In this episode, we talk with Maya Ackerman, an AI researcher and entrepreneur at the forefront of AI-assisted creativity. She shares how her journey from machine learning research to opera singing led to the creation of Wave AI, a company focused on co-creative AI tools for songwriting.

Maya Ackerman explores how AI can be a true creative collaborator, rather than just a content generator. She introduces the concept of “humble creative machines”, AI that enhances human creativity while keeping artists in control. We discuss AI’s theory of mind, biases in generative models, and the future of AI-assisted songwriting.

Dr. Maya Ackerman is a world-renowned expert in artificial intelligence and creativity, focusing on co-creativity and its transformative potential across industries. Ackerman is a researcher, entrepreneur, and professor of AI at Santa Clara University. She is the CEO of WaveAI, the company behind LyricStudio, a widely used AI songwriting tool. Her work has been featured in Forbes, NBC, and NPR.

So welcome, Maya Ackerman.

2

Pleasure to be here.

3

Thank you for having me.

4

First of all, you have a unique background that combines computer science and music.

5

I would like you to maybe begin by telling about your journey into AI and the intersection

with songwriting and kind of what inspired you to explore this blend.

6

Yeah, I studied machine learning in graduate school for my master's and PhD theoretical

foundations of cluster analysis, very math heavy.

7

Honestly, never intended, it never entered the field of AI because it was popular, it sort

of in a less popular branch of AI back in the time of machine learning, looking very much

8

at the mathematical foundations.

9

happened that I came across an opera singer who lived in Waterloo where I studied and I

started taking lessons from her and

10

opera singing.

11

I started performing semi-professionally.

12

So this happened in parallel to my PhD, which was really fun.

13

And I never really thought that I wanted these two worlds to collide.

14

I kind of felt that they needed to stay separate.

15

But in 2014, I attended a workshop called Information Theory and Applications.

16

It was in San Diego.

17

And there was this little session, little three talk session called computational

creativity.

18

And I remember making a mental note to attend it.

19

And there was Harold Cohen screaming about how people were calling his automated painter

creative.

20

But really he's a creative one.

21

He's a creative mind.

22

and I did not have a side, but I thought it was fascinating that there was a machine

painter and somebody was thinking about whether it was creative.

23

And there were a couple of other amazing speakers in that session, Jorain and Mark Riddell

and Jorain Wiggins, and they told me about computational creativity as a field and I just

24

fell in love with it.

25

I just fell in love with...

26

the idea of machines, creating machines, making art machines, helping people make art and

very very quickly came up with the project that will end up becoming my company which was

27

originally coming up with vocal melodies given lyrics.

28

So if I say something like...

29

I'd love to visit Norway, let's say, then how many ways can you sing it?

30

I mean, it's a universe, right?

31

Like, I'd love to visit Norway or I'd love to visit Norway.

32

Right?

33

So it helped me explore these possibilities.

34

And the rest is history.

35

So you kind of came from the technical side first and then into music kind of

simultaneously.

36

Yeah, yeah, kind of, you know, some people say that their past found them.

37

I definitely tend to feel this way, that my past found me.

38

And you were just starting to talk a bit about it that you started the company and started

the made kind of a tool to to generate melodies.

39

Could you tell a bit about a bit about that company?

40

Yeah, so it started off with that specific model.

41

So we built a machine learning model that did that, took lyrics and came up with different

ways to sing them.

42

And I found out to be so helpful.

43

I still remember we printed it on paper and we had this kind of funky notation because we

just...

44

The interface didn't matter, right?

45

We didn't need to serve it to users.

46

And I remember sitting by my piano and sort of deciphering that notation and playing it

out on the piano.

47

And almost instantly, like my songwriting was transformed.

48

Suddenly I could see so many more melodic possibilities than before.

49

It was clear pretty quickly that we have something here that could help more people.

50

But it still took us about three years to actually open up a Silicon Valley startup.

51

I remember sitting by my computer and filling out the paperwork to open up Wave AI and it

said company name and I'm like, we didn't even think about the name and I came up with

52

Wave AI sort of on the spot.

53

in 2017 before all the companies were something AI.

54

So that was really fun.

55

But overall, it's been a complicated journey.

56

It has not been straightforward at all.

57

Raising money, running a company, hiring people, advertising.

58

There is so much to learn to running a business, even if you have incredible technology.

59

It's a very steep learning curve.

60

Yeah, I can imagine.

61

It's a whole new world.

62

But what about the technical and creative challenges when designing an AI to be a

collaborator?

63

What kind of challenges did you face?

64

You know, I think it really expedited having a business, serving real users really helped

shape how I think about what AI should be.

65

I was always into co-creative.

66

I loved the idea, like instantly.

67

I loved the idea that a machine can be a co-creative partner.

68

I heard this phrase at the International Conference on Computational Creativity.

69

Somebody said it in passing during the first talk that I attended and

70

I was just, I mean was sold at that moment.

71

Wow, yes, of course, a machine can be a co-creative partner.

72

How did I not think about it myself?

73

But that could mean so many different things.

74

It could mean that it does most of it and then I change it up a little bit.

75

Or it could mean that I do almost everything and it just helped me in those two little

areas.

76

Or it could mean everything in between.

77

It could mean that...

78

Let's say if I'm writing a song, could mean that it helped with mastering and mixing and

composition and lyrics and vocalization.

79

Or it could mean that it just helped with one of them.

80

It let's say the lyrics.

81

Or it did all of those things, but just for the hook, right?

82

I it's just like working with another person.

83

It could mean so many different things.

84

And the more we served real people, the more I started understanding that the person needs

to be in the driver's seat.

85

And ultimately, it's all about, if done right, if done right, it's all about elevating

human capability.

86

So that even if you step away from the machine, you are now more capable, more creative or

smarter than you were before in that domain.

87

Yeah.

88

And does Wave AI do that for you?

89

So our sort bread and butter, most successful product is Lyric Studio.

90

And that was actually our third product.

91

So we started off with tools that, well, in different ways, we hit some stuff right, some

stuff not quite right with our first two products.

92

And Lyric Studio helps only with lyrics.

93

And...

94

It does it really well.

95

It does it in a way that makes people better pen and paper lyricists.

