All language subtitles for Why AI is Doomed to Fail the Musical Turing Test

What would it take for a machine to jam?

This question was first posed
in Computer Music Journal

in 1988 as a form of musical

turing test, a way of identifying human-like

intelligence in machines. And this test is

very simple. Blues and E flat,

one two, a 1234.

What would it take in this case for a

machine to convince me and you that it was

human? Well, very quickly, a number of

things need to happen in real time. First,

the machine would need to identify where one

was and entrain to my pulse using

not only auditory but visual cues. We're very

good at figuring out if other people are

locked in with our pulse. And so I would

need to sense that the machine was feeling

the groove. Two, the machine would

need to identify what it was that I was

doing on my bass guitar, and then fit that

within the paradigm of a twelve bar blues.

Am I doing the quick IV, for example,

or am I playing a II-V on the turnaround

instead of a V VI? So am I playing a

jazz blues versus a delta blues versus a

Chicago style blues? It would then need to

take that information and respond in kind in

an improvised solo, using meaningful blues

vocabulary. Now, music is not really a

language, but it certainly feels like a

language to those who play music. And so if

I ask a musical question, does it feel

like I'm getting a meaningful answer in

response? Does it feel like I'm

connecting with somebody, that there is a

real ghost in the machine on the other side

of the algorithm?

I'm fairly convinced that this will never

happen. No AI will be able to pass a true

musical Turing test. Call Ray Kurzweil.

Tell him hes a hack. Singularity, my ass.

Now, I'm not going to bullBASS you in this video.

I'm not going to appeal to some vague sense

of the uniqueness of human musical

creativity - "AI has no soul" - because it's

gonna get there, right? I mean, Red Lobster

is already using AI generated music in its

ad campaigns. And who are we to question the

aesthetic sensibilities of Red Lobster? Red

Lobster, you got the magic
touch. "Cheddar Bay Biscuits

I love you so much" There have
been some incredible advances in generative

AI from companies like Udio and Suno AI.

They let you generate full pieces of music

from text prompts. Just type in what you

want and it will do a pretty good job of

giving it to you.

"In Tommy's Shack Grill late at night.

Big Joe Flippin those paddies, fine.
American Cheddar melting just right."

That's horrifying. It's like spitting

on Muddy Waters' grave. Upon hearing these

kinds of results, many techno-optimists have

breathlessly extolled that the musical

Turing Test has been passed.

Machine musical intelligence is upon us.

Daddy Elon is so excited.

So why am I so doubtful here, right?

Why am I saying that music AI
will never pass the Turing Test?

Well, I think there's a
pretty profound category error

that's going on here that we need to always

be aware of going forward.

What generative AI does is not music.

Let me explain.

This video was brought to you by Nebula.

Hope of an extraordinary

aesthetic success based on extraordinary

technology is a cruel deceit.

Iannis Xennakis, 1985.

In 1950, Alan Turing first wrote about what

he called the Imitation Game, what we now

call the Turing Test. In this test, an

interlocutor asked questions of two

entities, one machine and one human. And if

at the end of a conversation - traditionally

done through a text prompt - the interlocutor

is unable to tell which one is the machine

and which one is the human, then we say that

the machine passed the Turing Test. It

displays human level intelligence. Now,

how might we take that idea and expand it

out into the world of music? Well, one

scenario that we showed at the beginning of

this video involves a "Turing Jam Session".

During the improvisation among an

interlocutor and two musicians, the

interlocutors task would be to identify who

is the machine and who is the human. I'm

really drawn to this test because jam

sessions are a great way to get to know

people, know what they're about, know their

musical taste. They're a lot of fun. They're a

social activity. But there are other

possible musical Turing tests. Christopher

Ariza loosely categorizes them as either

musical directive tests, which involve

ongoing musical interaction between the

interlocutor and two agents, and musical

output tests, which involve no interaction.

The listener simply judges the human-like

quality of a given musical output.

