TRASCRIZIONE 🔗
WEBVTT
1
00:00:00.100 --> 00:00:04.680
A client asked me last month: "If the AI designed our logo, do we own it?"
2
00:00:05.180 --> 00:00:06.060
And what did you say?
3
00:00:06.740 --> 00:00:10.480
I said: "Who designed it?" And the room went very quiet.
4
00:00:10.960 --> 00:00:13.440
That silence is our topic today.
5
00:00:23.020 --> 00:00:56.460
Welcome to ideation.zone. I'm Elisabeth, and this is the third episode of our season on ideas and artificial intelligence. Today we talk about ownership – who owns an idea developed with a machine, and whose ideas went into the machine in the first place. One thing before we start: none of us is a lawyer, and nothing here is legal advice. The court decisions we'll mention are recent, and several are under appeal. Yannis, Noa – hello.
6
00:00:57.020 --> 00:01:06.040
Hello. And I'll add a second disclaimer: I'm a psychologist. I know more about why people feel they own an idea than about whether they legally do.
7
00:01:06.600 --> 00:01:12.900
And I'm the one who has to put it in contracts. So I've read more court decisions this year than in my whole life.
8
00:01:13.460 --> 00:01:18.620
Then you're both more qualified than most people who talk about this. Let's start.
9
00:01:23.200 --> 00:01:27.880
There's a leading case in the United States, with a wonderful name for a painting. Noa?
10
00:01:29.040 --> 00:01:54.360
"A Recent Entrance to Paradise". A computer scientist, Stephen Thaler, built a system he calls the "Creativity Machine". It produced an image, and he applied for copyright – with the machine as the author, and himself as the owner. The Copyright Office said no. And in March twenty twenty-five, the court of appeals in Washington said no too, unanimously. Copyright needs a human author.
11
00:01:54.840 --> 00:01:56.240
It reminds me of the monkey.
12
00:01:56.740 --> 00:01:57.440
The monkey?
13
00:01:58.800 --> 00:02:19.940
A crested macaque in Indonesia picked up a photographer's camera and took a selfie – a brilliant one, grinning straight into the lens. An animal rights group sued in the monkey's name for the copyright. And an American court decided that animals can't own copyright. The monkey had the idea, the photographer had the camera – and nobody had the right.
14
00:02:20.480 --> 00:02:26.180
Machine, monkey – the law keeps saying the same thing. No human, no author.
15
00:02:26.700 --> 00:02:27.800
And patents?
16
00:02:28.400 --> 00:02:46.880
The same. Thaler also tried to name his AI as the inventor on patents – in the US, in Britain, and elsewhere. Rejected everywhere. And last year the American patent office said it clearly: AI is a tool, like lab equipment. The inventor is the human who conceived the idea.
17
00:02:48.140 --> 00:03:01.260
Yannis, in our first episode you talked about an AI co-scientist that proposes research hypotheses. If a machine proposes most of the technical details of an invention – who's the inventor?
18
00:03:01.640 --> 00:03:27.520
That's one of the exposé's open questions, and I find it genuinely hard. Patent law asks who had the "conception" – the definite idea of the complete invention. My guess: patent offices will start asking for the trail. What did the human ask, what did the machine propose, what did the human reject, test, change? The inventor becomes the person who can show the decisions – not the person who had the first spark.
19
00:03:28.060 --> 00:03:31.460
So the scientist becomes the curator of the machine's sparks.
20
00:03:32.240 --> 00:03:42.440
Curator and tester. Which, if you think about it, is what good research directors have always been. They rarely have the idea. They know which idea to bet the lab on.
21
00:03:42.900 --> 00:03:45.480
That's a bit sad for the person who had the spark.
22
00:03:45.860 --> 00:03:49.540
It's a bit sad for the machine too. But it doesn't mind.
23
00:03:50.140 --> 00:03:54.920
So what about prompting? If I write a very detailed prompt, am I the author?
24
00:03:55.440 --> 00:04:12.500
In the US, the Copyright Office said in January twenty twenty-five: with today's technology, prompts alone aren't enough. A prompt is an instruction. The machine decides how the result looks. It's like telling a painter "paint me a sunset over the sea" – you didn't paint it.
25
00:04:12.880 --> 00:04:16.420
But here's where it gets interesting, because other courts see it differently.
