Name: How Artists Are Rewriting AI’s Future | Fei-Fei Li, Holly Herndon and Mat Dryhurst
Uploaded: 2026-02-06T06:57:05.000Z
Duration: 13 min 19 s
Description: Thousands of YouTube videos with English-Chinese subtitles! Now you can learn to understand native speakers, expand your vocabulary, and improve your pronunciation...

Transition words & phrases

I do believe we're in a civilizational moment because AI and the surrounding technology is causing such a profound shift in the way we can live our lives, that not only it's a technological inflection point, it's going to cause an inflection point in our social structure.

As a scientist, I do feel a deep and profound sense of responsibility.

And because I'm one of the people in our generation that brought this science into the world, my contribution in AI started with realizing the importance of data.

We're constantly producing new behaviors, new ideas, new activities.

But the more profound question is just because individuals are producing data, whether they're creators or just going about everyday life, doesn't mean we can always just take it all the time for free and without the considerations.

There are science questions, there are art questions, and then there's also design questions.

I would say that design and art questions are more about the actual impact of AI in the world.

They're about imagining possible futures and trying to understand the consequences that they can have on the world.

We started working with AI about 10 years ago.

My background is in computer music and when I learned about AI in the audio domain, it felt like an opportunity to create sounds that had never existed before and also new ways of performance.

So I'm really interested in these really important historical moments that kind of open up a possibility for all kinds of artists to explore and kind of run with it and create new forms of art with.

We're working right now with the artists Holly Herndon and Matt Driehurst, who are incredible collaborators and at the forefront of art and technology.

The last few years of AI have been almost like 500 years of change as machine learning models have developed in scope and have been think learning and training and producing new kinds of sensory experiences and information.

And Holly and Matt are coming up with new and unorthodox ways of perturbing and disturbing those systems.

We were really building things from the ground up, which required questions like what's your data set?

AI models are actually just us in aggregate.

So an AI model output is both the product of everybody and nobody in particular.

Is you train a model properly, everyone's in there, and

the result you get is the responsibility of nobody producing.

Training data should be a deliberate act.

I had always wanted to make a kind of one to one voice model and it felt like the only right answer would be to use my own voice.

And so I sang from a series of timit scripts, which is basically a series text that cover all of the phonemes of the English language.

The computer will understand how your voice sounds.

While most people would want to protect their voice, we thought it'd be more interesting to do the opposite, to actually make Holly's voice public domain and give people the tools to use it as they wanted.

I trained an AI on my own voice, and now she can sing anything in multiple languages.

In constructing a different protocol for how training might take place, you're kind of pointing to a different world.

In 2024, Holly and Matt created a work called Play from Memory for an online exhibition at MOMA.

They were inspired by a 1920s form of music pedagogy for children that used what they called imitation games to teach children how to play music, but

And what you see is a series of fantastical images of instruments, for example, that never existed, and

And these are all generated through machine learning models that Holly and Matt fine tuned, trained and constructed to be able to almost speculate about what this kind of musical pedagogy could produce.

They were fascinated by the similarities between how people thought children could be trained to play music and how we think about training machines every step of the way.

They have confronted the challenges and issues of AI and consent, and

this is why they are really interested in participatory forms of music like choral music, even medieval hymns and call and response.

They've done a arena concerts where they have these huge crowds.

They ask to opt in into a kind of training set, and

it becomes a beautiful call and response.

And then Holly and Matt will use those sounds, recorded sounds, to then create new generative AI artworks.

And in the sense that training sets usually are taken from images and sounds and pieces of information that are just out there in the Internet, Holly and Matt are curating and crafting and asking people to participate in creating their own training.

As part of the Serpentine show, the Call that opened in 2024, we decided to create a choral singing model.

I think we approach AI as a kind of really sophisticated coordination technology.

And so that's why we look at earlier coordination technologies like choral singing or group singing practices, because it really helps us create a kind of through line thinking about the really sophisticated decisions that are made when human beings are standing together, singing and creating Music together.

If you've had that experience, you know that your contribution is important and you're playing a key role, but

that what you're creating is something that's greater than the sum of its parts.

We also created a data trust where people could join the trust and kind of negotiate how their data would be used for the project.

It was a really kind of like a full stack approach to training a model from the ground up, with everyone's consent.

We would probably argue that intelligence is something that happens between us.

And so a frustration of ours is that, yes, when people often talk about AI and art, they think about typing some words into a model you don't own and getting a funny looking image out.

And it's like, yeah, that exists and we've all got to exist in a world where that exists.

The actual value is an interplay between things versus a picture that you put on the wall.

In the last couple of years, I'm just so excited by this new frontier that we're exploring in AI that I call spatial intelligence.

Spatial intelligence, to me is a continuation of my lifelong work in computer vision, which really tackles the problem of percept and visual intelligence.

Vision is not just for the sake of passive viewing.

Vision is for doing things, for moving, to navigate the world, to manipulate things, to collaborate.

The first robot, the first companion robot that we acquired was the Aibo dog that was designed by Sony in 1999 and then it was released to the public a few years after.

And you could pet it and it would move around, but nothing much, but

it was enough to be really lovely and it was enough for you to have an endearing relationship with it.

I mean, right now robots are much more sensing, much more able to sometimes detect also emotions by reading facial features and able to have whether they are embodied or they're just a voice, they're able to really have a conversation that is much more, more real.

Our voice is such a key part of our identity.

So when we sing or when we speak, it's a very physical, personal experience.

And it translates into something that for Holly and Matt, can become almost disembodied and find a new form in the voices of others.

And that's a very beautiful transition, I would say, between the physical, embodied realm and the machine intelligences that they're using.

We are still at the dawn of the understanding of intelligence, both in terms of neuroscience as well as in terms of machine intelligence.

I do believe human intelligence and machine intelligence manifest in very different ways.

We not only survive, live, work, but we emote, we feel.

We have compassion, we have creativities that we don't know how to explain.

And to think about what human centered AI is, is to think about what we want this AI do for us.

Artists are able ask questions and ideate scenarios that the rest of us would never have thought of.

And in that sense, it's crucial for artists to take on technology at the highest levels, because we need those other perspectives and we need those other voices if we actually want these models to be as good as they could be.

You can't over engineer a potential outcome.

You actually do have to give models the ability to, you know, develop curiosity.

And this is why there's so much mistrust in AI because everyone's really unhappy with the way that the Internet turned out.

So if this is a new Internet, we want to make sure that we get it right this time.

Everybody has to have a say in AI it cannot be a tiny group of people telling the rest of the world, this is the AI we're creating and giving you.

That will cause a lot of unintended harms and consequences.

I feel that we need to be much more aware and conscious of our moves.

It's not only up to the artists themselves.

It's also up to the receivers of the art.

So the figure of the lonely artist that drops their art and lets the world deal with it doesn't really exist anymore.

We need artists really down in the trenches with us.