[music]

Alison: This is All Of It on WNYC. I'm Alison Stewart. Let's talk about the intersection of Artificial Intelligence and art. Right now, there's a piece on view at MoMA's first floor that uses AI to create digital artworks based on 200 years of MoMA's paintings. There was a work made with the help of AI on display at last year's Whitney Biennial, and that's just in the visual arts. New AI generators like ChatGPT can spit out a novel in seconds, or create an image from scratch in the style of Van Gogh just in the blink of an eye. A new song can be generated.

The future of the arts, and Artificial Intelligence, and how it'll affect human creativity is unclear. It does raise a lot of questions. Will AI-generated art just be a tool for artists? Is it a passing fad, or will it fundamentally change how we channel our creativity? Joining me now is Janelle Shane, a researcher, and author of the book, You Look Like a Thing and I Love You: How AI Works and Why It's Making the World a Weirder Place. She also runs the excellent blog, AI Weirdness. Hi, Janelle.

Janelle: Hi. Thanks for inviting me.

Alison: Listeners, what do you think about Artificial Intelligence, and creativity specifically? Have you messed around with any programs like ChatGPT to produce a piece of art, maybe a poem, a novel, an image? Call us with your experience, 212-433-9692, 212-433-WNYC. The phone number's 212-433-9692 212-433-WNYC. Of course, you can reach out to us on social media @allofitwnyc.

Maybe you're a full-time artist, writer, musician, or visual artist who has used AI as a tool for your work. What have you noticed? How have you used it? Maybe you fall on the other side of the creative arts, and you have concerns. We'd like to hear those as well. 212-433-9692. All right. Let's do some basics, if you don't mind, Janelle. People have used ChatGPT and DALL-E. You just input an instruction, and then moments later, there's a result. What's happening in that timeframe after you input, and then you get the output in that split second?

Janelle: Yes, so what you are talking to is basically, this big computer program that's been trained on a whole bunch of internet text, or in some cases, internet text and the pictures that appear next to them. Essentially, what it's trying to do is based on what it's seen on the internet training, try to predict what pictures would go along with the text you give it, or what would come next after whatever you wrote.

Alison: It's a lot about of its prediction.

Janelle: Yes. Is predicting trying to match what it's seen on the internet in the past. That's what it was rewarded for during training.

Alison: What is the impact of word choice on the end results?

Janelle: Yes, the impact is huge. If you are talking to ChatGPT or something like that, and you phrase your question very casually, you'll get maybe a more casual answer back. If you phrase it in a term of a joke, you might get a punchline back rather than a serious answer. You're building the first half of a webpage essentially, and it's filling in the second half, and what goes in the second half really matters. Depends on what you put.

Alison: I think people can intellectually figure out, okay, I understand that when it comes to the written word. How is the process different when it comes to visual imagery?

Janelle: Yes, for visual imagery still, you have certain images tend to appear on certain web pages in certain contexts. If you say, this is a picture my three-year-old drew you're going to get of a cat, you're going to get a way better thing than awkward-winning national Geographic photography of a cat in X background. It's accessing all of these different domains. That's been one of the challenges when you're trying to work with this thing, and try to get a particular visual effect is that it's seen a lot of terrible images in its internet training. You are simultaneously trying to direct it toward the images you want and away from all the terrible stuff that you don't want.

Alison: You used ChatGPT on your blog, AI Weirdness specifically using GPT-3, and you wanted GPT-3 to make Valentine's Day cards. Can you walk us through the creation of the Valentine's Day card, and what you wound up with?

Janelle: Yes, I decided to see if I could get it to do these roses are red violets are blue rhymes. As I've described, what I end up doing is setting up, in this case, it was a fake listing for a card somebody might be selling. Listing is for one antique Valentine's Day card. Interior reads, roses are red, violets are blue, and that's where I stop, and I say, "Okay, fill in the rest of it." I gave that task to various text-generating algorithms to see what I would get.

Alison: Would you share some of what you got?

Janelle: Oh, sure. Yes.

Alison: They're hilarious.

Janelle: Yes. I got roses are red, violets are blue, I would like to say I adore you. 8C one half of five-ampersand..2G [chuckles]. It specified that this card is illustrated with a leaping donkey, cornflowers, a butterfly, and foliage in the background,

Alison: A leaping donkey for Valentine's Day [laughs].

