The Chisel-Construct Cycle · AI Series · 06

How “Artificial Intelligence” Got Its Name

Han Qin · About 28 minutes · Published

1. A Request for Money

By 1955, this whole business had already gone by a dozen names or so.

The people making machines called it automatic computing systems. The people doing theory called it cybernetics, automata studies, or nerve nets. The newspapers called it electronic brains, mechanical brains, or thinking machines.

On August 31, 1955, another name appeared. It was printed in the title of a proposal asking a foundation for money.

The document was titled “A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence.” It ran to seventeen pages. The first five bore four names. Behind them were the four men's qualifications and the subjects each intended to study.

The way the pages are divided tells you what kind of document it is. Only five pages belong to all four men. In the other twelve each writes for himself. It is not a manifesto. It is a file assembled out of separate pieces. In the first five pages four people agree on what they will do together. Then each says what he plans to do on his own. Of the four, only one would appear in the record day after day that summer: the man who wrote the proposal.

Here are the four signers and the positions the proposal gives them.

John McCarthy, assistant professor of mathematics at Dartmouth College. Marvin Minsky, junior fellow in mathematics and neurology at Harvard University. Nathaniel Rochester, manager of information research at IBM. Claude Shannon, mathematician at Bell Telephone Laboratories.

Shannon has appeared once in this series. He wrote the 1950 paper on programming a computer to play chess. The third reason for choosing chess, in the first section of his paper, is that people generally think it takes thinking to play well.

Two of the four were at companies and two at universities. The proposal explains that the companies could cover the expenses of their own people. It was the university side that needed the foundation to pay.

This division would recur. Half the people doing the work drew company salaries, half drew academic ones, and the half that needed somebody to pay extra was often the academic half.

They asked for thirteen thousand five hundred dollars. They envisioned ten people staying two months.

The proposal's own term is summer research project. Research, not a conference. It was meant to bring a group to one place for a whole summer, not to convene a formal gathering for a few days.

It is a document from before the event. It was written before the summer to get the money approved and give a group of people a reason to stay in the same place for a while.

Later accounts have described this paper as the proclamation of a new discipline or the birth certificate of a field. There is no such sentence in the original. It does not define artificial intelligence as a discipline, call it a new science, or announce the birth of anything. An event is usually recognized as the birth of a thing after the thing is grown, and the people doing the recognizing mostly were not there.

And it is in the title of just such a document, written before the event, that the phrase artificial intelligence appears.

It is in the title, the opening paragraph, the introduction to the seven research topics, and the later pages on arrangements. The four men use it to describe a project, a line of research, and a problem.

Nowhere do they say, “We call this such and such,” or, “We are coining a word.”

A name came into use, without an announcement.

It is used as the name of something that already exists: a summer research project on artificial intelligence, the problem of artificial intelligence. It reads as if it referred to something already under way. And this piece of paper from August 31, 1955, is an early, verifiable instance of the name being used for a line of research.

They did not get as much money as they asked for.

On November 30, 1955, Robert Morison of the Rockefeller Foundation wrote to McCarthy. In response to the request for thirteen thousand five hundred dollars and two months, he offered seventy-five hundred dollars for something closer to five weeks.

The money fell by almost half, and so did the time. Ten people for two months was no longer on the table. The grant mattered because it went a long way toward determining what the summer would actually be. What took place was looser still than the grant provided for.

2. The Conjecture

A sentence in the proposal's first paragraph has been quoted a great deal. It is the first time in this series that the possibility of simulating any given feature in principle has been put down as the starting point of a piece of research.

The sentence says that every aspect of learning or other feature of intelligence can in principle be described precisely enough to make a machine that simulates it.

The machine is not the subject of the sentence. Its subjects are those features of learning and intelligence, and what the sentence wants to happen to them is for someone to describe them. The machine comes in very late, after the description.

In its place in the document, the sentence has two layers of qualification. Both are easy to lose when quoting it.

The first sits outside the sentence. The original says beforehand that the study will proceed on the basis of a conjecture. It is a conjecture, not an established result and still less a report on something already made. The conjecture comes first; the research comes after.