26
00:04:16.860 --> 00:04:36.260
China. In twenty twenty-three, the Beijing Internet Court looked at an image someone made with Stable Diffusion. He had used more than a hundred and fifty prompts, adjusted parameters, chosen the result that matched his intention. And the court said: that's enough human choice – he's the author.
27
00:04:36.880 --> 00:04:39.380
Almost the same situation, opposite answer.
28
00:04:39.800 --> 00:04:57.620
Opposite answer. And Britain has had a strange rule since nineteen eighty-eight – long before any of this – for "computer-generated works" without a human author. The author is the person who made the arrangements necessary for the work. Nobody quite knows what that means for a chatbot.
29
00:04:58.340 --> 00:05:27.240
Can I add the psychological side? Because legally, ideas were never owned anyway. Copyright protects the expression, not the idea. But people feel very strongly that ideas belong to someone. And there's an old finding in the sociology of science – Robert Merton called it the Matthew effect, in nineteen sixty-eight. Credit flows to those who already have it. When a famous and an unknown scientist find the same thing, the famous one gets remembered.
30
00:05:27.780 --> 00:05:29.180
And AI makes that worse?
31
00:05:29.820 --> 00:05:41.180
My worry is that it makes credit even more arbitrary. If the machine did half the work, whoever presents the result gets the credit – and usually that's the person who already has the platform.
32
00:05:42.340 --> 00:05:44.520
Noa, what do you actually do in your studio now?
33
00:05:45.500 -->
Load full transcript
00:06:22.396
Three things. First, we keep a decision log – who framed the problem, which AI outputs we picked and why, how we changed them. It sounds bureaucratic. It's actually the most useful document in the project, because it shows our thinking. Second, we transform everything. We never hand a raw AI output to a client as a final design. The more human changes, the stronger the claim – and honestly, the better the work. And third, we put it in the contract from day one: how AI-assisted work is credited and who owns what.
34
00:06:22.936 --> 00:06:33.996
The decision log is interesting from a research point of view too. It's the only evidence of what the human actually contributed. Without it, nobody can tell – not even you, after a month.
35
00:06:34.616 --> 00:06:35.816
After a week, honestly.
36
00:06:36.396 --> 00:06:49.216
And there's one more kind of protection people forget – confidentiality. An idea that you keep secret is protected as a business secret, no matter who or what generated it. As long as you actually keep it secret.
37
00:06:49.716 --> 00:06:52.916
Which is harder than it sounds when you paste it into a chatbot.
38
00:06:54.316 --> 00:07:16.276
Exactly. In twenty twenty-three, engineers at Samsung pasted internal source code and meeting notes into ChatGPT to get help. It became a scandal, and Samsung restricted the tools for a while. Now the first rule in my studio is: no client brief goes into a public AI tool. Only into the ones we have a contract with.
39
00:07:16.836 --> 00:07:22.536
So the danger isn't only that someone takes your idea. It's that you give it away yourself.
40
00:07:23.316 --> 00:07:26.996
Usually it's that. Nobody steals ideas. People leak them.
41
00:07:27.556 --> 00:07:39.116
There's an open question in this exposé I'd like you both to try. Should lawmakers create a new kind of right for AI-assisted works – weaker and shorter than normal copyright?
42
00:07:39.796 --> 00:07:58.376
My guess: something like that is coming, and it might even be reasonable. Copyright lasts seventy years after the author's death. For a logo that took twenty minutes with a machine, that's absurd. A short right – five years, maybe ten – would protect the investment without locking the commons for a century.
43
00:07:58.956 --> 00:08:12.095
I'm not sure. Every new right creates a new industry of people registering things. Imagine someone generating ten million slogans overnight and registering all of them. Then every human writer steps on a mine.
44
00:08:13.776 --> 00:08:22.676
That's a terrifying image. And a good objection. Then the short right would need a threshold – real human selection and change, documented.
45
00:08:23.216 --> 00:08:24.816
Then we're back at my decision log.
46
00:08:25.376 --> 00:08:28.896
Then your decision log becomes a legal document. Congratulations.
47
00:08:29.516 --> 00:08:30.956
My accountant will be thrilled.
