Will Robots Talk in Symbols? What Will They Say?

Machines have learned our words. Whether they can mean anything by them turns

on a question a California writer started asking a quarter century ago.

B Y L A R R Y M A C D O N A L D

__________________

Type a small red heart at the end of a message to a chatbot and, often as not, a heart comes

back. The exchange takes about a second. It may also be one of the stranger events in the

history of communication. A shape that people have carved into tree bark and stitched onto

pillows for centuries has just been handed to you by something with no pulse to quicken.

Did the machine use a symbol, or did it only return one?

That sounds like a seminar question. It is turning into a kitchen-table one. Machines now draft

our condolence notes, our wedding toasts, and our apologies, which are the kinds of speech

most heavily freighted with symbols. Robots with arms and legs are moving out of the

laboratory and into the warehouse, and their makers promise the living room is next. If these

things are going to talk to us for the rest of our lives, we ought to know what kind of talking it

is.

We have been here before, and we handled it badly. In 1966, the M.I.T. computer scientist

Joseph Weizenbaum wrote a program called ELIZA that did little more than turn a typist’s

statements back into questions. People confided in it anyway. Weizenbaum spent much of his

remaining career alarmed by how little it took.

◆

In 2001, a writer named John Fraim was posting essays to a Web site called Symbolism.org from

Northern California, where he had moved after what he describes as a narrow escape from the

entertainment business in Los Angeles. The essays are about the dot-com crash, the contested

Presidential election of 2000, the sameness of suburbs, and Marshall McLuhan. None of them is

about robots. Read today, they amount to a working theory of what a talking machine can and

cannot do.

Fraim starts with a definition. The usual one says that a symbol is something that refers to

something else. He tightens it: a symbol is something made by culture that points at something

outside culture. And the only thing outside culture, he argues, is nature. Spring. The ocean.

Night. Birth and death.

1Hold that definition up to a language model and a problem comes into view. The model is built

from text, and text is culture’s output. The machine has read more about spring than any

person ever will, and it has never once been cold. When it writes “spring,” the word points to

every other place the word has appeared. It points at the library. It does not point at the thaw.

The cognitive scientist Stevan Harnad named this predicament in 1990. He called it the symbol-

grounding problem, and compared it to trying to learn Chinese with nothing but a Chinese-to-

Chinese dictionary. You look up one squiggle and find more squiggles. You can ride that merry-

go-round forever and never touch the ground.

Fraim liked to quote a line of McLuhan’s: we don’t know who discovered water, but we can be

fairly sure it wasn’t a fish. A language model is the purest fish yet bred. It swims in our words

and has no shore to stand on. So the first answer to the title question is a deflating one. A

chatbot talks in symbols the way a player piano plays Chopin. The notes are all there.

◆

A robot is a different animal, because a robot has a body. It can bump into the table. It can run

low on power. And the body, it turns out, is where symbols begin.

One of Fraim’s shortest essays, “The Shape of Sequence,” draws on the work of Dan Acuff, a

consultant to toy companies. Acuff reported that children as young as eighteen months read

jagged lines as the bad guys and rounded lines as the good guys. That, he pointed out, is why

Mickey Mouse is assembled almost entirely out of circles. The German psychologist Wolfgang

Köhler had found something similar in adults back in 1929: shown a spiky shape and a blobby

one, and offered two made-up words, people agree with eerie consistency about which name

belongs to which.

Meaning, in other words, starts before language. It starts in the hand and the eye, with the

sense of what could hurt you and what will hold you.

Now picture the robots that companies want to put in your house. Look at the helpers in the

product videos, the vacuum under your couch, the speaker on your counter. They are

overwhelmingly smooth and round, with large dark faces and no teeth. Before a domestic robot

utters a word, its designers have made it say something to the eighteen-month-old who still

lives in each of us. It says: I am safe.

That is a symbol in the full sense. It points past culture to the body’s oldest knowledge. But

notice who is doing the talking. The robot is the message. The speaker is a design team.

◆

What will they say once they do use words? Fraim’s essays suggest three answers, and the first

comes from a piece of junk mail.

2In 1975, a freelance copywriter named Martin Conroy wrote a two-page letter to sell

subscriptions to the Wall Street Journal. It told of two young men who graduate from the same

college, join the same company, and meet again at their twenty-fifth reunion, one a department

manager and the other the president. The Journal mailed that letter for decades. By the figures

Fraim cites from the trade press, it brought in something like a billion dollars.