I have no doubt that generative AI will be

able to pass musical output tests. Red

Lobster's marketing team apparently thinks

so, too. But musical directive tests, where

there's a continuous interaction between

agents, put the emphasis on process, not

product. Alan Turing's original imitation

game envisioned a conversation between

agents, not just somebody passively looking

at text and determining whether or not it

was computer generated, whether chatGPT

did your homework, in other words. And so to

pass a musical Turing test, a machine has to

do what musicians do when they make music.

The process has to feel human. One way that

visual artists have already started to push

back at the flood of generative AI images is

to document their process and tell the story

of making the art. This approach will become

more and more common with musicians over the

next couple of years as a means of pushing

back against the inevitable tide of

generative AI slop. So back in 2008,

Jack Conte, current CEO of Patreon, was

releasing music with Nataly Dawn under the

name Pomplamoose. They were releasing these

YouTube videos in a format that they called

video song. Video songs had two one,

what you see is what you hear, and two, if

you hear it, at some point, you saw it.

Now, this seems kind of like obvious now

because this format is so ubiquitous. But

back then, back in 2008, it was a

revolutionary approach to releasing art.

You were seeing the music as it was actually

made - an appeal that I imagine will carry

some resonance in the future in the face of

generative AI. The music theorist William

O'Hara expands on this idea of showing your

work for musicians. Borrowing a term for the

ancient greek word for craft, he calls it

the techne of YouTube. A techne of YouTube

performance, then, is a form of music

theoretical knowledge that exists at the

intersection of analytical detail,

virtuosic performance ability, practical

instrumental considerations, and an

awareness of one's audience and the

communicative tendencies of social media.

Pomplamoose has since expanded on this idea

of techne in social media musical

performance by releasing these short form

video things where you see the musicians

actually working out the music in the studio

and joking around, you get to see how the

sausage is made, so to speak. Can we try

snapping a tight one? Let's try a tight one.

Sorry, you want to snap a tight one?

You also get to see what

exactly would be necessary for a machine

intelligence to do if it was to pass

the musical directive test.

It would need to understand
and respond to musical jokes.

It was just like, we're feeling this in two.

I just think about everything in one. One.

One is a 400 bar form.

It would need to understand
and take musical direction.

It would need to vibe with Jack Conte,

so to speak.

I talk about him all the time on this channel,

but the musicologist Christophe Small

talks about how music is not really a noun,

but in fact a verb. In his book Musicking,

he writes that music is not a thing at all,

but an activity, something that people do.

The apparent thing "Music" is

a figment, an abstraction of the action,

whose reality vanishes as soon as we examine

it at all, closely.

Singing protestant hymns in a church,

dialing in sick lead tones on a quad cortex

freestyling live on the radio,

facing the wall of death at a metal festival,

watching singers perform o
sole mio in a concert hall,

sitting in at a jazz jam session,

streaming yourself on Twitch
producing music in FL Studio,

its very to difficult to see what these

things have in common

besides somehow reacting to organized sound.

And any one of these activities might be a good

candidate for a musical directive test.

They are all different examples of ongoing

musical interactions.

Generative AI, on the other hand,

is very good at creating products,

musical recordings.

But that product is only ever useful for

passing the output test.

Passing the directive test, though, would

require AI researchers to treat music as a

verb, like Christopher Small suggests a

process, a thing that two or more people do

together, like in Alan Turing's original

imitation game. And when you do that, you

have to take a look at the dynamic

relationships between audiences and

performers and the technology they use and

the spaces that they make music in. And by

the way, all of this is just for western

music so far.

The whole thing is a lot.

I am not gonna pretend like I know

how AI works. I am but a simple bass player.

I have tried to read those papers and I am

just not smart enough. I do recommend

Valerio Velardo's videos on AI music if you

want to get into some of the technical weeds

about these things. But basically, the way I

understand it is that a large language model

will train on a bunch of data and then use

that data to try and accurately predict what

the next thing will be in a sequence. This

is basically what the computational

cognition model is for humans, which says

that we take information, information in

from the world through our senses as input,

and then we process it in our brain, and

then our brain outputs behavior. This is a

fairly outdated model, and one that I don't

feel like applies to how we think about

music. And if the point is to pass the

Turing test, the machine has to think like a

human.