Janelle: A leaping donkey, so traditional.

Alison: Yes, right. What have you observed about the creative side of AI through this experience? Want to share your observations?

Janelle: Yes. I actually found I liked working with the smaller, earlier algorithms better. I worked with an algorithm called GPT-3. That's been the creator's OpenAI have worked on that algorithm over the years since it was first introduced in 2020. The 2020 algorithm is a lot better at being creative, and interesting and generating something that's not too generic internet.

There are smaller versions of those algorithms, and the smaller versions are really fun. The very smallest version of GPT-3 said, roses are red, violets are blue, silent evening is full of regret. In the war room, everything is violet and tinsels formaldehyde. Satire is the thing, especially when it lacks the lampshade. That card was to be illustrated with a trifle gone to seed on the board.

Alison: That was a lot of [laughs] That's a perfect card for somebody somewhere. My guest is Janelle Shane. The book is called You Look Like a Thing and I Love You: How AI Works, and Why It's Making the World A Weirder Place. Her blog is AI Weirdness. Our phone lines are pretty full. Let's take some calls. Stan is calling from Brooklyn. Hi, Stan.

Stan: Hello, how are you.

Alison: I'm doing great. You're on the air.

Stan: Yes. I've been a drummer all my life, and when I got to New York City, the studio work was Happening. In the late 70s, in the early 1980s, they started to develop extremely good drum machines that you couldn't tell the difference between a real drummer, and a machine. What happened, I would estimate probably 90%, or 95% of the studio work disappeared immediately. Also, the invention of synthesizers that could sound very much like strings, violins, and so forth. Brass even. A tremendous amount of musicians lost studio work.

You would have to resort to incorporating technology into what you did. The musicians that survived would do that as well as the acoustic instruments would incorporate drum pads, drum machines, and all sorts of little manner of synthesized instruments that could be triggered, and so forth. That's how you'd survive. You'd have to incorporate it. Other than that, you'd be reduced live playing mostly [chuckles]. That's in a nutshell. It had a profound effect on musicians when the synthesis of sound came along.

Alison: Stan, let me ask you this. I can remember having a conversation with the songwriter who got a little defensive because I mentioned that a lot of songs, especially pop songs, tend to sound alike. He was saying to me because they're all using the same drum sounds, and pieces of certain digital drum pieces are being passed around. To your point, not only are people losing work, which is upsetting, I felt like aren't we losing some creativity if we're just passing around these sounds.

Stan: Yes. This is the thing. All these loops and sampled sounds are available, and anybody can get on a computer, and put together a track. That sounds extraordinarily good. Yes, you're right. There's a drummer called Clyde Stubblefield who was famous for doing the Funky Drummer beat. As that thing has been sampled and used 1,000 times, that would not be an exaggeration. I don't know what to say.

Alison: Thank you. No, You've given us a lot to think about. I really appreciate it, Stan. Thanks so much. I want to follow up with Janelle, from your points, appreciate you calling in. Is AI making music or is AI mimicking music?

Janelle: People are using AI to make music. I think you always have to keep in mind the person who's choosing what prompt to put into the AI or the person's choosing what music to train the AI on. Humans are not out of the picture. Even if the thing is turned on generating Bach 24/7, somebody decided to train it on Bach or decided whose interpretation of Bach should go into that. Are these things better than the human? It's interesting is the human-generated things.

That's another question. We've definitely seen in the history of various kinds of automation and artificial substitutions that something doesn't have to be necessarily better than what it's replacing.

Alison: Let's talk to Darren from Weehawken, New Jersey. Hi, Darren, thanks for calling in.

Darren: Hi, thanks for having me. Back in July, I used Midjourney to generate some images. My father had a depressive episode and I wanted to send him something maybe meaningful, but I don't have the artistic talent that the images that stood out. I wanted to personalize it for him. The prompt was something like, my father's a sailor, did a lot of sailing and said I aprreciate something to the effect of each man on a sailboat breaking through the clouds after a storm, epic 4K cinematic. It was just a way for me to try and quickly do something a little bit more meaningful than [unintelligible 00:12:23] all mediums that we were attempting to [unintelligible 00:12:27]

Alison: Darren, that's very creative of you. Darren, hope your family is doing well. Thank you so much for calling in with the candid story. My guest is Janelle Shane. The name of the book is You Look Like a Thing and I Love You: How AI Works and Why It's Making the World a Weirder Place. We'll talk more about AI and creativity. You can call in and share your stories. 212-433-WNYC, 212-433-9692. We'll get two more calls and have more with Janelle after a quick break. This is All Of It.

