and recounted
Sources & confidence (1)Erasing peaks that were never there
1947–2026Verified in the originalJapan
[Japan] In the Diet and in book titles, the second peak is in 1990 (about 3 years later than in English) and the third in 2019; rising again from 2023
From the Diet’s records of speeches and from Japanese book titles, we counted how many times and “AI” (recounted with parts of other words removed, so that it means AI) appear. In book titles, is highest in 1990 (37 books). In the Diet, peaks in 1985 (7 mentions) and 1990 (9), and was almost 0 from 1993 to 2014. In the Diet, “AI” meaning AI went 11 in 2015 → 593 in 2019 → down to 336 in 2021 → 761 in 2023 → 2,424 in 2025 (the year the AI law was made). “Generative AI” was 0 until 2022 and reached 172 in 2023.
If you search for just the two letters “AI”, the counts in the Diet records jump in 1996 (AIDS, in the drug-induced HIV scandal), 2009 (AIG, the insurance company) and 2012 (the AIJ case about pension fund management). In book titles it also picks up the word “love” (read “ai”). When we recount after checking what each one means, those jumps disappear.
What we still don’t know:The reason the counts fell in 2020–2021 cannot be pinned down to one cause (it overlaps with the period of the infectious disease epidemic).
What the figure shows:Simply counting the 2 letters “AI” creates peaks unrelated to AI in 1996, 2009, and 2012. Counting again with the corrected method gives 0 matches in those years.
Share remaining after recounting (%)
061959–2026 · Japan's National Diet
How the Diet's names for it changed
In Japan's Diet, when and by what name were thinking machines called?
For each word, we counted by year the Diet speeches that contain it. “Electronic brain” appears first in 1959, and even in its busiest year, 1966, only 4 times. appears from 1981 and reaches 33 in 1985. Try choosing words in the figure on the right.
first appears in 1983 (within the range we found). Masaki Moriya of the Nomura Research Institute named, as a technology of the 21st century, (translated). After that, from 1993 to 2014 there are almost 0 hits, and from 2015 it suddenly rises.
In 2017, “AI” in the sense of AI appears 268 times, overtaking at 145. In 2023, “generative AI” appears, 172 times.
height on the vertical axis＝log₁₀ (number of speeches＋1) (the＋1 is so that years with 0 speeches can also be drawn)
The counts for , , and are the number of search hits as they are (we did not check the content one by one). “AI” is the recounted number.
Sources & confidence (2)How the Diet's names for it changed
23 February 1983Verified in the originalJapan
[Japan] The term first appears in the Diet (within what was found): (translated)
In a joint committee hearing of the House of Councillors, Masaki Moriya of the Nomura Research Institute, who had been invited to the Diet to give his views, named “artificial intelligence and fifth-generation computers, which have recently been talked about a lot” (translated) as technologies of the 21st century. Within what the search found, this is the oldest appearance of in the Diet record.
What we still don’t know:This is the oldest one found in the Diet record, not the time the term was first used in society.
[Japan] In the Diet and in book titles, the second peak is in 1990 (about 3 years later than in English) and the third in 2019; rising again from 2023
From the Diet’s records of speeches and from Japanese book titles, we counted how many times and “AI” (recounted with parts of other words removed, so that it means AI) appear. In book titles, is highest in 1990 (37 books). In the Diet, peaks in 1985 (7 mentions) and 1990 (9), and was almost 0 from 1993 to 2014. In the Diet, “AI” meaning AI went 11 in 2015 → 593 in 2019 → down to 336 in 2021 → 761 in 2023 → 2,424 in 2025 (the year the AI law was made). “Generative AI” was 0 until 2022 and reached 172 in 2023.
If you search for just the two letters “AI”, the counts in the Diet records jump in 1996 (AIDS, in the drug-induced HIV scandal), 2009 (AIG, the insurance company) and 2012 (the AIJ case about pension fund management). In book titles it also picks up the word “love” (read “ai”). When we recount after checking what each one means, those jumps disappear.
What we still don’t know:The reason the counts fell in 2020–2021 cannot be pinned down to one cause (it overlaps with the period of the infectious disease epidemic).
What the figure shows:In Japan's parliament, names have passed the baton from electronic brain to fifth generation, artificial intelligence, AI, and generative AI. Artificial intelligence first appeared in 1983, and mentions rose sharply from 2015.
071975–2005 · Peak offset
Did Japan's peak come late?
Did Japan's second peak come later than the peak in English books?
In English books, is most frequent in 1987. In Japanese book titles, the most frequent year is 1990 (37 books). Comparing the peak years, Japan is 3 years late.
But a peak year moves with the ups and downs of a single year. Japanese books also have 30 each in 1987 and 1988. Smooth by averaging 3 years at a time, and Japan's peak becomes 1989, shrinking the offset to 2 years.
So let us compare the whole shape. Slide the Japanese line left and right and compute how similar the two lines' shapes are (the correlation coefficient). For 1975–2005, the best fit is with no sliding (0 years). The “offset” can be 3 years or 0 years depending on how you compare.
r＝Σ(x − x̄)(y − ȳ) ÷ √( Σ(x − x̄)² × Σ(y − ȳ)² )
x
value for English books
y
value for Japanese books (shifted by k years)
x̄, ȳ
averages. The closer r is to 1, the more the shapes match
The event cards below are written by comparing peak years (about 3 years late). Both are different views of the same data.
Sources & confidence (2)Did Japan's peak come late?
1947–2026Verified in the originalJapan
[Japan] In the Diet and in book titles, the second peak is in 1990 (about 3 years later than in English) and the third in 2019; rising again from 2023
From the Diet’s records of speeches and from Japanese book titles, we counted how many times and “AI” (recounted with parts of other words removed, so that it means AI) appear. In book titles, is highest in 1990 (37 books). In the Diet, peaks in 1985 (7 mentions) and 1990 (9), and was almost 0 from 1993 to 2014. In the Diet, “AI” meaning AI went 11 in 2015 → 593 in 2019 → down to 336 in 2021 → 761 in 2023 → 2,424 in 2025 (the year the AI law was made). “Generative AI” was 0 until 2022 and reached 172 in 2023.
If you search for just the two letters “AI”, the counts in the Diet records jump in 1996 (AIDS, in the drug-induced HIV scandal), 2009 (AIG, the insurance company) and 2012 (the AIJ case about pension fund management). In book titles it also picks up the word “love” (read “ai”). When we recount after checking what each one means, those jumps disappear.
What we still don’t know:The reason the counts fell in 2020–2021 cannot be pinned down to one cause (it overlaps with the period of the infectious disease epidemic).