Anybody who has ever performed knows that

getting stuck inside your head

is the worst possible thing.

Thinking too much means you

cannot react with meaningful musical ideas.

But your mind isn't blank, you're still

thinking about things.

Its just very fragmented.

One hip new theory in philosophy

of the mind that accounts for this is called

4E cognition, after the four E's

Embodied, Extended, Embedded, and Enacted

cognition. They represent a dynamic

relationship between the brain, the body,

and your environment.

The first of these is Embodied cognition.

Your body shapes how you think.

When it comes to music, this means

that if it sounds good, it's because it

feels good.

And this is backed by two decades of

music neuroscience research,

especially when it comes to auditory motor

coupling. The areas of your brain which

process movement through space are the same

areas of your brain that process rhythm and

music in general. The vestibular system that

governs balance influences your sense of

downbeat, where one is. If you're not

physically balancing your body, you might

lose the downbeat. A very common occurrence,

like what I just did here in this jam with

Rotem Sivan and James Muschler. You see me

swaying maybe a little bit too much there in

the background, and then the rhythm gets all

...floaty.

Cool, but maybe not intentional.

Failure is actually something that Alan Turing

identified as a means of getting a machine

to fool people into thinking that it was

human. If a machine is too good at answering

questions, it wont seem human.

To err is human,

and so to pass a Turing test,

an AI might need to lose where one is.

The second form of cognition is

Extended cognition,

where the world is your brain.

Thinking requires a lot of energy -

evolutionarily speaking - your brain gets

tired sometimes. And so humans have figured

out ways of extending our thought patterns

into the physical world as a means of

reducing cognitive load. The classic example

of this is writing. We write things down so

we dont have to remember them anymore,

freeing up cognitive capacity in our brain.

Plato famously complained about this, how

people were getting lazy because they relied

on writing too much. Smartphones are the

latest example of this, extending our brains

into the world with technology. We could say

that music notation is a form of extended

cognition, by letting us remember more

music than we would normally be able to with

our mere brains. Orchestral composers don't

have to remember every note for every

instrument that they have ever written, and

so they are freed from the constraints of

their own memory to imagine larger and

grander musical designs - music shaped by

extending our brains past their limitations.

An AI might need to show forgetfulness if

it's going to pass a musical Turing test.

The third form of cognition is

Embedded cognition.

Patterns of thought are
embedded in external systems.

One way to think of this is
how I have embedded my musical

vocabulary into a system of tuning for bass.

4th tuning. Like, I don't have to think

that much to be able to express myself with

my bass tuned like this.

The notes are just there where my fingers

expect them to be.

But if I was to detune my bass a little bit

to an unfamiliar tuning system,

I don't know where anything is anymore.

So my cognitive load has increased

as I hunt and peck for each individual note

on my instrument.

Anybody who's ever tried to type with a

Dvorak keyboard

knows what I'm talking about.

The patterns of thought embedded in

the technology that we use,

like the bass guitar,

guide our musical intuitions.

You can tell if somebody wrote a piece in a

digital audio workstation versus musescore,

for example, would an AI need to mimic this

pattern of embedded cognition to pass a

Turing test? I don't know, but this is how

we do it, you know?

The fourth form of cognition is

Enacted cognition.

Doing, is thinking.

You process the world through action.

It basically says that there are

certain activities which are meaningless

if you are passive,

like sports, for example.

We don't say that you're a good

soccer player if you spend a lot of time

thinking about soccer, because you have seen

a lot of other people do it, you know, you

kind of gotta get out there and actually run

and do the thing yourself.

The very first question from this video,

what would it take for a
machine intelligence to jam?

Is like asking,

what would it take for a machine
intelligence to play soccer?

And the answer is, a body.

We need replicants.