Fraim’s reading of it is the most useful thing in his archive. The content of the letter, he says, is

a sales pitch. The power is in the context. The story opens on a late-spring afternoon, at a

graduation. By its fifth word the reader is standing in the season of beginnings, at the modern

version of an initiation rite. Nobody notices. Everybody feels it.

Conroy found that formula once, by instinct, and it lasted a generation. A machine trained on

everything that ever moved us can find it on demand, for each of us separately, every morning.

The first thing robots will say, then, is whatever they are asked to sell, wrapped in spring.

◆

The second answer is that they may not need words at all.

In an essay called “The Democracy of Electric Action,” Fraim observed that actions had never

formed a language the way words had, and that electricity was changing that. Every click, card

swipe, grocery scan, and channel change was landing in a database somewhere. People were

voting hundreds of times a day without knowing it. Then he asked a question that reads

differently now than it did in 2001. Jung had built his idea of a collective unconscious on the

dreams of a handful of patients. What if the electric activity of a whole town were treated as the

mind of one patient? Or a whole country?

He was describing, near enough, the raw material that today’s systems are made from. And he

saw the fork in the road. That record of our actions could give a voice to people with no talent

for loud words. It had mostly been used, he noted, for persuasion and control.

A machine that reads us through our actions can answer in kind. It can speak through what it

shows us first and what it leaves out. That is talking in context rather than content, and it is the

kind of speech we are worst at hearing. Fraim had a phrase for our condition: the “trance of

content.”

◆

The third answer is the one that should keep an editor up at night.

Fraim borrowed an old philosophical sequence, the one, the two, and the many, and laid it over

American life. First came the mass culture of mid-century, with three television networks and a

few brands everyone shared. Then the Cold War’s world of two. Then the many: hundreds of

channels, a billion Web pages, an attic stuffed past the rafters. Youth produces, he wrote; age

3has to reduce. He predicted that the way out of the clutter would be reduction, and that the

same electronics that had overproduced our content would do the reducing.

A language model is a reduction machine. A billion pages go in. One paragraph comes out. The

many become one again, on every screen, in a calm and agreeable voice.

Fraim had already warned what that costs. In “Freedom to Be Equal,” he argued that Americans

had confused more choices with more difference. One mall looks like another mall. One suburb

looks like another suburb. If we all put our questions to the same few machines, and they

answer from the middle of everything that has ever been said, our symbols will start to look like

our suburbs.

What would be missing is place. Fraim closes one essay on a two-lane road in the Valley of the

Moon, a short drive from where I live in Sonoma County. A faded Gore-Lieberman billboard was

still standing there months after the election. Nobody had bothered to take it down. He could

not say quite what it meant, only that it meant something, the way the billboard over the valley

of ashes does in “The Great Gatsby.” Its meaning came from where it stood and how long it had

stood there.

A chatbot is from nowhere. A robot, though, has to be somewhere. One that worked the same

forty acres of vineyard for twenty harvests would know frost as a fact and not as a word.

Whether that would amount to having a symbol for it, nobody yet knows.

◆

There is one more audience to consider. What will robots say to each other?

In 2017, researchers at Facebook set two software agents to haggling over a pile of virtual goods.

Left to themselves, the agents slid out of English and into a clipped, repetitive shorthand that no

person could follow. The headlines said the machines had invented a secret language. The

plainer truth is that they had found a shortcut, and the researchers steered them back to

English.

That small episode says a great deal. A symbol, in Fraim’s sense, is gloriously inefficient. It

carries more than the message requires. Machines talking to machines have no use for the

surplus. Among themselves they will talk in code.

Symbols are what they will save for us.

So when the heart comes back across the screen, it is fair to ask who sent it. For now, the

answer is everyone who ever drew one. That is not nothing. It may be the largest chorus ever

assembled. But a chorus is not a voice, and the difference will matter most on the days we are

grieving, or voting, or deciding whom to trust.

4Fraim ended one of his essays by admitting that the thing he was hunting, the context behind

all our content, was “an elusive goddess.” The machines are hunting her now, too, with better

instruments than he had. What they report back will depend on a detail we have barely begun

to examine: what, exactly, we have been telling them about ourselves all this time.