The second is inside the sentence: in principle. And describing a thing precisely enough to simulate it is a matter of degree. What counts as enough precision depends on whether the thing can be simulated. There is no independent standard of precision already on the table.

The sentence thus has two stages. First the thing must admit a precise enough description. Then a machine is made from that description. A so that joins the two. All the weight is on the join. How precise is precise enough? Precise enough for a machine to take it on.

Strip those two layers away and you are left with an assertion: a machine can simulate all intelligence. Put both back and it is a working assumption. Suppose for now that it can be done. Then get to work.

It is ordinary enough to find a working assumption in a request for money. The people asking have to explain what they mean to try, not guarantee they will succeed. The people who wrote this one could hardly guarantee that. What they brought that summer was a few papers and an account of a program still being implemented on a computer. The claim they meant to put to work covered learning and intelligence as a whole. A conjecture bridged the gap. Requests for money often have that shape: little in hand, a large thing to say, and an in principle in between. The shape has been used ever since.

Set the sentence beside the earlier essays.

In 1843, a person drew a boundary for a machine that had not been built. It lay at the act of beginning: the machine had no pretensions to originate anything.

In 1887, a person listed two inherent incapacities of such devices: they could not find problems of their own, and they could not decide for themselves which of several routes to take.

In 1936, a boundary was proved. No one could move it.

In 1948, at the end of an internal report, someone wrote that once a thing was understood, its behavior could be predicted, and the method behind it was clear, it no longer looked intelligent.

In 1950, someone replaced a question while expressly saying its original form could not be abandoned altogether.

The five draw different boundaries. One concerns work a person gives a machine, another a theorem, another a question deliberately kept on the table when the wording changes. To put them side by side is not to say they are the same.

This time is different. The sentence says that the features of intelligence can in principle be described in a way that lets a machine simulate them.

It does not say they have been simulated. It does not say the ability to simulate is the same thing. These are claims quite far apart. To say you have done something, bring evidence. To say it is worth trying, you can start with a conjecture.

But when it is written as though it has already been done, everything else starts to look like engineering: a matter of time, computing power, and data. Over the next seventy years that kind of sentence recurs.

This sentence brings into research things previously taken to lie outside the boundary. At first it puts them into a conjecture. What cannot be simulated, or what remainder there might be once the simulation is done, is not answered in the proposal.

Four people wrote it to get money approved.

For the next seventy years the sentence was quoted again and again, still mostly in places where people were asking for money: proposals, plans. The qualifiers generally do not make the trip.

3. Seven Topics

The proposal lists seven numbered research topics. Here is the list as it stands.

First: automatic computers.

The second is not a noun but a question: how can a computer be programmed to use a language?

Third: neuron nets.

Fourth: the theory of the size of a calculation.

Fifth: self-improvement.

Sixth: abstraction.

Seventh: randomness and creativity.

Four items have to do with how to make the machine, how to carry out a calculation, and how large it might be. The other three have to do with what the machine should know how to do. The two kinds of thing appear side by side on one list, numbered in sequence.

The first is automatic computers. By 1955 such machines already existed, but the machines themselves still had to be listed as something to study. What a machine that could go on calculating by itself might do, and how to go about using it, were still open questions.

The second item later acquired a stable name as a discipline. In 1955 it was still a question. The third was then written as neuron nets rather than the expressions current later. The fifth said self-improvement, not the name the work would eventually come under. The fourth asks how big a job a calculation is. The question grew into a field of its own; here it is just number four. The seventh puts randomness and creativity together. That combination too came apart later, and the two words went their separate ways.

The gap between names and work is visible at a glance. People are ready to do the work, but the lines of research in which they will do it do not yet have names of their own. At least four of the seven now have their own conferences and journals. In 1955 they were four of the seven numbered lines on a piece of paper.

Something else is visible when the seven are laid out together. They are not a catalog of the branches of a discipline. They are a list made out of things four people separately wanted to do. Neuron nets came from Rochester's line of work. The size of a calculation came from Shannon's. Getting a machine to use a language and to improve itself came from McCarthy and Minsky. They were put on one piece of paper, numbered, and became seven topics.