96

And you know, this takes a certain attitude towards your users.

97

This takes a certain trust.

98

So our product is helpful for people all across the expertise spectrum.

99

It's helpful for beginners and it's helpful for people who are...

100

stars or at the extreme end of professional because it's a very very flexible system it

can inspire you or it can hold your hand more it's sort of anything in between but we

101

don't we sort of don't foster dependence in a way that perhaps other

102

some other AI products try to do.

103

So we are okay with the fact that some, that all of our users really become better at

lyrics writing.

104

And a lot of them choose to stick around even though they're better because the product is

still helpful.

105

And it's, it's a certain attitude towards our users where we, we want to build this kind

of, if you will, healthy relationship with them, where we want to empower them.

106

We want to make them better.

107

And we trust that those who, that enough of them are going to stick around, right?

108

That we don't need to.

109

make them dependent on us.

110

We don't need to make their songwriting impossible without Lyric Studio.

111

We don't want that.

112

Kind of like a little analogous to a healthy partner.

113

An unhealthy partner might want to make you dependent on them, get rid of all your

friends, make it so that you can survive without them or something.

114

All kind of abusive patterns.

115

Whereas a healthy partner will want you to be strong and independent and capable and

happy.

116

and trust that you will keep choosing them essentially and that's kind of a vague analogy

to the approach that we're taking that I really think more of the industry needs to take.

117

Do feel like that's the difference between using Lyric Studio and using Chat GPT for

instance?

118

So I actually have a very positive attitude towards ChatchebikT just for other use cases.

119

I love ChatchebikT.

120

I'm an avid user, I'm a paid user.

121

And I think it's precisely because for a lot of use cases, you get this kind of

experience.

122

I call this broadly Humble Creative Machines.

123

So for example, if I'm writing, you know, text and not poetry but prose, it can be really

fun to experiment with different styles, to ask it to modify something.

124

It's amazing at coming up with titles.

125

And in many contexts, you can have this kind of iterative experience with Chat-Shop-E-T,

but with lyrics,

126

It's just not, I guess it just hasn't been a focus area.

127

And fundamentally, when you're writing lyrics, if you are doing it as an art, right, then

you need to be creating.

128

And Chachi Pt sort of creates a whole thing for you all at once, which is a little too

much.

129

Imagine you're collaborating with a person on lyrics and you're saying,

130

I want to write about a sunny day, a sunny December day in California, right?

131

It reminds me of some childhood experience.

132

And your friend just writes the whole thing right away.

133

You're not really writing together.

134

It's too much.

135

And it might not be even the style that you want.

136

Chachapiti is very rigid in its lyrical style.

137

And again, I don't mean this as a profound criticism of Chachapiti.

138

It's not a core use case for them.

139

They focus more on business use cases and other things.

140

So yeah, it's very different from Chudge GPT.

141

Our product is really geared for that lyrics writing experience, whereas Chudge GPT has a

lot of other use cases that it's just phenomenal at.

142

Yeah, that's really interesting.

143

fact that it could provide too much information as you were saying, it's like, yeah, if

you collaborate with someone and they really loudly and takes kind of takes the word for a

144

long time, long time.

145

you get, I guess you will in Lyric Studio, you will get

146

shorter samples of sentences or words that are more inspirational.

147

line at a time model.

148

We do have something that can draft something a bit longer for you if you really want that

support, which sometimes you do to kind of like get going.

149

But it's a very simple looking experience, right?

150

So by the way, users like simple and for good reason, right?

151

They already have we have enough complexity in our lives.

152

User interfaces need to be simple and straightforward.

153

And so you have your notepad, which takes more space than anything else, inviting you to

focus on your own writing.

154

And then on the side,

155

You suggestions, one-line suggestions that take into account everything you've written,

the topics you've selected.

156

And you can tap into them at any time, into these suggestions.

157

You can also ignore them at any time, which is really, really important.

158

You don't want to feel like at every moment you need to be thinking about the AI, because

once you tap into your own creative process, and our users tell us that a lot, one line

159

can sometimes get two thirds of the song done, because it sends them...

160

on this

161

Does it mean that Alex wrote a song for me?

162

Probably not.

163

Does it mean that I gave it a prompt and then it wrote a song for me?

164

Well, no, that would be really weird, right?

165

It also doesn't mean that I'm now dependent on Alex.

166

If Alex made me more creative, I became more creative, even if I never see Alex again.

167

Which means that there is a certain, Which means that...

168

My creative process, my capabilities were front and center.

169

And so you need to create this sort of experience with AI if you want to genuinely benefit

people.

170

I mean, we are a small company and we are able to have a lot of users in a landscape that

has a lot of AI tools.

171

And that really shows how important it is to prioritize your users.

172

And really, there are long-term well-being, if you will.

173

In terms of AI being a creative collaborator or inspirator, are there any areas where you

find AI to be lacking in terms of creative potential?

174

everywhere.

175

Everywhere there's possibilities.

176

I'm kind of simultaneously in awe of what other people are building.

177

And at the same time, I see this this potential to make the system so much more useful by

focusing more on empowering people and less on demonstrating how cool AI is.

178

I call it humble creative machines.

179

You can have the most brilliant human being, but

180

Sometimes it's more appropriate for them, like for example as teachers in the classroom or

as managers.

181

to pull their genius back a little bit and utilize it to enrich and help elevate other

people.

182

And that's really, really key in human society.

183

And we actually get very upset at other people when they're too arrogant, even if their

brilliance, quote unquote, justifies it somehow, right?

184

And so that's what's happening with AI.

185

mean, it's genuinely amazing what it can do.

186

I look at text to image models.

187

But...

188

If you haven't, by the way, you've got to try mid journey.

189

I I just have no connections to that company, but the technology is incredible.

190

And I love it.

191

I love it so much that I use it.

192

I use it even though it's hard to use, but it's really not designed with the idea of

giving a person control.

193

They eventually built in in-painting, but even that's still difficult to use.

194

And you can sort of take a picture and ask it to modify it, but it modifies it how it

wants.

195

More than half the time it doesn't understand what I want.

196

And that despite being a wonderful system that I love, I love Mid Journey.