We need androids out there on the field.

Otherwise, it's just a supercomputer

thinking about soccer.

Without a body,

AI sports intelligence is meaningless.

You gotta have robots on the field,

processing in real time what's going on

with their robot bodies.

Without a body,

AI music intelligence is meaningless.

You gotta have robots on the bandstand,

processing in real time what's going on

with their robot bodies.

As the great jazz educator Hal Galper said,

we musicians are athletes of the fine muscles,

and like athletes of the larger muscles,

our meaning is created by our bodies doing things.

Like our siblings in sports,

we share in vivo, the struggle, the joy,

and the experience of our lived selves

moving through the world.

Embodied AI is a long way off, but I don't

see any technical reason why we can't have

musical terminators. Again, I'm not an AI

researcher, so I don't know any of the

actual nitty gritty with any of this stuff,

but, you know, it could happen. And I also

see how it would be possible to treat music

more as a conversation between human and

machine, passing a musical directive test by

valuing and creating meaning in the process

of musicking. So why am I so doubtful?

Why is the thesis of this video that

the musical Turing test will never be passed?

We will never have musical machine intelligence.

[CAPITALISM]

The presumed autonomous thingness of works

of music is, of course, only part of the

prevailing modern philosophy of art in

general. What is valued is not the action of

art, not the act of creating, and even less

that of perceiving and responding, but the

created art object itself.

You can sell a thing.

It's harder to sell the process.

If the process of making quality

recorded music can be made more efficient,

the market is incentivized to make the

process as efficient as possible.

Generative AI creates recorded music

extraordinarily cost effectively compared to

the other ways you might do it. It fully

automates a process that had previously

required human input, much the same way that

industrial capitalism automated making cars,

making food, and making things,

as long as those things are good enough. In

other words, as long as they pass the

musical output test, the processes are

relevant, only the product. There are

billions of dollars now being thrown at

developing generative models for language,

images, and now music. Because there are

potentially billions of dollars to be made

in the market. Spotify is now in on the

generative AI music trend. Theres just, you

know, no money to be made in passing a music

directive test. You're just making the

process to get to the product less

efficient. And I mean that very literally,

too. The current prize for passing an

improvisation based Turing test is only

$1,000. And with such weak market pressure

to do something like this - I mean, I guess

you could sell tickets to see this as part

of a live show - theres just no reason to

spend that kind of computational energy into

doing this. The cloud computing power

required to run large language models uses

absurd amounts of energy consumption is a

great bottleneck for AI. Based on the

extremely intensive computational demands of

training. The carbon cost of image

generation is staggeringly high, and raw

audio generation is slated to be much

higher. I mean, just to, I guess, put this

in perspective, I think a gigawatt, it's

like around the size of

a meaningful nuclear power plant,

only going towards training a model.

Who is going to build the equivalent

of nuclear power plants so that robots can

jam with me? It's just so much more

efficient to have humans do the jamming.

I'm old fashioned and

very idealistic about that. My feeling is

I'll outplay anybody using the machine

or I'll die. I don't care. The day that

the machine outplays me, they can plant me

in the yard with the corn. And I

mean it. I'm very serious. I will not permit

myself to be outplayed by someone using the

machine. I'm just not going to permit that.

You know, there are people that I respect

that use this technology to make beautiful

music that somehow captures what it means to

be human in the year 2024. And I think

that's exciting. But then there are people

that I do not respect, like the people who

run companies like Suno, Udio, and other AI

companies, who have a very accelerationist

mindset when it comes to this, it feels like

music is just one more box to tick on

the way to the singularity. Music is a

problem that technology can solve. There

seems to be a profound disinterest in the

artistic process, why music sounds the way

that it does. And so you get things like the

beautiful, rich history of the blues, a

Black American tradition, reduced to typing

into a text prompt. Delta blues about

cheeseburgers. That's why I refuse to call

this stuff music, because the technology

behind it is so aggressively anti human,

anti history, anti music. And that's why I

also feel like the musical directive test

will never be passed because the people

running the show just don't care. You know,

one of the things that I've learned over the

years talking about musicking on this channel

is that you can change your relationship to

music that you hate by choosing to musick it

differently. And I kind of hate this Red

Lobster tune because it kinda slaps red

lobster. You got me up every

single time.