English books: in the 1950s, from the 1960s. The 1987 peak was passed in 2018
We counted year by year how often particular words were used in English books. was most frequent in 1956, and around 1962 was overtaken by . Artificial intelligence peaked in 1987 (about 2.0 times per million words) and fell in the 2000s (about 0.72 in 2007). It then passed the 1987 peak in 2018 (about 2.3) and reached about 6.4 in 2022. In 1965, called the , it was only about 0.11, about 1/18 of the 1987 peak.
The popular story:“The 1960s were the golden age of AI” The original source says:Looking at how often the term appears in English books, 1965 is only about 1/18 of the 1987 peak. What was used a lot in the 1950s was a different term, .
What we still don’t know:The frequency of a word is not the size of the boom itself (names change, and the number of books differs from year to year).
What the figure shows:Comparing peak years puts Japan 3 years behind, but smoothing the data or comparing whole shapes can reduce the gap to 2 years or 0 years. The gap depends on the comparison method.
Correlation coefficient r for each number of years shifted
08Counting method · Peaks
How many peaks does a boom have?
If we count “boom” using only the lines of word counts, how many are there?
Hearing “first to fourth”, it sounds as if there were four booms. So what happens if we count peaks from the lines with a fixed procedure?
The counting method is this. Follow the line, and when it drops by a set share (the threshold) or more from its highest point, count one peak. A peak that does not reach 10% of the largest year's value is not counted.
Move the threshold in the figure on the right. In English books, with the initial threshold, the only peaks found are one in 1987 and the current rise, which has not yet come down. The 1960s are too small to count as a peak. In Japanese book titles, counting year by year, even small ups and downs of 5 or 10 books become peaks. Smooth over 3 years and two peaks remain, around 1989 and around 2018. The number of booms is not settled by looking at the line alone; it changes with the threshold you count by.
count as a peak ⇔ (peak height − the valley after it) ÷ peak height ≥ threshold
threshold
by what percent the line must fall to count as one peak
smoothing
averaging with the years before and after
Sources & confidence (2)How many peaks does a boom have?
1950–2022Verified in the original
English books: in the 1950s, from the 1960s. The 1987 peak was passed in 2018
We counted year by year how often particular words were used in English books. was most frequent in 1956, and around 1962 was overtaken by . Artificial intelligence peaked in 1987 (about 2.0 times per million words) and fell in the 2000s (about 0.72 in 2007). It then passed the 1987 peak in 2018 (about 2.3) and reached about 6.4 in 2022. In 1965, called the , it was only about 0.11, about 1/18 of the 1987 peak.
The popular story:“The 1960s were the golden age of AI” The original source says:Looking at how often the term appears in English books, 1965 is only about 1/18 of the 1987 peak. What was used a lot in the 1950s was a different term, .
What we still don’t know:The frequency of a word is not the size of the boom itself (names change, and the number of books differs from year to year).
[Japan] In the Diet and in book titles, the second peak is in 1990 (about 3 years later than in English) and the third in 2019; rising again from 2023
From the Diet’s records of speeches and from Japanese book titles, we counted how many times and “AI” (recounted with parts of other words removed, so that it means AI) appear. In book titles, is highest in 1990 (37 books). In the Diet, peaks in 1985 (7 mentions) and 1990 (9), and was almost 0 from 1993 to 2014. In the Diet, “AI” meaning AI went 11 in 2015 → 593 in 2019 → down to 336 in 2021 → 761 in 2023 → 2,424 in 2025 (the year the AI law was made). “Generative AI” was 0 until 2022 and reached 172 in 2023.
If you search for just the two letters “AI”, the counts in the Diet records jump in 1996 (AIDS, in the drug-induced HIV scandal), 2009 (AIG, the insurance company) and 2012 (the AIJ case about pension fund management). In book titles it also picks up the word “love” (read “ai”). When we recount after checking what each one means, those jumps disappear.
What we still don’t know:The reason the counts fell in 2020–2021 cannot be pinned down to one cause (it overlaps with the period of the infectious disease epidemic).
What the figure shows:A line graph can have 1 peak or 5, depending on the counting rule. With a fixed rule, English books show only the 1987 peak and the current rise.
091982–2023 · Numbers
Which era did the boom numbers point to?
When, and for which era, were numbers like and assigned?
The numbers were not all assigned together in the 2010s. In the full-text search of digitized books and magazines at the National Diet Library, the 1982 issue of the magazine already contains the phrase (translated). It refers to a boom about 20 years earlier.
But in 1985, people did not agree on how to count even the booms of that same decade, the 1980s. A book in March called it (translated). In the Diet in April, Takemochi Ishii of the University of Tokyo said, (translated). A July issue of the magazine (Railway Communications) listed (translated) as one of its recent feature articles.
Choose a number in the figure on the right and, for each source that used it, a line joins the year it was used (◆) to the era it points to (a band). Choose "Fourth" and the line that pointed to the future in 2017 points, in 2023, to the period after 2017.
What you think you knowThe popular story:People tend to think “The numbers first, second... were added later (in the 2010s)” The original source says:A 1982 magazine already had , so the booms were being called by number from the middle of the second boom. But how to count had not been settled
Notes on how we searched
For the magazines and books of 1982, 1985, 1993, and 1995, we checked only by text extracted by machine reading (the page photographs are limited to in-library viewing, so we did not confirm them). That is why the confidence grade is marked “Weakly confirmed”.
The oldest speech in the Diet proceedings where all three numbers appear together is from 2017 (within the range we found). In the full-text search of J-STAGE, which collects Japanese academic journals, the oldest text where appears was from 2014.
Sources & confidence (10)Which era did the boom numbers point to?
November 1982Weakly confirmedJapan
A 1982 magazine already has the “first artificial intelligence boom” (translated) ── the oldest numbered expression found
An article in the magazine (November 1982 issue) has a sentence saying that companies that had become disillusioned with the results of the “first artificial intelligence boom” (translated), which arose some twenty years earlier, had kept their interest in this field lost. So in 1982 the boom of about twenty years before (by this site's calculation, the early 1960s) was already being called the (translated). In a full-text search of books and magazines digitized by the National Diet Library, this is the oldest example found of an AI boom expression with a number (within what was found).
The popular story:It is often thought that numbers like first boom and second boom were added afterward (in the 2010s) The original source says:A 1982 magazine already says “first artificial intelligence boom” (translated), so booms were being called by number from the middle of the second boom
What we still don’t know:I have not seen the article's title, author, or pages (library use only). Because it is OCR (text read by machine), it needs to be checked against the printed page.
1 sources and how they were checked
President 20(11) (1982 · Approximate frame number in the National Diet Library Digit… · Original source · Could not be opened)
March 1985Weakly confirmedJapan
A 1985 book has (translated) (the 82 is thought to mean 1982)
Mr. Kaminuma's book (1985) contains a chart-like listing of years and events, with entries such as (translated) and “82 Second Artificial Intelligence Boom: Fifth Generation Project” (translated). In the column (thought to mean 1982; this site's estimate), “second artificial intelligence boom” (translated) and (translated) are listed side by side. The text right after is garbled in the scanned reading, but in another book edited by Kaminuma the same year, the same sequence is followed by (translated). This is the oldest (translated) wording we have found.