197

Again, I'm a paying user of that system, but there's so much left on the table because the

AI needs to be built from the ground up in a way that considers

198

that lets the user become the driver.

199

And that's missing.

200

And as a result, it's missing a lot of commercial opportunity and a lot of opportunity to

enrich human lives.

201

Why is that the case, do think?

202

I think it's just a way of thinking about things.

203

know, I think sometimes as consumers, and I'm sure that happens to me all the time in

industries where I'm not an expert, we just assume that it has to be this way.

204

It's this way because there's no other way to build it, but it's just not true.

205

It's just not true.

206

I'm not saying it would be easy to take something like a text to image model and make it

and put the user in a driver's seat and make it easy to interact with that system and make

207

it so the user could realize their vision.

208

more completely and be understood by that model.

209

I'm not saying it's easy, but it's possible.

210

And so from a business perspective too, think business leaders, investors need to begin

appreciating that Chant'ch'ipiti succeeded because it enables these experiences for

211

people.

212

Before we had Chant'ch'ipiti, we had these other...

213

other models called like large language models.

214

We had, for example, GPT-2, you would give it a line of text and it would come up with a

paragraph to follow up that line of text.

215

And it was fun, but you couldn't really interact with it.

216

You couldn't explain your intent.

217

You couldn't iterate with it on a block of text.

218

And so, you know, it really wasn't that popular.

219

It was somewhat popular, but it wasn't chat GPT popular.

220

And then they did this alignment process, which was a significant investment.

221

Alignment is difficult and complicated, they made a massive investment in that.

222

And they made it into something that humans can interact with.

223

And it blew up.

224

It blew up.

225

So there is evidence that making the communication between human and machine front and

center, giving the human control, can lead to massive success.

226

And we need more of that.

227

Yeah.

228

In my own experience with the chat GPT as well, that's really the case after the

possibility of making custom GPTs came out.

229

That's really when the usage of it exploded on my part.

230

Because for instance, with these podcast episodes, I can now transcribe the episodes and I

upload the text to one of my custom GPTs.

231

And for the first 15 episodes, I wrote my own show notes and intros and stuff like that.

232

And I also uploaded all those intros and then I can use this custom GPT to generate show

notes for new episodes based on the transcript and my earlier show notes.

233

So it's like in my style, but also it takes into consideration all the information of each

episode.

234

which is really handy.

235

And it also kind of speaks to what you're saying.

236

It's like the dialogue and the customization aspect that really makes it interesting.

237

And that's probably something that's missing in a lot of the generative AI music tools.

238

Although it seems like they're trying though, because both Suno and Udio has in-painting

options.

239

Although they don't really work that well in my experience for now at least.

240

It's like...

241

At least for professional musicians, kind of desire even more control.

242

I need to have like fingertip control of every aspect, including like which frequency

should be cut out on the snare, on the bridge, the second time around.

243

Like it has to be that detailed.

244

So in my own experiences and also my students who have tried out these

245

software is the best possibility for now is to use them more as a sample database because

then you still have control afterwards.

246

Mm.

247

Yeah, I like what you're saying.

248

think it's I think what you're saying is the rule, not the exception actually.

249

This experience that you're having of like, you're not giving me enough control.

250

is the common experience of frustration with AI system, which actually tends to lead to

low retention.

251

So the whole AI industry is notorious for low retention, kind of like one hit wonders.

252

People use it one, two, three times and then never come back, often just once.

253

And so giving users control is how you overcome that.

254

And I think companies are not taking this issue of control seriously enough, even though

everyone complains about it.

255

Well, everyone, just about everyone complains about it.

256

And it needs to be taken so seriously that it should impact how you build your models from

the beginning.

257

And when you're not doing it, it's like taking a cooked dish and trying to improve it

afterwards.

258

Forget about it.

259

Forget about it.

260

It's too hard.

261

That's why you're saying it's there, but it's not really there, but that's not good

enough, you know?

262

And I have so much respect for what Sunoo and Uryu did, just like for Mid Journey.

263

So I really mean it in a kind of like, I mean, let's do it even better.

264

We have something really cool.

265

Now let's make it useful for people.

266

Yeah, the cooked dish analogy really hit it with me.

267

It's like when you cook a meal and it's too much salt on it and it's like you can't go

back on that.

268

Or at least it's really difficult.

269

Like, let's think about a baby, right?

270

We don't teach the baby how to be smart and then teach it to interact with the world.

271

It learns everything by interacting with the world.

272

Interaction is central.

273

And these models are taught to be smart.

274

And then the makers are like, my god, now I need to give it to users.

275

And these users keep asking for more stuff, right?

276

So kind of, we just need to think about it correctly.

277

And when we think about stuff correctly, then we can serve people better.

278

How would you go about doing that building up a model from the start that takes this into

consideration?

279

We have an approach that we figured out.

280

Just to make sure we don't get disturbed again here.

281

We have an approach that we figured out for that.

282

That's definitely confidential.

283

But I think the real kind of thing that's missing in the industry is really just there's

plenty of smart people out there who can figure out solutions.

284

I have no doubt that there is sort of a lot to discover also still left and a lot of

different approaches that can be explored on the technical side.

285

But it's really mostly the intent and desire to take that seriously enough to make massive

investments in that dimension of the tools that is sort of the missing element and not so

286

much the technical know-how is something that's...

287

That's the trade secrets.

288

Yeah, I can't share it on a podcast, it's unfortunately because it's, you know, this

relates to my company.

289

But this is something that I think is something that we as an industry can figure out

across many, many different domains as we, mean, OpenAI figured it out for for ChaiGPT how

290

to align a large language model.

291

And that was

292

That was a massive accomplishment.

293

And in other industries, I the complexity level is similar.

294

It's not, in some sense, it's a little harder, but in some sense, it's actually a little

simpler.

295

Apology.

296

I can't provide something meaty for this question.

297

I recently spoke to another computer scientist called Cheyenne Dadman, who has made a

framework for this kind of endeavor, which really shortly sums up, has something to do

298

with using reinforcement learning and including the user in the loop.

299

So that personal preferences also is a part of the selection process when it comes to

training the model on data.

300

Okay, that sounds reasonable.