The proposal also sets out the arrangements, in a loose way. Circulate the proposal to people who might take part and have them say what they want to work on. The organizers then issue invitations. During the summer there will be regular research discussions and informal work by individuals or small groups.

It is not a day-by-day conference schedule. It is a way of giving some people a reason to stay in the same place.

What would take place that summer was not fixed even on paper. Two things were fixed: who would pay and what it was called.

There is something else the paper does not do.

In telling the funders what they intended, the four do not say they will get machines to think like people or copy the brain. The jobs they list are concrete: make machines use language, form abstractions, solve problems of a kind then left to people, and improve themselves.

On the third point, they do not say they will solve problems people cannot solve. They say problems still left to people. Left to makes the point clear. This is work that can be handed over. How much can be handed over is not yet known.

For the request they used the plainest terms. In the title they used the larger word.

4. That Summer

The place was Hanover, New Hampshire, home of Dartmouth College.

The records left from the time do not say on what day the project opened or closed, or who its official members were.

A weekly report by Trenchard More labels two weeks in the middle and latter part of July as the fifth and sixth weeks. Counting back puts the first around June 18. Ray Solomonoff's notes begin on June 22 and run into late August. People usually say six weeks. The notes give a stretch of time with frayed edges: a start in the middle or latter part of June, and somebody still putting things on paper in late August.

More's report lists who was there. From July 16 to 20, a whole week, it lists only four people: Etter, McCarthy, More, and Solomonoff. On July 23, five: Ashby, McCarthy, Minsky, More, and Solomonoff. On July 24 there were seven, with McCulloch and one other person added. On July 25 there were four. On July 26, five. On July 27, three: McCarthy, More, and Robinson.

Over the two weeks there were three to seven people there on any day. The names changed day to day. It is the one stretch of the summer with head counts by day.

McCulloch has appeared before too. He wrote that paper on nerve nets with Pitts in 1943. The paper itself lays out a string of qualifications, one of them being that formal equivalence of this kind does not explain what actually happens.

Solomonoff has a separate handwritten page headed with the people at the summer research project. It lists twenty names. It does not say they were official members or say when they arrived or left. It is a list of who had come by, in a notebook among formulas and calculations.

Some names have passed through the series already: Shannon, McCulloch, Ashby. Others will appear in the essays to come, among them the people who made the checkers program and the system for proving logical theorems that year.

Historians have since reconstructed a list of ten official participants from foundation records and other material. The lists are not the same kind of thing. One is a list somebody made in 1956 as people came through. The other was put together from archives half a century later to say who counted as a formal participant. It is not a roster from the time.

One entry in More's report says that Ashby had just finished a book manuscript called An Introduction to Cybernetics. The report goes on to record his discussion of the homeostat, random numbers, and problem solving.

The new name was already in print a year earlier on the proposal. In the room that summer the old name was alive. It had arrived with a book just finished.

His homeostat came up earlier: four coupled electromechanical units that rearranged their parameters after a disturbance to find a new stable configuration.

There are no official daily minutes and no complete schedule. There are Solomonoff's private notes, More's report on weeks five and six, papers and working drafts by individuals, and correspondence. More opens by explaining that he will go by name through what the people who were there that week were doing. These are separate reports of work, not minutes of a joint meeting.

Years later McCarthy recalled why things worked out this way. The plan for everyone to come and stay the whole summer did not happen. There was less money than they asked for, and people did not want to stay in Hanover all summer. A lot came only for a few days. He includes himself, Minsky, Solomonoff, and Newell and Simon among the ones who stayed longer or had something more developed in hand.

There is another string of dates and names among Solomonoff's notes: June 22; material for a talk on July 10; Ashby on July 23; More on July 24; Samuel on August 6; McCarthy on August 7; Bernstein on August 8; Minsky on August 13; Bigelow on August 15.

The record looks like one person staying in one place, hearing something from someone every few days and jotting it down. It was not one meeting packed after another. There was a stretch of loose days, with a talk now and then and often just three or four people listening.

5. A Machine Not Yet Running

Among the things brought to Hanover that summer was a paper.