What we still don’t know:We have not seen the whole chart (from what year to what year it runs, and what else is written in it). In the text data of what seems to be the same chart in , August 1986, the DNA double-helix model of Watson and Crick in '53 is followed by (translated). It seems to call the late 1950s the (translated; this site's estimate), but this needs to be checked on the printed page.
4 April 1985Verified in the originalSources disagreeJapan
In the 1985 Diet, AI is explained as being said to be (translated)
Takemochi Ishii of the University of Tokyo was called to the Diet (a committee of the House of Representatives) and explained, (translated). He said the first was an effort centered on the United States from the mid to late 1960s, and the second was Japan's Fifth Generation Computer project. The idea of counting booms was already around in the middle of the second boom.
What we still don’t know:We do not know who started the phrase (translated) before 1985 (this is the oldest instance found within the Diet records).
2 sources and how they were checked
Where sources disagree:Ishii gives the period of the first boom as “the mid to late 1960s.” Many Japanese sources of the 2010s say “the late 1950s to the 1960s” (such as the 2016 White Paper of the Ministry of Internal Affairs and Communications; see another row)
In 1985, one feature called the boom of the time (the way of counting had not been settled)
The feature announcement in the magazine “Railway Telecommunications” (July 1985 issue) lists (translated) and (translated). Meanwhile, in the Diet at the same time it was called “now in its second boom” (translated) (April 1985), and in a book “second artificial intelligence boom” (translated) (March 1985). Some people counted the 1980s boom as the (translated) and others as the (translated).
The popular story:The second AI boom is often said to have been called the second boom from the start The original source says:In 1985, some features called the same boom “The Arrival of the First Artificial Intelligence Boom.” The way of counting was not settled from the beginning.
What we still don’t know:From the excerpt alone, we cannot tell which magazine and which issue ran the feature (it says June issue and July issue, and appears inside the introduction of an article on classifying telecommunications equipment).
1 sources and how they were checked
Where sources disagree:This feature calls the 1980s boom (translated), while Mr. Kaminuma's (1985) calls it (translated), and Takemochi Ishii in the Diet (1985) calls it (translated).
A 1993 book says (translated): the oldest use found of the phrase (translated)
A 1993 book contains a sentence about research on reproducing the mechanisms of human perception on computers: (translated). It is the oldest example in which (translated) appears in the form rather than “artificial intelligence boom” (found within the range covered by full-text search of the National Diet Library Digital Collections).
What we still don’t know:We have not seen which researcher is speaking or which page it is on
A 1995 doctoral thesis says (translated): the oldest form of found
A doctoral thesis from Osaka University (1995; its chapter titles include and others) contains the sentence (translated). It goes straight on to a discussion of Generic Task, a research area (marked with the reference [Chandrasekaran86]). It is the oldest example in which a numbered expression appears in the form rather than “artificial intelligence boom” (found within the range covered by full-text search of the National Diet Library Digital Collections).
What we still don’t know:We have not seen which page of the thesis it is on (the material comes from the library's transmission service).
The earliest paper found that has the (translated) (2014; the first and second are there too)
An article by Takahira Yamaguchi of Keio University in the Journal of the Japanese Society for Artificial Intelligence, looking back at the Fifth Generation Computer. It calls the 1950s–60s the (translated), the 1970s the (translated), and the 1980s the (translated). It describes the passing of the baton from national basic research programs (CYC, CALO and others) to programs that solve companies' problems, like Watson and Siri, as (translated). In a full-text search of J-STAGE (a Japanese database of academic papers), it is the oldest document in which these three terms appear together (within what was found).
What we still don’t know:Newspapers, magazines and web articles are outside the scope of this search. It remains possible that a document before 2014 wrote .
Yutaka Matsuo's (2015): chapter names use the first, second and third AI booms
This is a book by Yutaka Matsuo of the University of Tokyo. It divides history by boom number, using chapter names such as , , , and . The first section of Chapter 2 is . The 2016 white paper of the Ministry of Internal Affairs and Communications notes in a footnote that its section on the history of AI was written with reference to this book and others.
What we still don’t know:We have not checked which years the book gives for the start and end of each boom (the years in the text and figures), because we have not opened the book.
2015: the of AI appears in Japan's parliament (as told from Yutaka Matsuo's talk)
In the House of Representatives' Committee on Economy, Trade and Industry, Representative Akimasa Ishikawa spoke. Saying he had heard from an associate professor, Mr. Matsuo of the University of Tokyo, at the Liberal Democratic Party, he introduced the idea that (translated). At the center of it is deep learning. He also said that, based on big data, computers will become able to work out for themselves the variables and algorithms that humans had decided until now. In the Diet minutes, this is the oldest remark found so far in which the AI appears.
What we still don’t know:We have not found the original record (source) of what Mr. Matsuo actually said at the Liberal Democratic Party meeting. What was confirmed here is that Representative Ishikawa relayed it that way in the Diet.
In 2017, the First, Second and Third AI Booms Are All Named Together in the Diet
In the Committee on Internal Affairs and Communications of the House of Councillors, member Takae Ito said, (translated). She went on through the (translated) of the 1980s (saying, as something she had heard, that at that time the Ministry of International Trade and Industry (MITI) spent 57 billion yen on the Fifth Generation Computer project) and said, (translated). This is the earliest remark found so far in which all three numbers appear together in the Diet's proceedings.
What the figure shows:Boom numbers were not all added in the 2010s: a 1982 magazine already used “first.” But in 1985, the same boom was called both first and second, so there was no settled way to count them.
101950–2026 · Periodization
Same number, different start year
From what year to what year are the booms and winters? Do sources agree?
Even with the same number, the years differ by source. The white paper of Japan's Ministry of Internal Affairs and Communications (2016) puts the third boom from the 2000s to the present. The picture scroll on the cover of the Journal of the Japanese Society for Artificial Intelligence (2023) draws it as starting around 2010, triggered by the 2012 image-recognition contest. The winter after the second boom also differs: the white paper says from around 1995, the scroll from around 1992 to around 2010. Even within the same white paper, one place puts the second boom in the 1980s and another in the 1980s to 90s.
English-language Wikipedia puts the numbers on winters, not booms, and gives the two winters as 1974–1980 and 1987–2000.
Move "Year to view" in the figure on the right and each source's name for that year lines up. For 2005, for example, the white paper says third boom, the scroll says winter, and the English explanation puts it in neither winter.
A phrase like “the 1980s” was drawn from the start to the end of that decade. The ends of the bands are this site's drawing, not the sources' words.
