AI Art: How artists are using and confronting machine learning | HOW TO SEE LIKE A MACHINE

Summary

The video script explores the intersection of AI and art, discussing how artists are using AI not just as a tool, but to challenge and redefine our understanding of technology. It delves into AI's potential for creativity beyond human biases, as seen in MoMA's 'Unsupervised' exhibition, and raises questions about the implications of AI on society, culture, and the environment. The conversation touches on the generative turn in AI, the importance of considering AI's cultural and political aspects, and the need for a broader perspective on AI's role in our lives.

Takeaways

• 🤖 **AI in Daily Life**: Many everyday technologies are powered by AI, yet there is limited understanding about how they work.
• 🎨 **Artistic Engagement with AI**: Artists are using AI not only as a tool but also to educate the public about its complexities and to provoke thought about free will and perception.
• 🧑‍🎨 **Subversion of Technology**: Artists are exploring how to use technology in unconventional ways, embracing it to create new expressions rather than rejecting it outright.
• 📊 **Supervised vs. Unsupervised Learning**: The script contrasts supervised learning, where humans label data, with unsupervised learning, which allows machines to find patterns without human guidance.
• 🌟 **AI as a Creative Medium**: At MoMA, Refik Anadol's exhibition uses AI to reimagine and speculate on artistic possibilities beyond human labeling and categorization.
• 🔮 **Machine Dreams**: The AI creates a complex classification system from MoMA's collection, exploring 'empty spaces' in the data to generate new, imagined content.
• 🌐 **Breaking Down Boundaries**: AI's multi-dimensional imagination transcends human biases and categories, blending past, present, and future in a convergence of ideas.
• 🔍 **Bias in AI Systems**: The discussion highlights the inherent biases in AI training data and the real-world implications of these biases.
• 🌱 **Evolution of Human-Machine Relationship**: The script touches on the historical evolution of how humans and machines have grown together, from early industrial fascination to modern AI interaction.
• ⚖️ **Ethical and Societal Concerns**: There are concerns about AI being deployed by large corporations leading to wealth and power consolidation, potentially exacerbating societal inequities.
• 🔄 **Rethinking AI Purpose**: Artists and researchers are exploring alternative uses for AI, such as creating collective dreams and consciousness, rather than just efficiency and profit.

Q & A

What is the general perception of AI among the public according to the transcript?

The transcript suggests that many people interact with AI daily but have little understanding of it. There is a sense of feeling captive or passive to technologies that are simply given to us.

How do artists view AI in the context of the transcript?

Artists view AI both as a tool and as a subject that people should understand more about. Some artists want to intervene in AI processes at a high level, exploring existential questions about free will and perception.

What is the difference between supervised and unsupervised learning as discussed in the transcript?

Supervised learning involves humans tagging information from the outset, while unsupervised learning allows the machine learning model to do the tagging on its own, often resulting in a 'black box' where the processes are not fully understood.

What was Refik Anadol's approach to the exhibition at MoMA?

Refik Anadol's approach was to use unsupervised learning to create artwork that does not mimic reality but instead dreams and speculates on the machine's imagination. He used MoMA's archive metadata to create a realtime software artwork that is always changing.

How does the AI in Refik Anadol's MoMA exhibition navigate the data?

The AI builds a complex classification system or map of MoMA's collection, groups data points, and navigates the empty spaces to speculate on what could exist, leading to a kind of dreaming and hypothetical scenario.

What breakthroughs in AI research were mentioned in the transcript?

The transcript mentions breakthroughs like OpenAI's DALLE, DALLE-2, ChatGPT, Stable Diffusion, and Midjourney, which are all supervised, labeled, multi-model AI algorithms allowing interaction.

What concerns do the speakers have about the use of AI in society?

The speakers express concerns about AI systems being used by large corporations, potentially leading to a consolidation of wealth and political power, and resulting in an inequitable society. They also discuss the biases in AI systems and the oversimplification of complex realities.

How does the transcript discuss the role of artists in the context of AI?

The transcript highlights that artists can bring a unique perspective to AI, using their historical and cultural understanding of images to contribute to the conversation in a way that the engineering computer science tradition does not.