301

Yeah, I mean there is...

302

it's very tempting for me to just share everything with you.

303

No problem, no problem.

304

We can talk about it at a later time when it's not that hot potato anymore.

305

Maybe a couple of years or something.

306

So I've read some articles that you've written and in one of them you discuss something

called a theory of mind.

307

which is a of a framework for understanding biases in AI.

308

And for those who hadn't read that article, could you sum it up and explain what it is?

309

Yeah, this is such a cool pivot, such a fun topic.

310

I've spent the past year thinking about, in a more fundamental level, sort of, let's zoom

out.

311

Let's zoom out, okay?

312

Zoom out of the details of our lives and our personal struggles and look at it, at what's

happening as a moment in humanity, right?

313

As a moment in human history.

314

So we had, you for a long time, human beings had an oral history, but we would pass on our

knowledge by talking to our children, right?

315

And then people started writing down stories and their thoughts and their knowledge

started becoming written down.

316

And then we put it up on the internet.

317

Okay.

318

And then we built a machine brain.

319

based on this knowledge that we put on the internet, right?

320

And this machine brain is us.

321

It is really an embodiment of what Carl Jung called collective consciousness.

322

And that's wild.

323

That is, that at a conceptual level opens up

324

possibilities that were just unthinkable before.

325

In some sense, when you interact with Chat-Chapi-Tea or Mid-Journey, you're interacting

with a collective consciousness.

326

You're not just interacting with OpenAI or Mid-Journey, with not just a product of this

company, you're interacting with a mind that learned from a huge part of humanity.

327

It's essentially found a way for us to bond with each other, to work together in a way

that was just impossible before.

328

everything that's amazing about these systems, like their intelligence and their

creativity, is our collective intelligence and creativity.

329

No human being can ever hope to consume such amounts of data.

330

No human being could ever hope to learn from such amounts of data.

331

But now we kind of have it in our fingertips, which is why it's so important.

332

that we make these models easier to interact with, because that's sort of the remaining

barrier for some of them.

333

But there is a flip side, and I think we sort of, somehow we to open our hearts and be

okay with that flip side, because it's just reality.

334

And that's the shadow, it's what Carl Jung called the shadow.

335

And each one of us has a shadow.

336

We have this...

337

This part of ourselves that we prefer to deny.

338

It's a part that was unacceptable in our family or in our culture, in our society.

339

And it could be even a talent.

340

Maybe somebody is amazing at sports, but they're in a family that really devalues that and

only values academic strength.

341

And so that person had to put that talent in their shadow.

342

Or maybe it's anger.

343

Maybe your family didn't let you be angry.

344

Maybe you're a strong woman and strong women are not allowed in your society.

345

Right?

346

So it's kind of like the shadow can have all kind of traits that are not necessarily good

or bad.

347

They just weren't allowed.

348

They're just not acceptable in your world for whatever reason.

349

And it's the same with the collective.

350

Right?

351

Our world is

352

incredibly sexist and racist and ageist.

353

mean, it's just a fact.

354

If you look at people who study this, this is not debatable in reality.

355

But it's also, these are all sound acceptable traits.

356

So we think we don't have them.

357

We have them while thinking we don't have them.

358

So they're in our shadow.

359

And because these models are trained on our data, they reveal our biases.

360

Very creatively, if you look at the text to image models, it's really fascinating if you

tap into this, how creatively it visualizes our biases.

361

And so instead of saying, how dare the machine have biases?

362

Who do we blame?

363

Let's yell at the developers and the entrepreneurs and the CEOs, right?

364

Which is kind of the reflex reaction.

365

I think what's even more profound, while improving the AI is a nice goal, but it's very,

very complicated for many reasons.

366

The more fundamental thing, really, what we should be thinking about first, thinking about

the AI should come second, we should be thinking about first, is how do we improve

367

humanity?

368

And it's kind of like an opportunity for us to look in the mirror, recognize the bias

within ourselves, so we can have a better humanity first and a better AI second.

369

Yeah, that's

370

And do you think, or how do you think this mirror can help humanity get better?

371

You know, we want to run some experiments on that.

372

I'm trying to think, which example should I give you?

373

I've done quite a bit of research on this.

374

So one example, kind of the most, kind of the simplest to share, I guess, is our study of

brilliance bias.

375

So I mean, I invite you and the audience to think for yourself, who comes to mind when I

say the word genius?

376

Most likely most of you thought of Albert Einstein initially, but if not, you probably

thought of some other guy.

377

That's because our culture teaches us that intellectual brilliance is a male trait.

378

When we think of a genius, we think of a guy.

379

Then if we think a bit longer, it can be like, yeah, you know, there are some smart women

too.

380

But it's a kind of like afterthought.

381

And it actually goes deeper than it appears.

382

It developed around age 6, were extensive studies done, the great majority of people

developed brilliant spies at age 6, it's prevalent, it's kind of ubiquitous, it's in the

383

water, it's in the air, and it causes serious problems for women trying to enter careers

like musical composition, computer science, math, physics, anywhere where any jobs that we

384

think geniuses do.

385

We tend to favor men and women are kind of taught that these jobs are not for them.

386

So it's real.

387

It's a real thing in our society.

388

It's not very well known, but it's a very serious bias and it's a really big problem for

women and for humanity that kind of misses out on a lot of intellectual brilliance that

389

can come from the other half of humanity.

390

Could just jump in there for a second, because you say the brilliance bias starts at about

six years of age, and you've just done studies on this.

391

How are the studies done?

392

Do you like ask six year olds, which person do you think is the smarter one, the male or

female, for instance, or how was it conducted?

393

Yeah, good question.

394

I love these studies.

395

It's such good work.

396

So one thing is pick groups.

397

We have a game, like say you're a teacher and you say or the scientist says, we're going

to be playing a game for very, very smart kids, really, really smart kids.

398

You get to pick teams.

399

So you pick teams for very, very smart kids.

400

People mostly pick, boys get picked first for those games for very, very smart kids.

401

Another variation was

402

I it might have been done at a slightly different age.

403

But another variation is asking people to draw very, very smart people.

404

And when you look at kids who are about age five, they draw their own gender.