Its title was “The Logic Theory Machine: A Complex Information Processing System.” The names on it were Allen Newell and Herbert Simon. A line at the foot of the page thanks Shaw as a partner in the work and says he is responsible for putting it on a computer.

The job was this. In the early twentieth century Whitehead and Russell wrote a large work in three volumes, Principia Mathematica. It set out to derive mathematics entirely from logic: begin with a few logical rules, without assuming what number is, and derive arithmetic step by step. The book took ten years to write. Its propositions are numbered so that what has been proved earlier can be used later.

To make the book the subject was to pick a job already done, with each step on the page. You could open the book to see whether things agreed. The choice followed the same line as Shannon's first reason for picking chess six years earlier: the permitted moves and the final aim were clear. His third reason was that people usually thought the task took thinking. Proofs of theorems were then in that category too. Six years later the task picked had changed, but not the reasons for picking it. The program was to find proofs of those numbered propositions on its own.

What it did was not arithmetic. It worked on strings of signs it could substitute for one another or combine by rule. Until then, this work had been done by people and counted as part of reasoning.

It did not write out every possible expression until it happened on the target. It searched selectively. Try matching the target against the axioms and expressions already proved. If that fails, use an inference rule to join two expressions already in hand. If that fails too, break the target down into a smaller problem, solve it, then put the pieces together again. Go first down the route that looks nearer the goal.

These procedures have something in common. Every one is something a person might do in a proof. The approach does not play to the machine's strengths. It copies the way people do it.

Formulas are held as symbolic structures and lists in two places, working memory and stored memory. To do something with a formula, take it out of storage and into working memory. Once the work is finished, write it back.

To keep the paths from multiplying, the program sets limits on how much they can branch and how much working memory it can use.

With no limits, the paths would grow into a thicket. With limits, some proofs that could have been found will not be. What the program gets in return is not a guarantee. It gets a chance to finish. The field would make this trade many more times.

When it cannot find a proof, its answer is that this one has not been proved. That is not a proof that it cannot be proved. The whole of the third essay lies between the two sentences. In one, the machine failed to find something. In the other, the thing cannot be done. One takes a failed search; the other takes a proof. The program does the first, with theory in its name. The people who built it knew the distinction and put it in the paper. As in the first two essays, the person who says where the boundary lies is often the one closest to the machine.

The paper brought to Hanover states a qualification of its own.

The opening says that it deals with the specifications of the system, not its implementation on a computer. Later it adds that the language in the paper has not yet been coded for the computer, though a very similar version is being coded.

Coding here means turning a procedure already written out into instructions that a machine recognizes. At the time somebody had to do this one line at a time.

That tells us how far the document they brought had got. It does not tell us that there was no computer run all summer. In a letter to Russell that October, Simon says the program had yielded its first proof about two months before. The letter does not say it ran in Hanover.

One familiar account has the Logic Theorist demonstrated at Dartmouth. The material from the time used here does not give a record of a complete on-site computer demonstration. The nearest thing is the very paper that says implementation is still in progress.

In later accounts the familiar picture is of somebody at a screen pressing Enter and a machine proving a theorem on the spot. Simon himself later described a different scene. Around Christmas 1955, at a crucial stage, they simulated the program by hand. Several people held cards, each playing a part of the program, passing its state among themselves.

The results appear in the following year's paper. The program found proofs for thirty-eight of the first fifty-two theorems in the section of Principia Mathematica on propositional logic. For one it produced a more concise proof than the book's. Shaw's name was added. Two names on the account of how to do it; three on the account of what it did.

As for Russell's reply: on October 2, 1956, Simon sent the report on the Logic Theory Machine to Russell. Russell replied on November 2 that he was delighted. If he and Whitehead had known of the method earlier, he said, it could have spared them a great deal of work. The texts of the letters were later printed in Simon's autobiography. Both have entries in the archives. Russell received a report. The letter is his reaction to hearing about the result, not his verification of the program.

The distinction has come up before in the series. In the first essay, Babbage's autobiography and the correspondence of 1843 can each establish something. What one source can establish cannot simply all be carried over to the other.