What is the 'generative turn' mentioned by Kate Crawford?

The 'generative turn' refers to a crucial inflection point where traditional understanding of fields like illustration, film, and publishing is rapidly changing due to the impact of AI and generative capabilities.

What is the significance of Trevor Paglen's work 'Behold these Glorious Times!'?

Trevor Paglen's work 'Behold these Glorious Times!' is significant as it exposes the algorithmic biases in AI systems by showcasing the training images used for object and face recognition, among other things, and highlighting the real-world implications of these definitions.

How does the transcript suggest we should consider using AI tools?

The transcript suggests considering radical approaches to using AI tools, such as making them inefficient or working against their intended purposes, to upend expectations and challenge the status quo.

Outlines

00:00

🤖 AI and Artistic Expression

The paragraph discusses the intersection of AI and art, highlighting how artists are using AI not just as a tool, but also as a subject for exploration. Michelle Kuo and Paola Antonelli emphasize the lack of understanding people have about AI despite its pervasiveness. They discuss how artists are intervening in AI's processes to explore existential questions about free will, human perception, and the unseen aspects of technology. Refik Anadol shares his experience with AI at MoMA, contrasting supervised learning with unsupervised learning, where AI creates without human-imposed labels. Anadol's work with MoMA's archive data showcases AI's ability to dream and speculate, creating a new reality that challenges human biases and categories.

05:02

🔍 AI's Biases and the Generative Turn

This section delves into the biases inherent in AI systems and the broader implications of the 'generative turn' in technology. Kate Crawford and Trevor Paglen critique the assumption of objectivity in AI, arguing that these systems are deeply influenced by their training data. They discuss their collaborative work examining the datasets used to train AI, revealing the cultural and political biases embedded within. Paglen's artwork 'Behold these Glorious Times!' is highlighted as a way to expose these biases. The artists argue that AI's simplification of complex realities is problematic, and they advocate for a more nuanced understanding of the world, one that acknowledges the multiplicity and complexity of images and labels.

10:02

🏭 The Historical Context of AI and Technology

The final paragraph provides historical context to the relationship between humans, machines, and technology. It discusses how artists have long been fascinated by and critical of technology, from the industrial revolution to modern AI. Kuo and Antonelli recount how artists like Marcel Duchamp redefined art in response to technological advancements, questioning the role of the artist and the nature of art itself. The paragraph also touches on the evolution of human-machine interaction, from early machine art exhibitions to modern AI's potential to create collective dreams and consciousness. Paglen and Crawford express concerns about the concentration of power in the hands of corporations and the potential for AI to exacerbate societal inequities. They also explore the idea of using AI in unexpected ways, challenging its intended purposes for efficiency and work.

Keywords

💡AI

AI, or Artificial Intelligence, refers to the simulation of human intelligence in machines that are programmed to think like humans and mimic their actions. In the video, AI is central to the discussion on how it's being used by artists to not only create art but also to explore existential questions about free will and perception. The script mentions AI tools like OpenAI's DALLE and ChatGPT, illustrating AI's role in creative processes.

💡Supervised Learning

Supervised learning is a type of machine learning where an algorithm is trained on labeled data. The video script discusses this concept to explain how AI learns from human-tagged information, like identifying a pencil from a dataset, and uses this learning to perform tasks such as creating a realistic image of a bluebird.

💡Unsupervised Learning

Unsupervised learning allows machines to find patterns in data without human guidance. In the context of the video, unsupervised learning is used by artists to allow AI to create without predefined categories or labels, leading to unique and unexpected outputs that challenge our understanding of reality and imagination.

💡Black Box

A 'black box' in AI refers to systems where the decision-making process is not transparent or interpretable by humans. The script mentions this to highlight the mystery and complexity within AI systems, especially in unsupervised learning, where even the creators may not fully understand how the AI arrives at certain conclusions.

💡Machine Learning Model

A machine learning model is a system that learns from data to make decisions or predictions without being explicitly programmed. The video discusses how such models are used by artists like Refik Anadol to interpret and transform data, such as creating a dynamic artwork from MoMA's collection using a custom software.