405

So girls draw girls as very, very smart people, and boys draw boys.

406

And then once you pass the age six, seven barrier, everybody starts to draw boys.

407

Yeah, so because it just society teaches us, remember in my own family, you know,

watching.

408

taught that at exactly like the age of six?

409

That's a valid question.

410

What happens?

411

There's a lot of...

412

Yeah, you start...

413

Well, a lot of kids are in kindergarten before that, but I definitely remember with young

people in my own family watching them start off without sexism and watching sexism get

414

poured into their heads.

415

It's...

416

Part of what makes sexism and racism and ageism and all that stuff so effective is that

society pretends it's not there.

417

We're told that we've made progress.

418

We're told that things are not so bad anymore.

419

But it's just not true.

420

It's just not true.

421

It probably varies a little bit between cultures.

422

I don't want to make it completely blanket.

423

I you're from Norway, probably one of the places that kind of do the best job at equality

for women.

424

But if you kind of look at my own experience in the United States and some other places

where I've lived...

425

Kids are taught that the boys are smarter.

426

They're taught that very explicitly in many, different ways.

427

They get that message from classmates.

428

They get that message from teachers in many ways.

429

It's just something that just gets hammered in over and over and over and over again.

430

And the family, even if you're trying to raise your kids without these biases, it's almost

impossible.

431

It's so pervasive in culture and it's and you watch a little child who is born without

this stuff And then it just gets shoved into their heads.

432

It's terrible.

433

It's terrible.

434

There's a reason why These are serious issues these are very very very serious human

issues and it's the more you start studying them and kind of reading the research and more

435

you're like my god, it's it's everywhere and it's all the time

436

And we don't see it because we're fish in the water.

437

Because we're just so used to it.

438

I a statement like, girls are just as good at math as boys is incredibly sexist.

439

Because it suggests that boys actually have some kind of an advantage.

440

And that's better than some of the stuff you hear.

441

So we started looking into AI.

442

And when you type the word genius into mid-journey, you mostly get men.

443

And so we did a more extensive study on that and it's, yeah, most of these models have

brilliance bias to very, very high degree.

444

Yeah, but in the process, it kind of gives an opportunity to self-reflect, to ask

yourself, what do I think of when I think of the word genius?

445

To kind of like remove the shame and just be honest with yourself, you know?

446

To what extent do I imagine the same thing that the model imagines?

447

Am I shocked to see the model create any women at all?

448

And I think if we're honest with ourselves, there's really an opportunity for growth here,

because the only way that you overcome biases is to become profoundly aware of them.

449

And then you can make a choice.

450

Once you know what's inside your head, you can start making some conscious decisions.

451

Yeah.

452

So when you experience the model as a mirror of a collective human consciousness, kind of

exposes the biases.

453

So it makes it easier to see.

454

then you don't have to, you don't have to take ownership of it personally either.

455

So then maybe it's less shame attached to admitting it.

456

Exactly, mean the shadow really operates by shame, right?

457

We cannot take full responsibility for the biases in us.

458

It's not fair.

459

We did not invent them.

460

They get shoved into our head.

461

If we really, really pay attention, the messaging is continuous.

462

It's every single day.

463

And so...

464

Yeah, there is no shame in having these biases.

465

But there is pride if you're willing, there should be pride if you're willing to look

within and admit them within yourself and willing to put in the work to notice it within

466

yourself so that you can make better decisions.

467

continuously wrapping it in shame, telling people they're terrible for having these biases

is...

468

It's just perpetuating the problem.

469

It's a collective problem.

470

We as a society have this issue.

471

It's not something that individuals invent.

472

For the most part.

473

I you have people in the extreme who work to perpetuate them, but that's really the

exception.

474

Usually the only way out is through.

475

And how do these biases show up in music and in the models trained on music?

476

that's such a good question.

477

Music is one of the most honest facets of culture, right?

478

You want to know about a culture?

479

Listen to its music.

480

And so it's all of these biases are in the music.

481

And the lyrics, the lyrics in particular are sort of...

482

which is why co-creation is better in a way.

483

At least put the user in a driver's seat so they can be inspired to go in direction that

they find most meaningful.

484

Yeah, that's a good question.

485

I've never actually done analysis specifically on bias in music, in generative music, but

I like that.

486

I think one of the biggest problems, one thing that sort of immediately comes to mind is a

little different from the kind of biases we were just discussing, but sort of a lot of the

487

models tendency to converge towards the mean.

488

So if you look at the lyrics in Chat-Chip-A-T, for example,

489

ChachiPT does a really good job filtering out very, very obvious sort of discriminatory

biases, but it's not very good at giving you diverse results.

490

for anything.

491

when it comes to creative tasks, one thing that's really important is to be creative, to

give different ideas, to go in different directions.

492

And this kind of like bias to go towards the mean, to give you the average, to give you

the most expected is really problematic when it comes to creative things.

493

We don't want a system that creates music to keep creating the same kind of music, to keep

reinforcing only the most popular styles, only the most popular choices.

494

And that's a different kind of risk that we

495

face with these creative systems.

496

And it's actually not that hard conceptually to make them do something else, to make them

more exploratory.

497

But a lot of companies choose not to.

498

And I think that's another sort of mindset shift that we need to do.

499

Yeah.

500

Maybe the intuition for companies is to make something that is commercially appealing

instantly.

501

So like you hit a button and you get a catchy pop song that sounds similar to what you're

already hearing.

502

that sounds appealing for what you were mentioning earlier when it comes to retention that

the users might use it.

503

once or twice or thrice and think it's fun and then it's over because it's not new or

fresh anymore.

504

And yeah, I think we need to give more credit to our users.

505

We need to believe in humanity a little bit more.

506

Yeah.

507

In Norway, we have this organization called Tono, which is related to copyright.

508

It manages the rights of composers, lyricists, songwriters, and music publishers.

509

And they recently introduced some guidelines when it comes to working with AI as a

collaborator.

510

And these guidelines stipulates that only works with human creative inputs are eligible

for copyright protection.

511

And if a composition or lyrics are entirely AI generated without human contribution, they

can't be registered with Tono.

512

However, if a work combines human and AI generated elements, the human creator can

register their portion while the AI generated part is assigned to Tono AI with a specific

513

IPI number.