Another story often told is that they made the program a coauthor of a paper, sent it to a journal of symbolic logic, and had it rejected. In the usual version it was rejected because the journal would not recognize a machine as an author. An earlier account gives a different reason: the editor thought a new proof of an elementary theorem was not important enough in itself.

The reasons are a long way apart. One is about whether a machine can count as an author, an argument about a boundary. The other is about whether this theorem is worth another paper, an ordinary editorial decision. Two ways to tell the same event. One has a plot. One does not.

The name the program went by in 1956 was not the one it had later. The terms in the paper are Logic Theory Machine and complex information processing system.

In January 1956, Simon announced it in a graduate class at Carnegie Institute of Technology. Over Christmas, he said, Al Newell and I invented a thinking machine.

A thinking machine. The words are not those on the proposal. It had been four months since the proposal was drafted, and for those four months the new name had stayed on a piece of paper asking for money.

6. The Name Did Not Catch On at Once

After that summer, the new name came into use unevenly.

A name takes hold not because someone makes it up but because other people start using it. In 1956 some of them did.

There were people who used it. On July 31, 1956, More's report was already headed “Dartmouth Conference on Artificial Intelligence: Report on the Fifth and Sixth Weeks.” He had also changed the summer research project of the proposal into a conference. Someone was writing down the name that would prevail as early as July 1956. The people taking part had changed the name while the thing was going on. Change research to conference and it sounds more like an event that has taken place. On the busiest day in the report there were still only seven people in the room. In December 1956, Minsky titled a draft at Lincoln Laboratory “Heuristic Aspects of the Artificial Intelligence Problem.”

There were people who did not use it too.

When Newell and Simon published their paper in September of the same year, its title was still the one they had brought to Hanover, the two parts about the Logic Theory Machine and a complex information processing system. Complex information processing was still how they described the work in the text. Shannon's paper that month still spoke the language of information theory. It was called “The Zero Error Capacity of a Noisy Channel.”

From September 10 to 12 there was a meeting at MIT formally called the 1956 Symposium on Information Theory. The roster has many of the names from the summer: McCarthy, Minsky, Rochester, Shannon, Newell, Simon, Selfridge, Solomonoff.

Much the same people, a little over three weeks later, in a different setting. The name of the meeting was still information theory.

There was no general name change in 1955 or 1956. Some people used the new word, some went on using the old ones, some used neither.

No vote on the name has turned up. A story often heard is that everybody at Dartmouth formally settled on the name of the field. The written material used here contains no record of a vote or a resolution. The phrase was already on the proposal asking for money in 1955, and some participants took it up in 1956. In the fall of 1956 several names were still in use at once.

Of the names that came before it, cybernetics had a book that had sold twenty thousand copies, automata studies had a publication McCarthy himself had helped edit, and electronic brains had the newspapers. Artificial intelligence had a proposal and a few people who went on using the phrase. It eventually became the usual name. At first nobody announced it. People used it. It was not settled at a meeting. It was written into a budget.

7. Why Make Up the Name?

The paper from 1955 does not explain where the name came from.

All the explanations we have are recollections by the man himself, years later. The later the recollection, the smoother the telling.

In an interview in 1979, McCarthy credited himself with the term while leaving a little room. He said he would not swear he had never seen it before.

The first part holds something back; the second says only what he can vouch for. He had not seen it, not that nobody had ever said it.

In an oral history on June 8, 2011, he said it again. As far as he knew, he had come up with the phrase while writing the proposal. He could not prove beyond doubt that he had never heard the words put together before.

Thirty-two years passed. The reservation stayed the same.

The same account offers two more points. He and Shannon had previously edited a publication called Automata Studies. He remembered Shannon thinking it would be too much to put a phrase like machine intelligence straight into a title. He also said that a name people used later, machine intelligence, was actually better in his view.

Fifty-five years on, the man who made up the name says he prefers the one people used later. The sentence is not often quoted. As he recalled it, the name came out of a choice between phrases, and one thing that mattered was that somebody thought the other one went too far.

The reason people quote most often comes from 1996. One reason for coming up with artificial intelligence was to get away from its association with cybernetics. The rest of the passage says that he thought cybernetics concentrated too heavily on analog feedback. He did not want to have to take Wiener as an authority or spend his time arguing with him.