💡Data Points

Data points are individual pieces of information or values in a dataset. The script uses the term to describe how AI systems group and analyze data points to create classifications or visualizations, such as the 'galaxy' created from MoMA's collection data.

💡Latent Space

Latent space in AI refers to the multi-dimensional space that represents the underlying structure of the data. Refik Anadol in the script talks about navigating the latent space of MoMA's archive to reconstruct potential AI dreams, indicating a complex and abstract representation of data.

💡Generative Turn

The 'generative turn' mentioned in the script refers to a pivotal moment where AI's generative capabilities are rapidly changing traditional creative processes in fields like illustration and film. It implies a shift towards AI-driven content creation and the implications it carries for the future of these industries.

💡Bias

Bias in AI refers to the unfair or prejudiced treatment of certain groups or concepts due to the data used to train the system. The video script discusses how AI systems are trained on datasets that can contain inherent biases, which then affect the AI's outputs and decisions, raising ethical concerns.

💡Multi-Dimensional Imagination

This concept from the script refers to the ability of AI to combine different elements, such as past, present, and future, or various materials, to create new forms of expression. It contrasts with human perception by suggesting a convergence of elements that we typically consider separate.

💡Collective Consciousness

Collective consciousness in the video is discussed as a potential outcome of AI algorithms using collective memories to create shared dreams and experiences. It suggests a future where AI might help humans connect and understand each other on a deeper level through shared digital experiences.

Highlights

AI is increasingly integrated into everyday life, yet there's limited understanding about it.

Artists are using AI not only as a tool but also to increase public understanding of AI.

Some artists aim to intervene in AI processes to explore existential questions about free will and perception.

Artists often take existing tools and use them in unconventional ways to challenge AI's intended use.

AI breakthroughs like DALLE-2 and ChatGPT are following a pattern of supervised learning.

In supervised learning, humans tag information, guiding AI's understanding and creation.

Unsupervised learning allows AI to create without human-imposed labels, encouraging machine imagination.

Refik Anadol's work at MoMA uses unsupervised learning to create a 'dream' of AI from MoMA's archives.

AI can create a complex classification system, forming a 'galaxy' of data points with empty spaces for potential creation.

AI-generated art can challenge traditional categories and reveal a multi-dimensional imagination.

AI systems are often assumed to be objective, but they are inherently biased from their training data.

Trevor Paglen's work examines the cultural and political biases embedded in AI systems.

Training data for AI systems can include personal moments, raising ethical concerns about privacy and exploitation.

AI's simplified labeling of complex images can lead to a 'bleached' version of the world.

Artists bring a unique perspective on meaning and imagery that computer science lacks.

The history of art and technology shows a continuous exploration of the relationship between humans and machines.

Designers have evolved to understand how to use machines, as seen in the evolution from OCR-A font to AI-generated concepts.

AI has the potential to solve hard problems, but there are concerns about wealth and political power consolidation.

The full life cycle of an AI system, like Amazon Alexa, shows the extensive planetary cost of AI.

There's interest in using AI tools in unconventional ways, potentially making them work against their intended purpose.

AI algorithms may redefine creativity and challenge our definitions of reality.

Transcripts

Michelle Kuo: Many of the things that we interact with daily are powered by AI in some way,

but there’s very little understanding about it.

Paola Antonelli: There are artists that are using AI as a tool and there are artists that

want the people to understand more about AI.

Kuo: We sometimes feel captive to or passive to the technologies that are simply given

to us.

And for a number of artists, they want to intervene in those processes at a high level

and think about existential questions of free will of human will, machine will, but also

questions of perception.

How can we see things that are actually not made for us to see?

One thing that artists have always been very good at is taking a tool that exists in the

world and making it do something it's not supposed to do.

When artists are looking at technology as another kind of tool to experiment with, or

maybe even subvert or divert, they're saying, I'm not just going to reject this technology

wholesale at all.

In fact, I might embrace it, but I might try and make it do something else.

Refik Anadol: This the show at MoMA right now how it looks like how many people there

like that's…wow.

Last year we saw amazing breakthroughs in AI research like OpenAI's DALLE, DALLE-2,

ChatGPT, Stable Diffusion, Midjourney.