514

I mean, I know you're not a copyright specialist, but I was wondering if you could give

your thoughts on this approach and how it might influence the way artists and AI

515

developers collaborate.

516

This is-

517

Not this is probably on the better end of what I've heard copyright agencies trying to do.

518

At least it acknowledges co-creativity as a separate use case and it's open fundamentally

to different types of co-creativity.

519

So in a very fundamental way, I think there is something very intelligent about it.

520

The one kind of serious weakness that I hope the agency will eventually have a chance to

consider is just how difficult it is to properly separate who did what.

521

when you have a profound collaboration.

522

So just imagine trying to do it with two people to say exactly who did what.

523

If you really work together.

524

And even if I tell you, you you and I can work on a song, but then we'll need to tell an

agency exactly who did what, what kind of impact that could have on our creativity and our

525

ability to work together effectively, constantly tracking who did what.

526

So there is something in this proposal that sort of flies against the way that

collaboration works if you have a system that really can collaborate deeply and

527

meaningfully.

528

And so I think it has that real shortcoming and that it's going to be challenging for the

most well-meaning people if they're using a really good AI system that can really

529

collaborate with you.

530

It's going to be challenging for them to tease that out and in best attempts to be honest.

531

If the system does something like mix or master, right?

532

Like something one well-defined task, then it's easy, right?

533

Be like, I did the lyrics and whatever the melodies and it mixed or mastered or whatever

it did, right?

534

But very often some of the best

535

collaborative opportunities are much more interlaced.

536

Like let's say in my system you and the model write lyrics and sometimes you know when

you're working with a person they can give you one little idea but it was actually really

537

big.

538

We have it all the time in academia when we write papers it's so hard to tease apart who

did what.

539

And so I think that actually that's actually a real hindrance to embracing

540

quality collaborative AI systems.

541

I know that that was of course not the intent of the agency, but it's that's something

that we need to take seriously.

542

Yeah, so that's a challenge.

543

In human collaboration between musicians, are different approaches to this kind of task.

544

And if you're going to bicker about who did what, then yeah, this riff was mine, but yeah,

you played an earlier iteration of it and it might lead to a lot of conflict and probably

545

stuff happening that is kind of a hindrance of creativity.

546

So most musicians I talk to and play with usually agree beforehand that it's going to be a

50-50 split no matter what happens.

547

Just because then it's out of the way, you don't have to think about it while

collaborating or composing.

548

And I find that to be really the way to go because even though, let's say I...

549

I'm supposed to compose something with a musician friend and we're sitting in the same

room and I'm the only one playing, the other person isn't doing anything.

550

It would still impact me just the fact that the other person is there.

551

You could get kind of psychic about what happens if it's like some sort of non-physical

thing happening or...

552

Or if it's just the mental part of having a presence of another person that makes it, but

it certainly is an effect.

553

So I guess what I'm trying to say is that would probably be the case with collaborating

with AI as well.

554

So as you were saying, it's impossible to detangle.

555

So I'm just imagining this.

556

Cause I've composed a good deal of music and registered it with Tono and it's completely

trust-based.

557

You open a webpage and then you write, okay, this song lasts for this and that long.

558

And I have 50 % of the rights and this other person have 50 % of the rights or something.

559

And with an AI would do the same.

560

I could hardly see myself being incentivized to giving the AI a fair share of the song,

even though I know I used it in the process.

561

And I'm not sure if that's the way to go anyway.

562

At least if one were to look at it merely as a tool, because I don't give any writing

credits to my guitar.

563

either or writing stuff

564

I'm actually in full agreement with you.

565

think if you have really good tools that you can really collaborate with where this sort

of teasing apart is not even possible, then thinking about it as a tool and trusting

566

artists that they're engaging meaningfully is long term, that's the way to go, I think.

567

Just because anything else...

568

It's just not realistic.

569

It either blocks creativity or it...

570

People are gonna end up being...

571

Even honest people are gonna end up being unable to be transparent in this kind of context

because it's just so difficult to be transparent when it comes to collaboration on exactly

572

who did what.

573

It's so hard to know for anybody, you know, listening who's meaningfully collaborated on

anything with another person.

574

mean, they say if you ask people what percent of the project they did, let's say you have

three people who worked on the project, you ask each one how much percent you did, you

575

hope you most of the time end up with over 100 % when you add it up.

576

It's just, you know, when you have this meeting of the minds, whether it's with another,

with a machine or another person, you just gotta let, you gotta let, you know, one plus

577

one be bigger than two.

578

You gotta let the magic happen and...

579

society and when it comes to copyright offices eventually I think we'll have comfort that

it's okay to create this way and that it's in no way

580

diminishes the work, I guess, it doesn't make it less of an authentic process.

581

Because there are so many, especially as these tools evolve, are so many ways to give even

more of yourself when these tools are there to help you.

582

And yes, sometimes somebody could cheat and use it in a silly way where they don't do

anything, but I think that will show in the quality of the work.

583

But of course, now we're kind of getting bogged down into the details that maybe don't

make as much sense in regards to two of those guidelines.

584

But what makes sense is the part of entirely AI generated music.

585

That's kind of the point, I think, basically.

586

Especially when it comes to these large companies that obviously have trained their models

on

587

copyrighted data.

588

It remains to be seen what happens with trials and stuff, but what do you think about the

ethical considerations when it comes to training models on copyrighted music?

589

Yeah, that's the big question, right?

590

It's so, so, so, tricky.

591

It's so tricky.

592

think the most important thing, the most important thing is to not be hurting artists.

593

Even if you don't train on their data.

594

For example,

595

whole thing was OpenAI and it's Carla Johansson where they found somebody who sounded like

her and they were going to use Carla Johansson look-alike voice without using any of her

596

data.

597

That's also not OK.

598

In the end of the day, the use case of imitating a person without their involvement,

without their permission, without revenue sharing with them is unethical, regardless of

599

how you slice this, regardless of the details of the technology.

600

So I think that's the most important thing.

601

I think we need to be able to tell AI companies the stuff that we're not OK with them

doing.

602

Part of the reason why there is so much focus on how they do it is because there's more

precedent on trying to prevent companies or telling companies how to create their sauce.