48
00:08:35.476 --> 00:08:46.636
Second exposé, and now we turn it around. Not who owns what comes out – but whose work went in. Noa, there was a landmark case in Munich last year.
49
00:08:47.336 --> 00:09:14.796
GEMA against OpenAI. GEMA is the German society that manages music rights. They showed that ChatGPT, with simple prompts, reproduced the lyrics of well-known German songs – almost word for word. And in November twenty twenty-five, the regional court in Munich said: if the model can reproduce the lyrics like that, it has memorised them. And memorising them inside the model is itself a copy.
50
00:09:15.396 --> 00:09:17.256
A copy inside the machine.
51
00:09:17.836 --> 00:09:24.776
Inside the parameters. That was the striking part. Not just the output – the model itself contains the work.
52
00:09:25.336 --> 00:09:28.896
And one week earlier, a British court said almost the opposite.
53
00:09:29.436 --> 00:09:46.276
Getty Images against Stability AI, in London. Getty argued that the image model contained copies of their photos. And the court found, on the evidence, that the model's weights didn't store copies of the images. It learned patterns, it didn't keep pictures.
54
00:09:46.996 --> 00:09:49.676
Two courts, one week, two answers.
55
00:09:50.236 --> 00:10:04.716
But I don't think they contradict each other as much as it sounds. Song lyrics are short and distinctive – a model can learn them by heart. Millions of photos are not stored that way. It depends on what kind of work and what the model can actually reproduce.
56
00:10:05.316 --> 00:10:20.656
In the Getty case, there was a famous detail from the beginning – the generated images sometimes showed a distorted version of the Getty watermark. That's how they noticed. The machine had seen that watermark so often that it thought it belonged to the picture.
57
00:10:22.135 --> 00:10:24.156
That's a lovely accidental confession.
58
00:10:24.716 --> 00:10:25.756
And in America?
59
00:10:26.576 --> 00:10:51.096
Two judges in San Francisco, summer twenty twenty-five. In the case against Anthropic, the judge said training on books you bought legally is fair use – it's transformative, like a reader learning. But building a library from pirated copies is not. The company later agreed to pay a settlement of one and a half billion dollars – roughly three thousand dollars per book.
60
00:10:52.776 --> 00:10:53.616
Three thousand dollars a book.
61
00:10:53.996 --> 00:11:09.856
And in the case against Meta, Meta won – but only because the authors hadn't shown evidence of harm. The judge wrote that in many cases, copying protected works to train AI without permission will be illegal, because the AI can flood the market for the originals.
62
00:11:10.516 --> 00:11:28.516
I have a personal interest here, I should say. I'm a journalist. Everything I've written for thirty years is online somewhere, and it's very likely in some model. The New York Times is suing over exactly that. And I don't know what I feel about it.
63
00:11:29.096 --> 00:11:30.456
What's the part that
bothers you?
64
00:11:30.816 --> 00:11:44.876
Not that a machine learned from me. I learned from thousands of other journalists. It's that nobody asked, and that the machine now writes in a minute what I need a day for – partly because it read me.
65
00:11:45.456 --> 00:12:05.536
That's exactly how illustrators feel. When the "in the style of" trend exploded – everyone turning their photos into the style of a famous Japanese animation studio – a lot of illustrators I know were heartbroken. Style isn't protected by copyright in most places. But they spent twenty years developing theirs.
66
00:12:05.956 --> 00:12:07.936
What happened to your illustrator friends?
67
00:12:08.364 --> 00:12:35.584
Different things. One stopped posting her work online altogether – which is sad, because that's how she found clients. Another went the other way: she sells licences of her style to one model company, with her name on it. She says: "If they're going to learn me, they're going to pay me." A third one moved to embroidery. Real thread, real hours. She says nobody has trained a model on her hands yet.
68
00:12:36.064 --> 00:12:41.764
That's three very different strategies. Withdraw, license, or move to what can't be copied.
69
00:12:42.244 --> 00:12:44.864
And all three feel like losses to them, somehow.
70
00:12:45.604 --> 00:13:10.544
And here's the counter-argument, which I think is serious. All creativity is recombination. Arthur Koestler wrote a whole book about it in nineteen sixty-four – "The Act of Creation". He called it bisociation: new ideas come from connecting two frames that were separate. Newton wrote that he stood on the shoulders of giants. If learning from existing work were theft, no designer, no scientist could work.