Where sources disagree
Even within the scroll, a passage on 2017 onward says (translated), and the society has not decided on it as an organization.
In the 1985 Diet session, Takemochi Ishii described the first boom as American-centered attempts from the mid to late 1960s. Most sources from the 2010s put it in the late 1950s to the 1960s.
Sources & confidence (3)Same number, different start year
2016Verified in the originalSources disagreeJapan
The Ministry of Internal Affairs and Communications' white paper (2016) explains the history of AI in a table of the first, second and third booms and the winters
The Ministry of Internal Affairs and Communications' white paper (the 2016 edition of the White Paper on Information and Communications) shows the history of AI in a table of three booms and two winters. The first boom was the late 1950s to the 1960s (search and inference); the second was the 1980s (giving knowledge; including the Japanese government's large “Fifth Generation Computer” project), with a winter again from around 1995; the third runs from the 2000s to the present (machine learning and deep learning). It also wrote that in the past two booms, society's expectations rose above technical limits, and (translated; divergence means a gap). It says the same caution is needed for the present third boom.
What we still don’t know:The source of the chart is the ministry's commissioned survey (2016). A note says the historical part of the text drew on an article by Hitoshi Matsubara (then president of the Japanese Society for Artificial Intelligence) and on Yutaka Matsuo's (2015).
2 sources and how they were checked
Where sources disagree:Within the same white paper, Chapter 1 (page 5) puts the second boom at “the 1980s–90s,” while Chapter 4 (page 235) puts it at “the 1980s” with “a winter again from around 1995.” Figure 4-2-1-5 gives the development of ELIZA as “1964” (the Ministry of Education, Culture, Sports, Science and Technology, 2024, says 1966. The year of ELIZA is confirmed in the u04 record).
January 2023Verified in the originalSources disagreeJapan
How the Japanese Society for Artificial Intelligence's “History Scroll of Artificial Intelligence” (2023) Divides the Booms and Winters
Over one year, the Japanese Society for Artificial Intelligence drew the history of AI as a picture scroll on the covers of its journal. Its division of time is: picture 1 = from the start of the field to the first boom (late 1950s–1960s, fading from the late 1960s as limits came into view); picture 2 = from the winter to the second boom; picture 3 = the second winter (around 1992 to around 2010); picture 4 = the first half of the third boom (around 2010–2016, triggered by the 2012 image recognition contest); picture 5 = the second half of the third boom (2017 on; (translated); the original uses a formal word for “by now”) through the generative AI boom; picture 6 = the future as drawn by researchers.
What we still don’t know:Even in the scroll, the (translated) is only put as (translated), and the society has not decided on it.
4 sources and how they were checked
Where sources disagree:On the start of the third boom, the scroll says “starting with the 2012 image recognition contest” and “from around 2010.” The Ministry of Internal Affairs and Communications (2016) says “2000s,” and the Ministry of Education, Culture, Sports, Science and Technology (2024) says “around 2010.” On the winter after the second boom, the scroll says “around 1992 to around 2010,” and the Ministry of Internal Affairs and Communications (2016) says “from around 1995.”
Calling booms by number, “first, second … boom”, is mostly a Japanese way of counting (English numbers the )
Word frequencies in English-language books (Google Books Ngram) show that appears from 1983 and rose sharply in the late 2010s. Expressions that put numbers on winters, such as and “second AI winter”, also exist. On the other hand, , and did not appear in countable amounts. In Japanese, “first artificial intelligence boom” (translated) can be seen in a 1982 magazine, (translated) in a 1985 book, and (translated) in the Diet in 1985. From 2014, the first, second and third AI booms were used in an academic society journal, and they also appear in a chapter title of a 2015 book by Yutaka Matsuo and in government white papers from 2016 on.
The popular story:People tend to think “first to fourth AI booms” is a worldwide way of counting The original source says:Numbered “booms” are mostly a Japanese expression. English-language books put numbers on the winters (“The AI Winter”: first/second AI winter) and do not number booms. The U.S. DARPA uses a different way of counting,
What we still don’t know:We have not looked at how this is counted in Chinese (such as phrases for a “third wave” of AI) or Korean. Ngram covers only books through 2022, so we do not know the English expressions from 2023 onward.
What the figure shows:Even the same boom number covers different years in different sources. For example, 2005 belongs to the third boom in a white paper, a winter in an illustrated history, and no winter in an English account.
112016 · The white paper's explanation
Hope versus ability: the “divergence”
Why do booms end? How did the government white paper explain it?
The 2016 white paper of Japan's Ministry of Internal Affairs and Communications explained that in the past two booms, society's expectations exceeded AI's technical limits. It then said the booms are (translated) (the word means a gap or a distance), and that the same caution is needed for the third boom.
The figure on the right is this site's moving version of that explanation (all values are made-up numbers). The line for what AI can do (green) rises now and then like stairs. The line for expectations (purple) swells beyond it. The state of being too high continues, and when it becomes clear, expectations shrink.
Use the sliders to change how easily expectations swell and how many years it takes to notice the gap. In this figure, the more years it takes to notice, the higher expectations swell and the bigger the fall. If expectations swell only a little, no winter comes.
shrink ⇔ (expectation − what AI can do)＞threshold, lasting d years
d
years until the gap is noticed. The threshold is 30% of what AI can do (a made-up number)
The figure is one way of reading the white paper's explanation, not evidence that booms always ended this way.
Sources & confidence (1)Hope versus ability: the “divergence”
2016Verified in the originalSources disagreeJapan
The Ministry of Internal Affairs and Communications' white paper (2016) explains the history of AI in a table of the first, second and third booms and the winters
The Ministry of Internal Affairs and Communications' white paper (the 2016 edition of the White Paper on Information and Communications) shows the history of AI in a table of three booms and two winters. The first boom was the late 1950s to the 1960s (search and inference); the second was the 1980s (giving knowledge; including the Japanese government's large “Fifth Generation Computer” project), with a winter again from around 1995; the third runs from the 2000s to the present (machine learning and deep learning). It also wrote that in the past two booms, society's expectations rose above technical limits, and (translated; divergence means a gap). It says the same caution is needed for the present third boom.
What we still don’t know:The source of the chart is the ministry's commissioned survey (2016). A note says the historical part of the text drew on an article by Hitoshi Matsubara (then president of the Japanese Society for Artificial Intelligence) and on Yutaka Matsuo's (2015).
2 sources and how they were checked
Where sources disagree:Within the same white paper, Chapter 1 (page 5) puts the second boom at “the 1980s–90s,” while Chapter 4 (page 235) puts it at “the 1980s” with “a winter again from around 1995.” Figure 4-2-1-5 gives the development of ELIZA as “1964” (the Ministry of Education, Culture, Sports, Science and Technology, 2024, says 1966. The year of ELIZA is confirmed in the u04 record).