Type a small red heart at the end of a message to a chatbot and, often as not, a heart comes

back. The exchange takes about a second. It may also be one of the stranger events in the

history of communication. A shape that people have carved into tree bark and stitched onto

pillows for centuries has just been handed to you by something with no pulse to quicken.

Did the machine use a symbol, or did it only return one?

That sounds like a seminar question. It is turning into a kitchen-table one. Machines now draft

our condolence notes, our wedding toasts, and our apologies, which are the kinds of speech

most heavily freighted with symbols. Robots with arms and legs are moving out of the

laboratory and into the warehouse, and their makers promise the living room is next. If these

things are going to talk to us for the rest of our lives, we ought to know what kind of talking it

is.

We have been here before, and we handled it badly. In 1966, the M.I.T. computer scientist

Joseph Weizenbaum wrote a program called ELIZA that did little more than turn a typist’s

statements back into questions. People confided in it anyway. Weizenbaum spent much of his

remaining career alarmed by how little it took.

◆

In 2001, a writer named John Fraim was posting essays to a Web site called Symbolism.org from

Northern California, where he had moved after what he describes as a narrow escape from the

entertainment business in Los Angeles. The essays are about the dot-com crash, the contested

Presidential election of 2000, the sameness of suburbs, and Marshall McLuhan. None of them is

about robots. Read today, they amount to a working theory of what a talking machine can and

cannot do.

Fraim starts with a definition. The usual one says that a symbol is something that refers to

something else. He tightens it: a symbol is something made by culture that points at something

outside culture. And the only thing outside culture, he argues, is nature. Spring. The ocean.

Night. Birth and death.

1Hold that definition up to a language model and a problem comes into view. The model is built

from text, and text is culture’s output. The machine has read more about spring than any

person ever will, and it has never once been cold. When it writes “spring,” the word points to

every other place the word has appeared. It points at the library. It does not point at the thaw.

The cognitive scientist Stevan Harnad named this predicament in 1990. He called it the symbol-

grounding problem, and compared it to trying to learn Chinese with nothing but a Chinese-to-

Chinese dictionary. You look up one squiggle and find more squiggles. You can ride that merry-

go-round forever and never touch the ground.

Fraim liked to quote a line of McLuhan’s: we don’t know who discovered water, but we can be

fairly sure it wasn’t a fish. A language model is the purest fish yet bred. It swims in our words

and has no shore to stand on. So the first answer to the title question is a deflating one. A

chatbot talks in symbols the way a player piano plays Chopin. The notes are all there.

◆

A robot is a different animal, because a robot has a body. It can bump into the table. It can run

low on power. And the body, it turns out, is where symbols begin.

One of Fraim’s shortest essays, “The Shape of Sequence,” draws on the work of Dan Acuff, a

consultant to toy companies. Acuff reported that children as young as eighteen months read

jagged lines as the bad guys and rounded lines as the good guys. That, he pointed out, is why

Mickey Mouse is assembled almost entirely out of circles. The German psychologist Wolfgang

Köhler had found something similar in adults back in 1929: shown a spiky shape and a blobby

one, and offered two made-up words, people agree with eerie consistency about which name

belongs to which.

Meaning, in other words, starts before language. It starts in the hand and the eye, with the

sense of what could hurt you and what will hold you.

Now picture the robots that companies want to put in your house. Look at the helpers in the

product videos, the vacuum under your couch, the speaker on your counter. They are

overwhelmingly smooth and round, with large dark faces and no teeth. Before a domestic robot

utters a word, its designers have made it say something to the eighteen-month-old who still

lives in each of us. It says: I am safe.

That is a symbol in the full sense. It points past culture to the body’s oldest knowledge. But

notice who is doing the talking. The robot is the message. The speaker is a design team.

◆

What will they say once they do use words? Fraim’s essays suggest three answers, and the first

comes from a piece of junk mail.

2In 1975, a freelance copywriter named Martin Conroy wrote a two-page letter to sell

subscriptions to the Wall Street Journal. It told of two young men who graduate from the same

college, join the same company, and meet again at their twenty-fifth reunion, one a department

manager and the other the president. The Journal mailed that letter for decades. By the figures

Fraim cites from the trade press, it brought in something like a billion dollars.