71
00:13:11.124 --> 00:13:17.464
True. But Newton didn't stand on the shoulders of every giant at once, in a second, and then sell the view.
72
00:13:18.924 --> 00:13:21.404
That's a very good sentence. I have no reply to that.
73
00:13:21.844 --> 00:13:24.884
Write it down, you said you don't have data for everything.
74
00:13:25.444 --> 00:13:26.644
So where's the line?
75
00:13:27.084 --> 00:13:49.004
The courts seem to be converging on one: learning versus keeping. Learning patterns from lawfully acquired works – broadly accepted, as for people. Keeping or reproducing specific works – song lyrics, a photo, a melody – that's where they see a copy. Add two more conditions: where did you get the data – bought or pirated – and did the creator say no?
76
00:13:49.544 --> 00:14:05.064
That's the opt-out. In Europe, under the AI Act's code of practice, model providers are supposed to respect machine-readable opt-outs – a line on your website saying "don't train on this". Which only helps if you know about it and have a website.
77
00:14:05.584 --> 00:14:09.464
And for someone in a workshop, using AI to find a slogan?
78
00:14:10.024 --> 00:14:37.784
That's where it gets very practical. The most common things people ask AI for in brainstorming are exactly the risky ones – slogans, names, short melodies, logos. Short, distinctive things a model might have memorised. A slogan that sounds perfect might sound perfect because it already exists. So in my studio, every name or slogan from AI gets searched before anyone falls in love with it.
79
00:14:38.244 --> 00:14:44.164
That's the fixation problem again, from our first episode. Once the team loves it, nobody wants to check.
80
00:14:44.844 --> 00:14:55.424
The exposé asks whether collective licensing – like GEMA does for music – could become the standard way to pay people whose work trains AI.
81
00:14:56.084 --> 00:15:20.564
My guess: yes, for some fields, and it's the most realistic solution. Music has done this for a hundred years. A radio station doesn't negotiate with every songwriter – it pays a collecting society, and the society distributes the money. You could do the same for training data: model providers pay into a fund, and creators get paid according to how much of their work was used.
82
00:15:21.124 --> 00:15:31.864
The hard part is "how much was used". With radio you count how often a song was played. With a model, you can't easily say which training text contributed to which answer.
83
00:15:32.484 --> 00:15:40.684
True. So maybe a rougher key – how much of your work is in the dataset, not how often it's "used". Imperfect, but better than nothing.
84
00:15:41.304 --> 00:15:49.824
I'm sceptical it will reach the illustrator who isn't in any society. The well-organised get paid; the scattered don't. That's the Matthew effect again.
85
00:15:51.044 --> 00:15:52.924
Yes. That's the weak spot.
86
00:15:53.544 --> 00:16:24.463
Let me gather this. On ownership, you agree that the law everywhere draws the line at human contribution – and disagrees about how much is enough. A prompt alone, probably not; a hundred and fifty deliberate choices, maybe. And you agree that the most useful thing a team can do is document what the humans actually decided. Noa doubts that a new short right would help; Yannis thinks it could, with a real threshold.
87
00:16:25.104 --> 00:16:26.224
And a decision log.
88
00:16:26.844 --> 00:16:47.464
And a decision log. On training data, you agree that the line runs between learning and keeping – and that slogans, names and melodies from AI need checking before anyone falls in love with them. Where you differ is whether collective licensing will reach the people who need it most.
89
00:16:48.084 --> 00:16:49.784
I hope it does. I doubt it.
90
00:16:50.344 --> 00:16:52.964
I hope it does. I'm going to lobby for it.
91
00:16:53.524 --> 00:17:03.724
So here's the question I'd like to leave with you. Think of an idea you're proud of. Whose work made it possible – and did you ever tell them?
92
00:17:04.504 --> 00:17:27.264
The exposés, with all the cases and sources, are on ideation.zone. Next time, the last episode of this season: what AI does to thinking itself – to critical judgement, to a machine that always agrees with us, and to cultures that see the world differently. Yannis, Noa – thank you.
93
00:17:27.924 --> 00:17:31.464
Thank you, Elisabeth. That was the most legal I've ever been.
94
00:17:32.724 --> 00:17:34.684
And I'll send you both my decision log.