OpenAI, Napster, and iTunes

I’m not the first to note the similarities between today’s AI landscape and the Napster era of the music industry. OpenAI, and many of its rivals, have created powerful tools fueled by potentially infringing training data. In fact, OpenAI has said that “it would be impossible to train today’s leading AI models without using copyrighted materials.”

One way today’s AI picture differs from the Napster era is Apple’s role as an AI laggard. The prevailing wisdom is that Apple is far behind the OpenAIs of the world. But that perspective may be equal parts a misremembering of history and a misread of Apple’s position in AI.

In fact, I think Apple may be primed for another iTunes Music Store moment.

In my recollection, Napster was at its nadir when Apple, via the iTunes Music Store, created a legal alternative to the music sharing service. The combination of copyright lawsuits from the music industry and Apple’s easy-to-use, legal alternative caused the end of Napster.

But my memory was faulty. Napster was released on June 1, 1999. A few months later, Metallica sued Napster. By July 2001, a court injunction had forced Napster to shut down its service.

iTunes was released six months before Napster’s injunction, but the iTunes Music Store, the service that made digital music legal, was released in April 2003, long after Napster had collapsed. Apple was just getting into the music business as Napster ended. To put it in perspective, Napster was on life support before the first iPod was announced.

The lesson here is one summarized well in Baldur Bjarnason’s excellent book on AI: “With rare exceptions, there is no first-mover advantage in software.”

Like Napster, OpenAI faces significant legal scrutiny and is banking on fair use defenses. This situation raises concerns over the sustainability of current AI models because of their reliance on potentially infringing data.

Perhaps the biggest challenge to OpenAI comes from the New York Times (NYTimes) lawsuit alleging copyright infringement. In the lawsuit, the NYTimes alleges not only that OpenAI trained ChatGPT on NYTimes content without permission, but that ChatGPT will reproduce sections of articles nearly word-for-word.

ChatGPT isn’t the only AI model susceptible to this problem. And it isn’t limited to text. Gary Marcus and Reid Southen documented how easy it is to generate copyrighted images without even invoking a specific copyrighted property. All they had to do was ask the AI for a popular movie screencap.

I don’t know what the outcome of these legal challenges will be, but it isn’t out of the realm of possibility that they could mean the end of OpeAI like they did to Napster before it.

Which is why I found it notable that less than a week before the NYTimes sued OpenAI, they broke the news that Apple has approached multiple news publishers, including the NYTimes itself, to license their content for use in Apple’s upcoming AI efforts.

If OpenAI loses, or even if the risks of using OpenAI and its competitors seem too high for corporations, there is going to be a market for legal alternatives. Apple has experience convincing reluctant industry players to license their content to provide that alternative.

Apple has additional advantages beyond its wherewithal to license content. First, it has years of experience with on-device machine learning. It has shipped millions of devices with its Neural Engine—its custom neural processing unit used to power artificial intelligence.

Having dedicated hardware for neural networks enables Apple to do things that OpenAI cannot. Using on-device chips instead of data centers could greatly reduce the cost and environmental impact of AI.

In addition, Apple has placed an emphasis on user privacy and considers it a differentiator from competitors like Google and Facebook. By contrast, it is unclear how leading AI companies will comply with privacy laws like GPDR and it’s right to be forgotten. Processing AI locally instead of sending customer data and queries to the cloud greatly reduces these privacy concerns.

While Apple is rightly perceived as behind its competitors when it comes to artificial intelligence—you have to look no further than Siri’s lack of comprehension to many prompts to know this is true—there is plenty of evidence that Apple has ramped up its AI efforts.

In the last few weeks, Apple released a research paper describing a technique to use flash storage to run large language models on a phone. It has also released two open source AI libraries including a multimodal generative AI. All signs point to big AI announcements from Apple at WWDC later this year.

From what we know of Apple’s AI efforts thus far and its mixed history of cloud-based services, one would be tempted to bet against them. They don’t have the best track record.

But when I look at Apple’s efforts to license content, their large install base of neural engines, and their history of providing a safe alternative to Silicon Valley’s “move fast and break things” culture, I think people may be surprised at what comes next.

Plus, when it comes to many of the leading AI products, I’m reminded of what Steve Jobs said of Napster and its successor Kazaa, “It’s stealing. And it’s best not to mess with karma.”