Two things are there. One is a difference in technique. The other is not wanting to get into an argument. He writes the second alongside the first and says it plainly.

In the sources found here, this explanation can only be traced back to the publication of 1996. Whether he said the same thing when he made up the name is a different question.

In the oral history of 2011 he described the technical difference more concretely. Wiener's work, he thought, centered on the mathematical treatment of feedback. He wanted to go another way: start from facts and arrive at goals through logical inference.

For the next two or three decades the two routes had different centers of gravity. They were never cut off from each other altogether.

Wiener appeared earlier in the series. In the introduction to a book in 1948, he said the field had no common terminology, not even a name of its own. So he made one out of the Greek word for a steersman.

Seven years later, somebody made another so as not to have to use the first.

There is a neater story. Wiener declined to take part in the summer project, so McCarthy deliberately made up another name. It reads like a private feud. It is not the same thing as what McCarthy said.

It is far easier to tell the story of a name as a bout of personal pique than as a parting of ways over technique. It is easier to remember too. The version that is easier to remember travels farther. It has happened several times in this essay: the live demonstration, the rejection of a machine as an author, and now this.

Seen this way, the first boundary the name draws is not between the machine and the person. It is between two groups of researchers.

Yet in the room in Hanover that summer, the group supposedly outside the line was still inside, in the person of a man with a book he had just finished. The report records what he said alongside what everybody else said, without ranking them. The line in the name had not yet come down in that room. It would take more years to become a real line in lists of grants and conferences.

8. What the Summer Did Not Produce

What was left at the end of that summer?

The papers from the time do not contain the following things.

No common declaration. No statement of principles signed by all the participants or by a fixed group of them. The one programmatic statement signed by the four was still the document prepared to ask for money.

No common definition. The nearest thing is a note McCarthy wrote in September 1956 beside a list he kept. It defines the problem for the present purpose as getting a machine to behave in such a way that the behavior would be called intelligent if a person did it. The definition begins with a qualification: for the present purpose.

The note does not say what intelligence is. It says what kind of behavior will be called intelligent. The thing under discussion has moved from the machine's side to the side of the person speaking. The last section of the manuscript in the drawer from 1948 says the same thing. Those pages were still in the drawer. It would take twelve more years for them to appear in print.

No joint research plan. The organizational scheme in the proposal was a plan from before the event. No new version was adopted in the summer of 1956.

No joint report afterward. The 1955 proposal anticipated something written that could be sent to interested people. Nothing of that kind came together.

Nor did a common line of work emerge. More's report records Ashby's newly completed manuscript on cybernetics. Solomonoff wrote down his own ideas about induction. Newell and Simon brought a system for proving logical theorems. Rochester was working on neural models.

A familiar later account has Dartmouth committing itself to a symbolic approach and shutting out the other one. No such decision is visible in the material from that summer. What is visible is four or five kinds of work in one room at the same time, everyone writing his own thing, nobody signing on to anybody else's.

This is the shape of the written remains: one joint proposal from before the summer, and a pile of working papers by individuals and groups, personal notes, progress reports, and letters.

Many later accounts have taken the paper from before the event for the conclusions afterward. A proposal was read as a result.

What is left is a list of absences: an opening date, a closing date, an official roster, a recording, a photograph with a date on it.

A name came first. What it named did not yet have a common definition, method, or result that summer.

The gap between a thing and its name has appeared several times in the earlier essays. Each time the name came later. A machine is made, then somebody names it years afterward. People do a job, then years later there is a name for that line of work. Somebody wrote Turing machine in a review the following year. Cybernetics was made up after related discussions and several formal meetings.

This time it went the other way. The name came first, by a whole year.

For the next seventy years the scope of the name kept changing. When it was made, what fell under it was a piece of paper asking for money, seventy-five hundred dollars in funding, and a room with three to seven people coming and going on a given day.

The paper said: in principle, it can be done.

As for what remained outside the boundary, nobody counted it that summer. Part of what remained was sitting in the same room with a manuscript it had just finished. Its name was cybernetics.

The machines are still being built. Each generation says again where the boundary lies.