What the figure shows:The white paper explained that expectations exceeded what AI could do, and the boom ended when that gap became clear. In the diagram, taking longer to notice lets expectations rise higher and then fall further.
Gap (expectation − what AI can do) and threshold
122017 · Checking a forecast
Testing the peak forecast against counts
How does the 2017 forecast of the boom's peak match the counts that followed?
In 2017, in the foreword to a special feature of the Japanese Society for Artificial Intelligence's journal, Koichi Takahashi and Tomohiro Inoue wrote that the third boom, as a boom in the literal sense, peaks in 2016 or 2017, and afterward will settle into steady industrial use. They predicted that the next one would be (translated) (general-purpose AI means AI that can do many kinds of work as people do).
In the figure on the right, choose a line and move "Year of checking". If you check in 2022, in Japanese book titles peaks in 2017 (38 books) and in the Diet in 2018 (194 speeches), a shape close to the forecast. But counting “AI”, the Diet reaches 593 in 2019, falls in 2020–2021, and in 2023 reaches 761, exceeding the earlier peak. The answer changes with which word you count and when you check.
What you think you knowThe popular story:The fourth AI boom is the boom of AI that makes text and pictures (generative AI) The original source says:In the 2017 journal of the Japanese Society for Artificial Intelligence, (translated) meant a boom still in the future, expected to be triggered by general-purpose AI. What a number points to can change later
Sources & confidence (2)Testing the peak forecast against counts
September 2017Verified in the originalJapan
In 2017, the (translated) Meant a Future Boom Triggered by the Arrival of General-Purpose AI (AI That Can Do Many Kinds of Work)
In the preface to the special feature (translated) in the journal of the Japanese Society for Artificial Intelligence, Koichi Takahashi and Tomohiro Inoue introduced Inoue's paper as one discussing (translated), a boom beyond the one at the time (general-purpose AI means AI that can do many kinds of work the way a person does). The same text also says that, as a boom in the literal sense, the boom of the time (the third) would peak in 2016 or 2017 and then settle into steady industrial work.
The popular story:People tend to think “the fourth AI boom = the boom in generative AI such as ChatGPT” The original source says:In the 2017 journal of the Japanese Society for Artificial Intelligence, the (translated) meant a future boom that had not yet come, one (translated). What a number points to can change later
What we still don’t know:When the peak of the third boom came depends on what is measured (number of papers, funding, number of articles, and so on). We have not yet checked with numbers whether the 2017 forecast (a peak in 2016–2017) was right.
1 sources and how they were checked
On the Special Feature “AI Society Theory” (Koichi Takahashi, Tomohiro Inoue · 2017 · p.641 Introductions to the papers of the special feature (E: Econ… · Original source)
January 2023Verified in the originalSources disagreeJapan
How the Japanese Society for Artificial Intelligence's “History Scroll of Artificial Intelligence” (2023) Divides the Booms and Winters
Over one year, the Japanese Society for Artificial Intelligence drew the history of AI as a picture scroll on the covers of its journal. Its division of time is: picture 1 = from the start of the field to the first boom (late 1950s–1960s, fading from the late 1960s as limits came into view); picture 2 = from the winter to the second boom; picture 3 = the second winter (around 1992 to around 2010); picture 4 = the first half of the third boom (around 2010–2016, triggered by the 2012 image recognition contest); picture 5 = the second half of the third boom (2017 on; (translated); the original uses a formal word for “by now”) through the generative AI boom; picture 6 = the future as drawn by researchers.
What we still don’t know:Even in the scroll, the (translated) is only put as (translated), and the society has not decided on it.
4 sources and how they were checked
Where sources disagree:On the start of the third boom, the scroll says “starting with the 2012 image recognition contest” and “from around 2010.” The Ministry of Internal Affairs and Communications (2016) says “2000s,” and the Ministry of Education, Culture, Sports, Science and Technology (2024) says “around 2010.” On the winter after the second boom, the scroll says “around 1992 to around 2010,” and the Ministry of Internal Affairs and Communications (2016) says “from around 1995.”
What the figure shows:A forecast written in 2017 that the peak would be in 2016 or 2017 can look right or wrong depending on which words are counted and when the result is checked.
131984 · AI winter
began as a warning
When did the phrase appear, and with what meaning?
An AI winter is the name for a time when hope shrinks and research funding and interest fall. The oldest record found is a panel discussion at the 1984 conference of the American Association for Artificial Intelligence (AAAI-84). The chair, Drew McDermott, gave a worst-case scenario in which, five years later, neither the Strategic Computing Initiative nor Japan's Fifth Generation had worked out, and no research funding at all would go to anything related to AI. He said, (translated). The scenario also includes everyone changing the name of their research to something else. If that happened, even while research continued, word counts would look smaller.
In the figure on the right, open "English book counts and sources" to see word counts in English books. AI winter appears little by little from 1983 and rises greatly in the late 2010s. Second AI winter first appears in 2000, and first AI winter in 2011.
English books number the winters. Calling the 1970s the is a later way of speaking. For second AI boom, third AI boom, and fourth AI boom, there was not enough volume to count.
What you think you knowThe popular story:People tend to think “The cut in research funding in the 1970s was called an ‘AI winter’ at the time” The original source says:The oldest record is from 1984, and it was the name of a worst case that might come
Sources & confidence (2) began as a warning
1984Verified in the original
1984: was spoken of as the name of the worst thing that might come
At a panel discussion at AAAI-84, the moderator, Drew McDermott, said that researchers were worried that expectations for AI were too high and that it would someday end in trouble. As the worst-case scenario he gave this picture: in five years the strategic computing initiative fails miserably, Japan's Fifth Generation also does not work out, the Japanese government quickly pulls out of computers, the start-up companies all go bankrupt, and no money goes to anything related to AI. He described this as . He urged them to warn themselves so that it would not come to that, and to tell the world. In other words, “The AI Winter” was a phrase that appeared in the middle of the second boom as a warning about the future. Calling the decline in research funding in the 1970s the is a label added afterward.
The popular story:It is often thought that the drop in research funding in the 1970s was called an at the time The original source says:The oldest record found is a 1984 discussion, where it appears as the name of the worst thing that might come in the future (the speaker said ). The (the 1970s) is a label added afterward. English-language books number the winters, as in , but is hardly used
What we still don’t know:Who first said is not known (even in the discussion it was “by some”). What book the small value for 1983 in the Ngram comes from has not been confirmed.