603

And also, I mean, there's other issues.

604

You can force, you can say that you can't train on unlicensed data, right?

605

Let's say you have this kind of law.

606

But then, what if a company licenses a million books for a million dollars?

607

Right?

608

A dollar a book.

609

Did that help any of the writers?

610

Is any writer going to benefit from 50 cents or a dollar?

611

Probably not.

612

Who is benefiting?

613

Organizations that have copyright to large catalogs.

614

Is that who we're trying to protect?

615

Maybe.

616

I don't think that's how most people think about it.

617

If you want artists to benefit, need to be...

618

the companies need to pay significant amounts per piece of work.

619

And the economics may or may not work for that.

620

So it's, I feel like we kind of, and I mean that respectfully.

621

I mean, I have so much respect for creators and what they make.

622

But sometimes a lot of the public discourse about it is really, really naive.

623

Just because a company pays for the data doesn't mean that any meaningful money goes to

the creators.

624

And so when we fight for stuff, we need to understand what we're fighting for.

625

Or are we secretly just trying to get rid of these companies altogether?

626

Like, what is our real motive?

627

Or are we trying to prevent them from imitating artists?

628

Or are we trying to make sure that because sometimes I feel like what people what what the

discourse is trying to go for is just just misses a lot of important information.

629

Anyways.

630

I don't want confuse the audience.

631

I'm definitely on the side of the artists in a sense that I think that it should always be

about human creators and anything that the companies do that abuses creators, that hurts

632

them, is not acceptable.

633

I don't know if this was helpful.

634

Yeah.

635

It just tricky a lot of the discourse is naive and it's sort of Benefiting another company

you think it's benefiting you but really just benefiting the rights holders, right and

636

it's sort of

637

Have you heard any good arguments on the other side?

638

being of which other side here.

639

would be that maybe in a philosophical space in terms of collective consciousness and in

terms of ideas being the main thing.

640

I can make that argument.

641

I just don't know if I if I how much I support it but the other argument is Okay, here's

kind of another way I can argue it in the way that I've argued it in the past which which

642

again big question marks It's a problem is ownership right it's the way our world the way

you eat in this world is by having money, right?

643

and so if

644

Sharing without making money.

645

It has no no space in the way that we've organized ourselves, right?

646

And so, you know this opportunity that this these models that are being created which

really belong to all of us to begin with Don't actually belong to all of us.

647

These big companies are profiting from them.

648

And so sometimes in this kind of like financially organized world where

649

Open AI is benefiting from this knowledge created by everybody.

650

We end up hyper-focusing on you must share the revenue with all of us because we've all

created it, right?

651

Which is really understandable.

652

It's an understandable reflex.

653

Of course, it ignores the fact that it's impossible to share with 10 % of the internet,

right?

654

It's just not realistic.

655

But yeah, is a lot of magical things happen when we are willing to aggregate our data and

build a brain from them.

656

And there's a lot of magic that happens, which sometimes we become blinded to because we

are so worried about not getting our fair share.

657

But all of it, every aspect of it is understandable.

658

The way the reason we think this way is because our society forces us to think this way.

659

And we don't know how else to be humans without trying to protect what's ours.

660

And we don't know what to do when the only way to create something amazing makes it

impossible to meaningfully share with everybody who contributed to the data set.

661

Because the data set has to be too large.

662

We don't have a precedent for it.

663

We don't know what to do.

664

It upsets us.

665

It blinds us a little bit to the magic of what is being created because we're so upset

that it's not being shared fairly.

666

And so really there is no model for this.

667

We as humanity don't know how to solve this.

668

Really cool stuff is being created and not everybody who contributed benefits.

669

And that's new territory for us.

670

Yeah, I've heard some advocate for a solution where you have a model that has some kind of

backtracking function so that when you generate something you could kind of trace where it

671

came from and then you could just share revenue based on the material it was trained on.

672

Okay, but it's not true though, right?

673

Like, I've heard this too.

674

There is like, an entire ecosystem, okay?

675

Companies getting funding that have all kind of solutions to this problem, because it's a

real problem, right?

676

But it's not realistic.

677

It's entirely unrealistic.

678

And it's not just a solution, and it's a cool idea, right?

679

But it's a whole universe of solutions like it, and there's like whole set of them.

680

And none of them make any...

681

Respectfully, I feel that there are these serious shortcomings, because if you think about

the human brain, and in many ways these machine brains are similar,

682

We have no idea what we're inspired by.

683

I'm sorry, but we don't.

684

We don't.

685

We might think we do.

686

It's like, let's say if you love Taylor Swift and you're a musician, you might be like,

I'm probably influenced by her, but probably you are influenced by whatever you heard as a

687

little child.

688

And good luck tracing it.

689

Good luck tracing it.

690

Like, it's just not how brains work.

691

Brains don't know how they know stuff.

692

Brains don't know what they're inspired by.

693

And whatever system claims that it does know,

694

probably does so with incredible inaccuracy.

695

It might as well be making up what its source material are.

696

And the ridiculous thing is that people might be happy with it.

697

They might say, they might just believe the system on how it attributes this.

698

They might be happy that there is an attribution, even though most likely the attribution

is false.

699

Right?

700

So it's really more of an attempt to appease this human need for fair attribution, for

fair sharing of the pie.

701

then it is a genuine fair solution.

702

is it, is this technology?

703

We're not ready for it.

704

Our economic and social systems are just not ready for it.

705

And it's, and it's, that's why I think, I mean, as a kind of temporary solution, if you

will, let's just make sure that what it does in the end of the day doesn't hurt anybody

706

directly, at least, that it doesn't imitate any specific artist, that you can't log in and

say,

707

give me music in the style of Maya Akerman, you know, and cut me out of the loop, not

share anything with me, not ask for my permission, but have people be able to imitate me.

708

That would make me upset.

709

That would make any artist legitimately upset.

710

Disallow that.

711

Make that illegal.

712

Make it so that these systems help human beings and don't displace jobs.

713

Let's fix the real problems.

714

Let's think about it for a second.

715

Let's allow ourselves to think about it as a black box.