[music]

Alison: You're listening to All Of It on WNYC. I'm Alison Stewart. My guest is Janelle Shane. She wrote the book You Look Like a Thing and I Love You: How AI Works and Why It's Making the World a Weirder Place. Her blog is AI Weirdness. We're taking your calls. The conversation for the rest of the hour is about the intersection of AI and creativity. Janelle, you can type in, you say, "Hey, write a song in the style of Beyonce with lyrics like the Beatles." It made me think of this clip that when I was at MTV was legendary. It was Vanilla Ice trying to explain why his song did not rip off Queen. This is 15 seconds of hilarity. Take a listen.

Vanilla: We sampled them from them, but it's not the same baseline. He goes ding ding, ding, ding, ding, ding, ding, ding, ding, ding, ding, ding, ding, ding, ding. That's the way theirs go, ours goes ding, ding, ding, ding, ding, ding, ding.

Alison: That one more extra ding which is my silly way of getting to legal and copyright implications, if you give these kinds of instructions.

Janelle: This is really interesting. There's some active lawsuits that are going on right now, which will be really interesting to see where they go. One thing that has been recently published about some of these newer image-generating algorithms is that they do tend to sometimes reproduce the images from their training data and a little bit more often than the algorithms were used to in the past where like, oh, it all goes into a big mash and you'll never see original images back again.

Now, not only is there a question of consent to be part of the data set but there is also a chance that you can get some of these very similar images back and the person interrogating the AI to make their work may not realize it

Alison: is this situation where the law is not going to be able to keep up because things change so quickly?

Janelle: It will definitely be a challenge. Part of that too is some of the people making the laws barely understand the Internet, let alone what's going on now with AI. it is definitely a challenge.

Alison: That it's totally fair. Jessie on Twitter wrote, Wired Magazine breathlessly claimed in a recent article that no artists will lose their job because of AI, accepting the fine art world commercial artists, designers, illustrators, photographers, animators are the most at risk. Look at what desktop publishing alone did to typesetters and pre-press workers. AI will trade convenience and blinding speed or actual skills, and probably for pennies, no more deadline weight or limit on revisions. Jessie, thank you for your opinion. Let's talk to Howard from Long Island City, who hosts a trivia night and did something interesting. Hi, Howard.

Howard: Hi. Good afternoon. I was calling just to give an example of how I use ChatGPT to write a trivia category and I was blown away by just what the results that it shot at me.

Alison: Tell us a little bit about it. What do you think about it? How did you react to them?

Howard: I thought I made a mistake and then I realized that the prompt that I gave ChatGPT was so that the AI was so smart that it provided wro-- I thought it had provided wrong answers but I told it to list 10 trivia questions for a crowd of people in their 20s and their corresponding answers, thinking that it was going to give me the questions and the right answers but instead, it gave me what I believe were answers that people in their 20s would probably respond to the question.

Alison: That's interesting.

Howard: Yes. One of the questions that said was, who was the first video to reach 1 billion views on YouTube? It said the answer was Baby Shark. At the trivia night, a few of the people said, "No, that's not true, it was Gangnam Style." Of course, I looked it up and it was Gangnam Style. Chat GPT actually gave wrong answers to questions that I guess I thought people in their 20s would get. I was just fascinated by some of the questions and answers that it had.

Alison: That is fascinating. Howard, thank you for calling in. Janelle, do you have a response to that or anything about that subject [unintelligible 00:17:43] [crosstalk]

Janelle: It will try to give you exactly what you asked for, whether you meant to ask exactly that or not. The other thing too is that these things tend to make up whatever sounds good and they're more rewarded for sounding good and sounding fluid than for being correct. Even had you worded it differently, you would have had to fact-check his answers for sure.

Alison: Let's talk to Vernon from Staten Island. Hi, Vernon, thank you for calling in.