2 sources and how they were checked
The Dark Ages of AI: A Panel Discussion at AAAI-84 (Drew McDermott, M. Mitchell Waldrop, Roger Schank, B. Chandrasekaran, John McDermott · 1985 · p.122 Drew McDermott's opening remarks · Original source)
Calling booms by number, “first, second … boom”, is mostly a Japanese way of counting (English numbers the )
Word frequencies in English-language books (Google Books Ngram) show that appears from 1983 and rose sharply in the late 2010s. Expressions that put numbers on winters, such as and “second AI winter”, also exist. On the other hand, , and did not appear in countable amounts. In Japanese, “first artificial intelligence boom” (translated) can be seen in a 1982 magazine, (translated) in a 1985 book, and (translated) in the Diet in 1985. From 2014, the first, second and third AI booms were used in an academic society journal, and they also appear in a chapter title of a 2015 book by Yutaka Matsuo and in government white papers from 2016 on.
The popular story:People tend to think “first to fourth AI booms” is a worldwide way of counting The original source says:Numbered “booms” are mostly a Japanese expression. English-language books put numbers on the winters (“The AI Winter”: first/second AI winter) and do not number booms. The U.S. DARPA uses a different way of counting,
What we still don’t know:We have not looked at how this is counted in Chinese (such as phrases for a “third wave” of AI) or Korean. Ngram covers only books through 2022, so we do not know the English expressions from 2023 onward.
What the figure shows:The earliest record of “AI winter” is from 1984, when it named a worst-case situation that might lie ahead. Calling the 1970s the “first AI winter” came much later.
Years the words appeared, and the era they point to
141985–2026 · How it spread
Who said ?
When a famous story spreads, can the person who said it change along the way?
English Wikipedia writes that at the 1984 conference warned of an AI winter. The source it cites is Crevier's 1993 book.
But open the record of the discussion (AI Magazine, 1985) and the first person to bring up the words was the chair, Drew McDermott, and he introduced it as . The panel was five people: Waldrop, Ohlander, Schank, Chandrasekaran, and John McDermott. Minsky's name appears only once, as a participant in a different gathering.
The figure on the right lines up the names that appear in each source. Press "Compare with the list of panelists" and names not on the list get a mark. A famous person's name can creep in as a story is passed along in shorter form. Going back to the original record shows it.
This site has not yet opened Crevier's book. It remains possible that Minsky gave the same warning somewhere else.
What you think you knowThe popular story:“‘AI winter’ was a phrase Schank and Minsky warned of at the 1984 conference” The original source says:In the record, the person who brought up the words was the chair, Drew McDermott, and he introduced it as being called that by some. Minsky is not among the panelists
Sources & confidence (2)Who said ?
1984Verified in the originalSources disagree
The story that was a phrase Minsky and Schank warned about → In the record of the discussion, the person who brought up the phrase was the moderator, Drew McDermott
English Wikipedia says that at a 1984 conference warned of an AI winter (the source is Crevier's 1993 book). The record of the panel discussion (AI Magazine, 1985) shows that the person who first brought up the phrase was the moderator, Drew McDermott, who introduced it as . The panelists were Waldrop, Ohlander, Schank, Chandrasekaran, and John McDermott, and Minsky's name appears only once, as a participant in a different gathering (a seminar of Asimov's).
The popular story:The phrase is often said to have been warned of by Schank and Minsky at the 1984 conference The original source says:In the record, the person who brought up the phrase was the moderator, Drew McDermott, and he introduced it as . Minsky is not among the panelists
What we still don’t know:Crevier's book (1993) has not been opened. It remains possible that Minsky made the same warning at another event of the same conference, or on another occasion.
2 sources and how they were checked
Where sources disagree:Wikipedia (a statement sourced to Crevier 1993): Schank and Minsky warned. Record of the discussion: the person who brought up the phrase was the moderator, McDermott, and Minsky is not among the panelists.
1984: was spoken of as the name of the worst thing that might come
At a panel discussion at AAAI-84, the moderator, Drew McDermott, said that researchers were worried that expectations for AI were too high and that it would someday end in trouble. As the worst-case scenario he gave this picture: in five years the strategic computing initiative fails miserably, Japan's Fifth Generation also does not work out, the Japanese government quickly pulls out of computers, the start-up companies all go bankrupt, and no money goes to anything related to AI. He described this as . He urged them to warn themselves so that it would not come to that, and to tell the world. In other words, “The AI Winter” was a phrase that appeared in the middle of the second boom as a warning about the future. Calling the decline in research funding in the 1970s the is a label added afterward.
The popular story:It is often thought that the drop in research funding in the 1970s was called an at the time The original source says:The oldest record found is a 1984 discussion, where it appears as the name of the worst thing that might come in the future (the speaker said ). The (the 1970s) is a label added afterward. English-language books number the winters, as in , but is hardly used
What we still don’t know:Who first said is not known (even in the discussion it was “by some”). What book the small value for 1983 in the Ngram comes from has not been confirmed.
2 sources and how they were checked
The Dark Ages of AI: A Panel Discussion at AAAI-84 (Drew McDermott, M. Mitchell Waldrop, Roger Schank, B. Chandrasekaran, John McDermott · 1985 · p.122 Drew McDermott's opening remarks · Original source)
What the figure shows:In the original panel record, moderator McDermott introduced the words “AI Winter,” and Minsky was not a panelist. Famous names can enter a story as it is passed along.
152017 · DARPA
Three waves sort AI by how it's built
How did a US research agency count the history of AI?
John Launchbury of DARPA, the research agency of the US Department of Defense, explained AI in a 2017 slide deck as . The first wave is where engineers write rules to represent knowledge (Handcrafted Knowledge). The second wave builds statistical models and trains them on large amounts of data (Statistical Learning). In the third wave, the system itself builds models that explain situations (Contextual Adaptation). No division by years is given. The AI that learns from data, which was the center of the third boom in Japan, falls in DARPA's .
The figure on the right is the spiral example from the slides. Trying to separate the two arms with a rule set by people (distance from the center) does not work. Stretch the space and one line can separate them. The slides explain that a second-wave neural network separates them by stretching and squashing the space layer by layer.
(x, y) → (θ, r)
r
distance from the center
θ
angle turned (add 2π for each full turn). The way the figure stretches is this site's own choice. Second-wave AI also learns how to stretch from the data
What you think you knowThe popular story:The numbering of AI history (first, second...) is the same worldwide The original source says:DARPA's waves are types of how systems are built. Kai-Fu Lee's (2018) are four kinds of uses that advance at the same time
Sources & confidence (2)Three waves sort AI by how it's built
February 2017Verified in the original
The American DARPA Way of Counting: (2017) — Its Numbering Is Off by One from Japan's Booms
Slides by John Launchbury of DARPA, the research agency of the U.S. Department of Defense (given as U.S. Department of Defense on the DARPA page), explained AI as . The first wave = Handcrafted Knowledge (engineers write rules to represent knowledge). The second wave = Statistical Learning (engineers build statistical models and train them on large amounts of data). The third wave = Contextual Adaptation (the system itself builds models that explain situations). The slides sort AI by how the systems work and give no date boundaries. The AI that learns from data, which was the center of Japan's , corresponds in DARPA's scheme to .