Like these are all actually following a very similar pattern, which are extremely supervised,

extremely labeled, multi-model AI algorithms allowing us to interact with them.

Kuo: In quote unquote supervised learning, which is the more conventional mode of machine

learning.

Humans are the ones that are essentially tagging these bits of information from the outset

saying, here's a picture of a pencil, this is a pencil.

AI is quite good at saying, "If you tell me to create a blue bird, I'll, I'll show you

a picture of a bluebird that looks real."

Anadol: But this was to me not very inspiring to be honest.

To me, what was more inspiring is what happens if you don't use a technology as it's imposed

to us, but used a different way.

And in this context unsupervised, the exhibition now at MoMA is actually doing something different.

It's not exactly following the labeling data or try to mimic reality.

It is trying to dream and speculate an imagination of a machine.

Kuo: So unsupervised learning means that actually you are letting the machine learning model

do that tagging based on its own sort of...

Antonelli: Who knows?

Kuo: Black box.

Yeah, it's a, actually it's..a lot of it is a black box, like we don't actually know what's

going on in there.

Anadol: For "Unsupervised," we took the entire metadata of MoMA archives, which is around

138,000

Anadol: images.

Kuo: And Refik used that data to create custom software artwork that would interpret and

transform MoMA's collection data.

He has created a large scale presentation of this realtime software artwork.

It's almost like a performance.

It's always changing.

The machine learning model has built an incredibly complex classification system or map of MoMA's

collection.

It has decided it's going to group a number of data points over here and a number of data

points over here, and you create a kind of galaxy, but in that galaxy there's a lot of

empty space.

So then the machine learning model in concert with Refik's team is sort of navigating through

that empty space and saying, "Nothing exists here, but what could exist?"

And that is where the kind of speculative and hypothetical and even what we might call

a kind of dreaming starts to take place.

Anadol: I should really show you this.

This is like real next level.

So what you see on the left side is completely not real.

These are all AI-made potential AI dreams.

So what we can do here, I'm literally flying in the latent space of MoMA archive and reconstructing

these potential dreams.

Once AI starts to create this new reality, we learned that there is not any borders between

these biased categories that we need as humans to understand things.

That is a truly multi-dimensional imagination.

It is blending past, now, and future.

It is blending multiple materials.

It's just convergence of things that we thought they are independent.

Antonelli: The AI is not using our value system, it's not using our points of reference.

It's just peering into another type of mind.

Kate Crawford: I think we are at a crucial inflection point right now.

I've been calling it the generative turn.

It's a moment where what we previously understood as how everything from illustration to film

directing to publishing works is all about to change very rapidly.

There is this assumption that AI systems are somehow highly scientifically objective that

they are presenting a view on the world that is somehow unmediated.

But of course the opposite is true.

You know, these are systems that are profoundly skewed from the absolute beginning, from the

data that they're trained on.

This is something that I worked with Trevor Paglen on, where we studied the training data

sets that are used to teach AI systems to see the world.

Trevor Paglen: My interest in AI is not at all like, oh, what kind of wizbang kind of

image can you make using it?

That's actually totally uninteresting to me.

What's interesting to me is looking at AI as not only a set of technical systems, but

technical systems that have culture built into them that have politics implicitly built

into them and trying to unpack that.

Kuo: Trevor Paglen's work “Behold these Glorious Times!” is a kind of hypnotic video

that is basically flashing at you many different images that are used to train AI.

Paglen: The age of machine learning is kind of characterized by building AI systems and

computer vision systems where the idea is that you have a lot of images of different

things and then you use the neural network to find patterns across those images.

So "Behold these Glorious Times!"

looks at those kinds of training sets, these much larger databases that are used for things

like object recognition, forms of contemporary face recognition, but other things as well.

For example, gesture recognition or gait recognition.

So the video installation goes back and forth between just looking at the images by themselves

at this very fast speed and then starting to get a glimpse of what the machine learning

system is doing internally to try to distinguish these images from one another.

Kuo: Trevor is an artist who is really laying bare the algorithmic and inherent biases in

many AI systems, but also the ways in which these definitions have real world implications,

many of which are obviously terrifying.