716

For a second, let's forget about how it works and make sure that what it does is positive.

717

Because when we try to tell the companies how to build their AI, we're really, most of us

are operating in a respect to any field that we know nothing about.

718

And so it's very, very easy to just say stuff that doesn't make sense and demand things

that are not gonna help, demand things that are impractical and ultimately get nowhere,

719

right?

720

Instead, let's demand that these

721

companies do things that are don't do things that are egregiously bad thinking about it as

a black box.

722

I think that's a good point of view to integrate.

723

Even if people are going to continue thinking about data and trying to desperately find

solution, solutions to that.

724

I understand that reflex, that desire.

725

But we also need to be thinking about what is this AI actually doing in our world?

726

Because that's something that's obvious to us.

727

That's something that we can demand control over.

728

Different point of view.

729

We don't tell people how to learn.

730

We don't tell human beings how to learn.

731

But we have a ton of rules of what you are and you're not allowed to do.

732

So I think that perspective, the black box perspective, I'm not allowed to go out there

and offer services.

733

I'm going to write songs for you in the style of, I don't know, the Beatles or whatever.

734

Like, I don't know.

735

I think that would be illegal.

736

I'm pretty sure I'm not allowed to impersonate people.

737

There's a lot of rules on what you are and what you're not allowed to do for human beings,

even though we don't control her.

738

people's brains are formed and how they learn.

739

And we kind of need to take that same attitude, at least in part, to AI, because I think

we're going to get further even as we try to figure out the other pieces.

740

Alright, Maya Ackerman everybody.

741

Yeah, nice.

742

But I mean, I guess what you're saying is we as a whole or humanity or something needs to

tell the companies what not to do and enforce kind of some rules about what we want AI to

743

do with our music and do to our world.

744

Do you think that is realistic?

745

I think any of it would be difficult because some of the AI companies are just really,

really rich.

746

So in that sense, it's very difficult.

747

But telling them how to build the AI is even more difficult.

748

It's even more difficult because...

749

I just see so many mistakes in the public discourse on what people assume is possible.

750

People assume it's realistic, that's just not.

751

And I'm just not seeing that being productive so far.

752

It doesn't mean it's not worth the effort, but I'm not seeing it working.

753

But going around and telling OpenAI, hey, don't use Scarlett Johansson's voice.

754

Even if you don't train on her data, don't imitate her voice.

755

That was successful.

756

They stopped doing that.

757

And they finally stopped and it wasn't clear it was going to work, but they did stop

making it so easy.

758

Like it used to be with Dali.

759

You could say, give me art in the style of, you know, insert your favorite artist or

photographer.

760

That's how they advertise their services initially.

761

But they stopped doing that.

762

Somehow that campaign was successful.

763

So we're actually seeing some success telling companies, hey, we don't want your AI doing

X, Y, Whereas a campaign saying we don't want your AI built in way ABC.

764

There have been some success, but that has been a really slow, really really slow effort.

765

So yeah, I don't care how someone's brains work.

766

Don't break into my house, you know, don't do illegal things.

767

Don't care, right?

768

Like it's a listen to my music, but don't claim to be me, right?

769

Like it's we have a model when it comes to people.

770

Yeah.

771

And those were some good examples of public opinion shaping how these companies conduct

their business.

772

Could you imagine some financial models that takes creators into consideration and also is

appealing to the public in terms of using?

773

Yeah, I mean the part that I think leaves actually a really big opportunity is for the

artists who are interested Only for those who are interested.

774

It's a very very important piece of the puzzle If somebody wants to offer up their style

if somebody wants to offer up their voice, which is such a personal delicate decision

775

Right not to be taken lightly Then I think there is a ways to compensate those artists

really well Right you say you go in and you say, you know, I want a song

776

I'm just going to use my own name to not implicate anyone else, right?

777

I want to use Maya's voice, right?

778

Then I should be compensated for that if I choose to engage in that.

779

And I think for the bigger artists, there is an opportunity to make a lot of money.

780

And even for smaller artists who have a really cool style or really cool voice or

whatever, something that they think other people will want to use in their songs.

781

there's an opportunity to make a lot of money by collaborating with AI companies for the

people who want, who really want that kind of thing.

782

Or maybe like, maybe you're okay with people using your lyrical style and you're not okay

with them using your voice, right?

783

That way you get to license exactly what you feel comfortable licensing.

784

And then you can imagine if that becomes hot, suddenly millions of people are.

785

Writing something in a way that leverages your style and maybe that's exciting for you.

786

And maybe it sounds absolutely I'm sure it sounds absolutely disturbing to a whole bunch

of other people and that's okay Grimes already started leveraging that People are making

787

music using using her voice and last I looked She gets half the royalties and some songs

have become successful.

788

So this is already a precedent for people who want to get involved in that way or maybe an

artist could

789

develop a different sound, a different style, different from the one that they use, and

they would want to license this new thing that they created.

790

So there's really so much possibility here for how an artist could scale how many people

they could kind of collaborate with through these AI systems.

791

So there are exciting possibilities.

792

There are ways to make money.

793

But for most people, it's...

794

I hate saying it, but for most people, it's not gonna be by just licensing their music.

795

Just because...

796

because of how the economics work.

797

I hate to be the bearer of bad news, but if you can imagine, like if you need to license

10 million songs, 20 million songs to build a really powerful model, you can't pay

798

hundreds of dollars per song.

799

It's because of how much money are you really gonna make on that model and how quickly

these numbers blow up very, very quickly and companies need to operate within the

800

constraints of what they're able to raise.

801

It's really terrible.

802

It's a little bit upsetting for me to have to kind of verbalize all of this, but I think

it's important for us as a society to understand the real bottlenecks here so that we can

803

move in a direction that hopefully, at least ultimately, we're happy with.

804

And maybe what you want to say then is I hate these AI models.

805

I don't want them at all.

806

Right.

807

And that's a legitimate perspective as well.

808

whatever perspective you have needs to be sort of it helps if it's informed by

practicality.

809

I think we'll call that the final statement.

810

So you so much for joining the podcast, Maja.

811

This was so fun.

812

Thank you for the fantastic questions.

813

Really enjoyed it.

814

Alright so I'll hit the stop button then.