There's like a phrase extension too

disgusting.

Now, I owe a massive amount of context for

everything that we've talked about here

today to a video that I saw from the science

creator Tibees - Toby Hendy, where she goes

over Alan Turing's original 1950 paper,

where he first details his imitation game.

She highlights just how visionary Turing's

paper truly was, like how he predicted that

machines would need a degree of randomness

in them to evolve to have human like

intelligence. This has actually turned out

to be an essential part of modern machine

learning. She also covers how Turing thought

of potential objections to the idea that

machines could become intelligent,

including some weird ones like Turing felt

like he seriously needed to address the

prospect of extrasensory perception.

Anyway, this video was a great one, giving

me some extra context about the history of

machine learning. And you can find it and

many more like it over on my streaming

service, Nebula. Nebula is a creator-owned

streaming service that was originally

started as a means for creators to make

interesting videos and essays free of the

constraints of the recommendation algorithm.

But it's since organically grown into,

like, one of the genuinely best places on

the Internet for curious folks to find

exclusive, interesting, and enriching

content. Like for example, on there you'll

find science creators like Tibees and Jordan

Harrod, who does some fantastic stuff with

AI. You'll find amazing video essayists like

the OG video essayist herself, Lindsay Ellis

is on Nebula. And also Jacob Geller. If you

have never seen a Jacob Geller video essay,

I highly recommend you check out some Jacob

Geller. Like some of this stuff is so

beautiful. I think he's one of the the best

people in the game making video essays. Go

check out some Jacob Geller. You'll also

find some of my fellow music creators that I

deeply love and respect, like the queen of

jazz education herself, Aimee Nolte is

on Nebula. You also have the wonderful music

theorist 12Tone making videos over

there. I genuinely love the community of

creators over on Nebula. They are such a

wealth of inspiration for me, and I know

they will be a wealth of inspiration for you

too. If you're already on Nebula, we're

making it easier to find content for both

new creators and existing favorites. There

are categories now, news, culture, science,

history, podcasts and classes.

Each category is kind of like its own mini

service. Like, I have some classes over

there. I have a class on vlogging, which you

might enjoy. I also have a class that I did

with Aimee Nolte on jazz and blues that I know

you will enjoy. If you like my nerdy music

theory channel, if you sign up using my link

at Nebula.tv/adamneely, or use the

link below, you can support me and all the

creators over on Nebula directly and get

Nebula for 40% off annual plans,

which is as little as $2.50 a month.

What's exciting, though, and genuinely

unique to nebula, I think, is that now

Nebula is offering lifetime time

subscriptions, which means, yes, now until

the singularity, the end of time. Thank you,

Ray Kurzweil. You can be enjoying Nebula and

all it has to offer. There are some big

concept, high octane Nebula originals to be

excited for coming this summer. Like

Identiteaze, the debut short film from Jessie

Gender coming this June. And of course, Jet

Lag season ten is now in production. You

actually might remember Toby from season

five of Jet Lag The Game, but she'll also be

appearing in the latest season alongside

Ben, Adam, and Sam from Wendover. I'm a fan

of Jet lag the game, by the way. It's kind

of like reminds me of tour. It feels almost

weirdly comfy watching them run around

because it's like me running around Europe.

Anyway, $300 gets you lifetime access to all

of this great stuff and everything that

Nebula will ever produce

from now until the end of time.

I love reading that, by the way.

Now until the end of time.

That's. That's fun.

I'm very excited for the future of Nebula,

and I think you will enjoy this

community that aims to engage the world in a

more meaningful human way.

Thank you so much for watching. You can sign

up today for either 40% off annual plans or

$300 off lifetime access. And until next

time, guys,