Paglen: There's moments towards the end of "Behold these Glorious Times!"

where you're seeing training sets that are made out of things like family home videos

or extremely personal moments in people's lives, and they've been incorporated into

training sets to understand, "Oh, this is a mother like putting down a baby, so we wanna

understand what a mother with a baby looks like so that we can try to sell them diapers

or whatever we want," right?

And so you're seeing the ways in which certain kinds of content are being ingested into machine

learning systems in order to try to capitalize on learning what can be extracted from those

moments of intimacy.

Crawford: It still strikes me as preposterous that we're assuming that a single image can

be given a label, a single word when we know about the multiplicity and complexity of a

single image, the idea that we can so benignly label something as a chair and then a person

as say a debtor or a kleptomaniac.

These are things that are literally happening today in data sets.

And the risk there is that we're starting to see a very simplified and in some ways

just really bleached version of the world.

Paglen: We have all kinds of allegorical and kind of squishy meanings attached to all of

the things that we look at in our everyday lives.

Those are informed by our own histories, our cultures, our own memories.

And so that question of meaning making is all over the place.

Artists, what we bring to the party is thousands, if not tens of thousands, of years of thinking

about what the hell an image is.

The kind of engineering computer science tradition does not have that.

This is a place where artists are bringing voices to the conversation that I think are

quite urgent.

Antonelli: We're talking about AI today, but in truth, the fascination and the fear that

humans have with machines and with technology has been amplified and examined and explored

by artists and by designers since technology existed.

Kuo: Early in the 20th century, artists were fascinated by industrial production and what

that meant for someone like an artist.

Suddenly the most skilled human technician was actually outpaced by an apparatus like

a camera or some kind of forming machine that used a conveyor belt.

And artists like Marcel Duchamp said, "Well, wait a second, I'm actually gonna radically

redefine what an artist does and what art even is."

And he took a ready-made industrial object, like a bicycle wheel, and stuck it on a stool

and said, "This is art, because I say so."

In one fell swoop, he realized this kind of crisis of the artist in the 20th century,

which is what is art if it's not technical facility or total realism.

These are things that suddenly felt scary but also exciting.

Antonelli: I would like to just move to the beginning of the Museum of Modern Art and

in 1934, there was a show called Machine Art in which pieces of machinery were unveiled

and placed on white pedestals against white walls, and the beauty of the machine revealed

to the world.

And slowly but surely, designers have been trying to really understand how to use machines.

So OCR-A from the mid 1960s was a font that was designed to be understood by machines.

A few decades later, it's instead the machines trying to make concept as readable as possible

by humans.

It's an evolutionary process in which humans and machine kind of grow together.

Paglen: There are definitely a lot of hard problems that AI can absolutely help solving.

Having said that, and going back to the idea that the context in which these tools are

always being deployed is by huge corporations, they worry that there is a huge potential

for a massive consolidation of wealth and political power, and I'm concerned that that

adds up to an increasingly inequitable society.

Even if the problems that we want to solve can be solved, it's always about capitalism,

not technology, right?

Crawford: This system is drawing on all of us.

It's drawing on our labor, on our voices, on the earth, on the meteorological layer,

on energy, on water.

So for anatomy of an AI system, we really wanted to show the full life cycle of an AI

system, and in this case, we chose the Amazon Alexa.

So every time you turn to Alexa and say, "Hey Alexa, what's the weather tomorrow?"

You are bringing into being this incredibly complex network that starts all the way back

in the mines, where the rare earth minerals are being extracted through to the end of

life of these devices to really show that full planetary cost of an AI system.

I'm really interested in quite radical approaches of how people could use these tools in ways

that they've never been designed to use.

How might we upend the expectations that these tools are for work or for efficiency, and

to try and find ways to make them inefficient, to find ways to actually make them work against

themselves?

Anadol: I also believe that AI algorithms may have a different purpose.

It's kind of this finding the language of humanity by using collective memories to create

collective dreams and eventually collective consciousness.

The near feature that is coming right now very much will be questioning creativity,

questioning who will define what is real or not, and I think we have to be ready.