The popular story:People tend to think “the numbering of AI history (first boom, second boom…) is the same all over the world” The original source says:The American DARPA sorts AI into “Three waves of AI,” and a (translated) is not a period of time but a difference in how the technology works. AI that learns from data is the in Japan but at DARPA
What we still don’t know:We have not watched what is said in the video (YouTube). The slides give no date boundaries (only the example of DARPA's robot car competitions in 2004 and 2005).
The English (2018, Kai-Fu Lee) Are Four Kinds of Uses Going On at the Same Time, Not Numbered Booms
In an article in Fortune, Kai-Fu Lee sorted the ways AI is used into : Internet A.I., business A.I., perception A.I. and autonomous A.I. The article says the four are . Even when English says , it points to something different from Japan's (translated).
The popular story:People tend to think “in English too, ‘4th wave of AI’ = the fourth boom” The original source says:Kai-Fu Lee's , which is often quoted, are not an order of eras but four kinds of uses going on at the same time
What the figure shows:DARPA's “Three waves of AI” divide AI by how it is built, rather than by date. Even spirals that human-written rules cannot separate can be split by 1 line if the space is stretched.
Bars of the slide's four forces (lengths read by eye by this site)
This is the same mechanism as "Try it", written as Python code. Press ▶ Run and the result appears below. Change the numbers and run it again.
u01_spiral.py
161995– · Hype cycle
Fitting a template of peaks and valleys
Does a new technology always follow “peak of hope → valley of disappointment → settling down”?
The research firm Gartner publishes a figure that shows how hope for new technologies changes, as a single curve (the hype cycle). In Gartner's 2019 view for Japan, artificial intelligence was placed in the “Trough of Disillusionment” stage. According to a paper that studied this figure, the curve is made by adding a peak-shaped curve that stands for people's excitement and an S-shaped curve that stands for a technology's maturity.
In the figure on the right, move the year of the peak and the year of the S. The heights are decided by calculation to fit the English-book data best. Fit to the 1975–2010 data alone, and the template overlaps well. But predict 2022 with that template and you get only about 1/8 of the real value. Overlapping past data and predicting the future are different things.
In 2016, Dedehayir and Steinert checked the figure against Gartner's own records and reported a mismatch between the records and the figure's template.
hope(t)＝A × peak curve(t)＋B × S curve(t)
A, B
heights of the curves (calculated by least squares to fit the data)
t
year
What you think you knowThe popular story:Every new technology follows the order of a peak of hope, a valley of disappointment, and settling down The original source says:Checked by research, it could not be said to follow the template
The five stages of the hype cycle
The paper explains the figure in five stages: the trigger, the peak of inflated expectations, the trough of disillusionment, the slope of enlightenment, and the plateau of productivity (this site translated the English names).
According to the paper, an ordinary technology is said to go around once in 5 to 8 years. But when the records were examined, one round took about 22 years on average. The paper also raises the question of whether two separate phenomena, excitement and technological maturity, can be added on the same vertical axis.
Sources & confidence (3)Fitting a template of peaks and valleys
2016Verified in the original
The pattern “expectations rise to a peak, pass through a valley of disappointment, and settle” did not fit exactly when checked by research
A study checked, using Gartner's own records, whether the research firm Gartner's chart of an expectations peak and valley (the Hype Cycle) really happens. It reports that there were mismatches between the actual records and the shape of the chart.
The popular story:“New technologies always follow ‘peak of expectations The original source says:valley of disappointment→settling down’” → when checked by research, it could not be said to follow the pattern exactly
What we still don’t know:“The same flow repeats each time” is a shape that appears in history, and cannot be called a fixed law. We do not know whether the fourth boom will follow the same shape
In 2019, a Japanese Research Firm Places Artificial Intelligence in the “Trough of Disillusionment”
In , published by Gartner Japan, artificial intelligence was placed in the “Trough of Disillusionment” (the period when disappointment spreads after expectations have grown too high). It explains that this was because trial uses had led people to face reality instead of expectations.
What we still don’t know:The shape of the Hype Cycle curve is the research firm's own view and has not been confirmed as a law.
1 October 2025Verified in the originalSources disagreeJapan
2025: in the same research firm's view for Japan, generative AI is at the (on the boundary with the trough of disillusionment in the figure)
The main text of Gartner Japan's wrote that AI agents and generative AI are (translated). Meanwhile, the generative AI dot in the figure sits on the line between the peak and the trough of disillusionment, and the headline of one news article reported (translated).
The popular story:News reports that “generative AI entered the trough of disillusionment in 2025” The original source says:The main text of the original announcement says “peak”, and in the figure it is on the boundary
What we still don’t know:Where the fourth boom is on the curve right now differs depending on who looks, and is not known.
2 sources and how they were checked
Where sources disagree:The announcement's main text says “peak”, the news headline says “Trough of Disillusionment”. The dot in the figure is on the boundary.
What the figure shows:The hype cycle shape fits past data well, but its prediction for 2022 is only about 1/8 of the actual value. Fitting the past and predicting the future are different tasks.
What is inside the template: the peak curve and the S curve
172026 · Weight of a timeline
Japan inside an English AI timeline
Which eras and which countries does an AI timeline write about in depth?
We counted the data of an English AI timeline you can touch and explore (by GLOBAÏA, a Canadian organization). Its introduction says , but the data has 539. Of these, 325 (about 60%) are events from 2012 onward. Events centered on Japan number 7 or 8 (about 1.5%). There is no indication of how certain each event is.
The figure on the right stacks the 539 by year. Move the slider to change the dividing year and it shows what percent of all events fall in or after that year. The purple dots are the years of Japan-centered events.
A timeline changes how history looks depending on what it chooses to include. Newer eras tend to have more events chosen. Please look at this site's timeline with the same eyes.
share from the dividing year on＝number of events from that year on ÷ total number of events
total number of events
539 (of which 45 are before 1940)
Years were counted as the integer part of the data's year value.
Sources & confidence (1)Japan inside an English AI timeline
14 September 2026Verified in the originalSources disagree
The earlier Canadian group GLOBAÏA's AI timeline has 539 entries; only 7–8 events are centered on Japan, and 60% are after 2012
We counted the data of an English-language AI timeline that you can touch and explore. There are 539 events (its introduction says ). Events centered on Japan number 7–8 (about 1.5%), and events after 2012 number 325 (about 60%). It does not show how certain each event is.
The popular story:The note from when this site was started said (translated) The original source says:the data has 539 events
2 sources and how they were checked
Where sources disagree:The page's introduction says 'more than a hundred curated events'. The data has 539 events
What the figure shows:About 60% of the 539 entries in an English AI timeline are from 2012 onward, and only 7–8 center on Japan. What a timeline includes changes how history looks.
181982–1995 · OCR text
Searching through OCR text
How can we search the text of old magazines, and is there a trap?