Vernon: Hi. I'm both a musician and a visual artist. You got a caller a few calls ago who talked about drumming. I don't want to push back but I remember that time when the first drum machines came in. People were freaking out. What happened was that some drummers for low-level kind of work, things like where they just needed a basic beat, they may have been replaced but those drum machines really did not sound like drums. A lot of those drummers became the best drum programmers because they actually play drums so they know exactly where the hi-hat should be.

Say a drummer like Pumpkin who was a legend in the Bronx who programmed the beat for [unintelligible 00:19:10] MCs.

The technology has always been a part of music advancing. The thing that I'm concerned about is things like if I put in a really clever prompt to an AI, who owns my prompt because these are not closed systems. It's not like, you can buy the app and it's just on your computer or on your iPad. We're actually paying when we do the subscriptions. We're paying to train the AI. What happens to the intellectual property of our clever prompt that we think is so much different from somebody else's?

Alison: Vernon, do you play guitar?

Vernon: I do.

Alison: Is this who I think it is?

Vernon: Yes, it is who you think it is.

Allison: Everybody, this is Vernon Reid from Living Color [chuckles]. Hi Vernon. Long time no talk [laughs].

Vernon: I know. In fact, I'm actually going to be doing a performance tonight with James Blood Ulmer at MoMA. They're doing a whole thing about the JAM Gallery. Linda Goode Bryant's gallery. I'm going to be doing a live conduction in the style of Butch Morrison, Greg Tate with Burnt Sugar featuring James Blood Ulmer. That's an answer to AI. Do something completely random.

Allison: Vernon, thank you for calling in. Nice to talk to you again.

Vernon: Good talking to you too.

Allison: When we think about this, Janelle, we have a lot of calls about the legality, about the copyright issues, about who owns what. I'm not sure if you necessarily have the answers, but what questions do you have about ownership? About intellectual property?

Janelle: There's ownership questions all over the place. I think the wrong interpretation would be to say that the AI itself owns anything because it really is not thinking. It's just a prediction model. It's trying to predict what it's seen on the internet, but there's a lot of questions about who owns the training data, who chooses whether there's stuff goes into the training data without their consent. There's been this long tradition of people's writing on the internet or people's YouTube videos going into these big training data sets and many cases commercial training data sets without being asked before. I think I've seen more and more pushback against that. It's interesting because people claiming ownership over their photos that went into the training data really have some excellent points.

Allison: Let's talk to Daniel from Brooklyn. Slide in one more call. Hi, Daniel.

Daniel: Hi. I wanted to comment on the production of artwork from AI. I'm a visual artist. I'm a painter actually and when I see work done by AI, I immediately can tell pretty much, especially if you go up close to it because almost all of it is printed flat like a inkjet print. It doesn't have the ability to actually embody what it is that a human being can take from there in yourselves and put on a flat surface. It just doesn't.

Allison: Daniel, thanks for calling in. In our last moment, Janelle, what is a question you think people should be asking about the intersection of art and creativity?

Janelle: I think the question is what is acceptable as a substitute for human-made art. I think there are a lot of companies now that are trying to cut costs by pushing an inferior product and trying to see if that's going to get accepted. In some cases, you can't tell. In many cases, as his last caller pointed out, you can tell and the longer you look at it, the worse it gets. I think this will be really interesting to see how this develops and what is acceptable as a cheap substitute.

Allison: The name of the book is You Look Like A Thing and I Love You: How AI Works and Why It's Making the World a Weirder Place. Definitely check out the blog AI Weirdness. Check out those Valentine's cards. They are wild. Janelle Shane has been our guest. Janelle, thank you so much for taking our listeners' calls and sharing your expertise. We appreciate it.

Janelle: Oh, thank you very much.

Allison: That is All of It for today. Tomorrow we are live in The Greene Space with Jordan Carlos, author, Emma Straub, the Food Collective, Ghetto Gastro, as well as Yo La Tengo performing. I'm Alison Stewart. I appreciate you and I appreciate your listening and I'll meet you back here tomorrow, downstairs in The Green Space with an audience.

[music]

Copyright © 2023 New York Public Radio. All rights reserved. Visit our website terms of use at www.wnyc.org for further information.

New York Public Radio transcripts are created on a rush deadline, often by contractors. This text may not be in its final form and may be updated or revised in the future. Accuracy and availability may vary. The authoritative record of New York Public Radio’s programming is the audio record.