Fraim’s reading of it is the most useful thing in his archive. The content of the letter, he says, is

a sales pitch. The power is in the context. The story opens on a late-spring afternoon, at a

graduation. By its fifth word the reader is standing in the season of beginnings, at the modern

version of an initiation rite. Nobody notices. Everybody feels it.

Conroy found that formula once, by instinct, and it lasted a generation. A machine trained on

everything that ever moved us can find it on demand, for each of us separately, every morning.

The first thing robots will say, then, is whatever they are asked to sell, wrapped in spring.

◆

The second answer is that they may not need words at all.

In an essay called “The Democracy of Electric Action,” Fraim observed that actions had never

formed a language the way words had, and that electricity was changing that. Every click, card

swipe, grocery scan, and channel change was landing in a database somewhere. People were

voting hundreds of times a day without knowing it. Then he asked a question that reads

differently now than it did in 2001. Jung had built his idea of a collective unconscious on the

dreams of a handful of patients. What if the electric activity of a whole town were treated as the

mind of one patient? Or a whole country?

He was describing, near enough, the raw material that today’s systems are made from. And he

saw the fork in the road. That record of our actions could give a voice to people with no talent

for loud words. It had mostly been used, he noted, for persuasion and control.

A machine that reads us through our actions can answer in kind. It can speak through what it

shows us first and what it leaves out. That is talking in context rather than content, and it is the

kind of speech we are worst at hearing. Fraim had a phrase for our condition: the “trance of

content.”

◆

The third answer is the one that should keep an editor up at night.

Fraim borrowed an old philosophical sequence, the one, the two, and the many, and laid it over

American life. First came the mass culture of mid-century, with three television networks and a

few brands everyone shared. Then the Cold War’s world of two. Then the many: hundreds of

channels, a billion Web pages, an attic stuffed past the rafters. Youth produces, he wrote; age

3has to reduce. He predicted that the way out of the clutter would be reduction, and that the

same electronics that had overproduced our content would do the reducing.

A language model is a reduction machine. A billion pages go in. One paragraph comes out. The

many become one again, on every screen, in a calm and agreeable voice.

Fraim had already warned what that costs. In “Freedom to Be Equal,” he argued that Americans

had confused more choices with more difference. One mall looks like another mall. One suburb

looks like another suburb. If we all put our questions to the same few machines, and they

answer from the middle of everything that has ever been said, our symbols will start to look like

our suburbs.

What would be missing is place. Fraim closes one essay on a two-lane road in the Valley of the

Moon, a short drive from where I live in Sonoma County. A faded Gore-Lieberman billboard was

still standing there months after the election. Nobody had bothered to take it down. He could

not say quite what it meant, only that it meant something, the way the billboard over the valley

of ashes does in “The Great Gatsby.” Its meaning came from where it stood and how long it had

stood there.

A chatbot is from nowhere. A robot, though, has to be somewhere. One that worked the same

forty acres of vineyard for twenty harvests would know frost as a fact and not as a word.

Whether that would amount to having a symbol for it, nobody yet knows.

◆

There is one more audience to consider. What will robots say to each other?

In 2017, researchers at Facebook set two software agents to haggling over a pile of virtual goods.

Left to themselves, the agents slid out of English and into a clipped, repetitive shorthand that no

person could follow. The headlines said the machines had invented a secret language. The

plainer truth is that they had found a shortcut, and the researchers steered them back to

English.

That small episode says a great deal. A symbol, in Fraim’s sense, is gloriously inefficient. It

carries more than the message requires. Machines talking to machines have no use for the

surplus. Among themselves they will talk in code.

Symbols are what they will save for us.

So when the heart comes back across the screen, it is fair to ask who sent it. For now, the

answer is everyone who ever drew one. That is not nothing. It may be the largest chorus ever

assembled. But a chorus is not a voice, and the difference will matter most on the days we are

grieving, or voting, or deciding whom to trust.

4Fraim ended one of his essays by admitting that the thing he was hunting, the context behind

all our content, was “an elusive goddess.” The machines are hunting her now, too, with better

instruments than he had. What they report back will depend on a detail we have barely begun

to examine: what, exactly, we have been telling them about ourselves all this time.

One thought on “Will Robots Talk in Symbols? What Will They Say?”

Leave a Reply