The phrase from 1982 was found in the full-text search of digitized books and magazines at the National Diet Library. The search looks through text that a machine read from photographs of the pages (OCR).
But machine reading makes mistakes. In Mr. Kaminuma's 1985 book, a place that reads (translated) in another book from the same year was garbled (〓列アーキテンチャー). The mark〓stands for a character that could not be read. If you search for exactly the same letters, this book is not found.
So we count how many characters must be fixed to make a match (edit distance) and search so that slightly different text still hits. In the figure on the right, increase the number of fixes allowed. More items are found, but different words start to hit too.
d(i, j)＝min{ d(i−1, j)＋1, d(i, j−1)＋1, d(i−1, j−1)＋ (1 if the characters differ)}
d(i, j)
the fewest fixes needed to match the first i characters of the search word with the first j characters of the sentence
Many of the sentences in the figure are examples made by this site (marked as made-up).
Sources & confidence (2)Searching through OCR text
March 1985Weakly confirmedJapan
A 1985 book has (translated) (the 82 is thought to mean 1982)
Mr. Kaminuma's book (1985) contains a chart-like listing of years and events, with entries such as (translated) and “82 Second Artificial Intelligence Boom: Fifth Generation Project” (translated). In the column (thought to mean 1982; this site's estimate), “second artificial intelligence boom” (translated) and (translated) are listed side by side. The text right after is garbled in the scanned reading, but in another book edited by Kaminuma the same year, the same sequence is followed by (translated). This is the oldest (translated) wording we have found.
What we still don’t know:We have not seen the whole chart (from what year to what year it runs, and what else is written in it). In the text data of what seems to be the same chart in , August 1986, the DNA double-helix model of Watson and Crick in '53 is followed by (translated). It seems to call the late 1950s the (translated; this site's estimate), but this needs to be checked on the printed page.
A 1982 magazine already has the “first artificial intelligence boom” (translated) ── the oldest numbered expression found
An article in the magazine (November 1982 issue) has a sentence saying that companies that had become disillusioned with the results of the “first artificial intelligence boom” (translated), which arose some twenty years earlier, had kept their interest in this field lost. So in 1982 the boom of about twenty years before (by this site's calculation, the early 1960s) was already being called the (translated). In a full-text search of books and magazines digitized by the National Diet Library, this is the oldest example found of an AI boom expression with a number (within what was found).
The popular story:It is often thought that numbers like first boom and second boom were added afterward (in the 2010s) The original source says:A 1982 magazine already says “first artificial intelligence boom” (translated), so booms were being called by number from the middle of the second boom
What we still don’t know:I have not seen the article's title, author, or pages (library use only). Because it is OCR (text read by machine), it needs to be checked against the printed page.
1 sources and how they were checked
President 20(11) (1982 · Approximate frame number in the National Diet Library Digit… · Original source · Could not be opened)
What the figure shows:Errors in machine-read text from old documents prevent exact searches from finding it. Searching by how many letters need correcting can find it, but allowing too many changes also finds unrelated words.
19This site's way of reading · Confidence
How the color labels were decided
What do the colored labels (confidence grades) on this site's events mean?
Each event on this site carries a label for how far it has been checked. Verified in the original source means we opened the original source and checked down to the page or paragraph. Confirmed in a secondary summary means we could not open the original source but confirmed the same content in two or more reliable summaries. Weakly confirmed means there is only one summary, or a disagreement between sources is not settled. Anything from memory alone or guesswork alone is not listed as an event.
The figure on the right is the path that decides the label. Press this chapter's events below and the route that led to that color lights up. Of this chapter's 23 events, 18 are verified in the original source. The 5 events with the numbered phrases from 1982, 1985, 1993, and 1995 could not be checked against photographs of the pages and were confirmed only by text extracted by machine reading, so they are marked weakly confirmed.
Where sources differ on a year or a way of saying something, we add a “Sources disagree” label separate from the color, and write both without choosing one.
Sources & confidence (2)How the color labels were decided
November 1982Weakly confirmedJapan
A 1982 magazine already has the “first artificial intelligence boom” (translated) ── the oldest numbered expression found
An article in the magazine (November 1982 issue) has a sentence saying that companies that had become disillusioned with the results of the “first artificial intelligence boom” (translated), which arose some twenty years earlier, had kept their interest in this field lost. So in 1982 the boom of about twenty years before (by this site's calculation, the early 1960s) was already being called the (translated). In a full-text search of books and magazines digitized by the National Diet Library, this is the oldest example found of an AI boom expression with a number (within what was found).
The popular story:It is often thought that numbers like first boom and second boom were added afterward (in the 2010s) The original source says:A 1982 magazine already says “first artificial intelligence boom” (translated), so booms were being called by number from the middle of the second boom
What we still don’t know:I have not seen the article's title, author, or pages (library use only). Because it is OCR (text read by machine), it needs to be checked against the printed page.
1 sources and how they were checked
President 20(11) (1982 · Approximate frame number in the National Diet Library Digit… · Original source · Could not be opened)
1950–2022Verified in the original
English books: in the 1950s, from the 1960s. The 1987 peak was passed in 2018
We counted year by year how often particular words were used in English books. was most frequent in 1956, and around 1962 was overtaken by . Artificial intelligence peaked in 1987 (about 2.0 times per million words) and fell in the 2000s (about 0.72 in 2007). It then passed the 1987 peak in 2018 (about 2.3) and reached about 6.4 in 2022. In 1965, called the , it was only about 0.11, about 1/18 of the 1987 peak.
The popular story:“The 1960s were the golden age of AI” The original source says:Looking at how often the term appears in English books, 1965 is only about 1/18 of the 1987 peak. What was used a lot in the 1950s was a different term, .
What we still don’t know:The frequency of a word is not the size of the boom itself (names change, and the number of books differs from year to year).
What the figure shows:The site's colored labels show how far each original source has been opened and checked. Of this chapter's 23 entries, 18 were checked against originals, while 5 rely only on machine-read text and have weaker confirmation.
Closing
What this chapter still doesn't know
The number of times a word appears is not the size of the boom itself. When the name changes (electronic brain → artificial intelligence → AI → generative AI), the same way of counting can no longer see it. The number of books and papers itself also grows year by year. The two letters “AI” also become part of other words (AIDS, AIG, people's names), so we looked at each case one by one, made a standard, and recounted. We cannot pin down a single reason why “AI” in the Diet fell in 2020–2021 (it overlaps with the pandemic years).
The history of the numbering also leaves open questions. Who began using the phrase (translated) before 1985? Who first said ? How are booms counted in Chinese or Korean? English book counts run only to 2022, so we also don't know how English wording changed after that.
How many booms there were changes with the standard you count by. In the next chapter we go back to before the name was born, and begin with the dream of thinking machines and the history of calculating machines.