When news broke last year that AI heavyweight OpenAI and Axel Springer had reached a financial agreement and partnership, it seemed to bode well for harmony between folks who write words, and tech companies that use them to help create and train artificial intelligence models. At the time OpenAI had also come to an agreement with the AP, for reference.
Then as the year ended the New York Times sued OpenAI and its backer Microsoft, alleging that the AI company’s generative AI models were “built by copying and using millions of The Times’s copyrighted news articles, in-depth investigations, opinion pieces, reviews, how-to guides, and more.” Due to what the Times considers to be “unlawful use of [its] work to create artificial intelligence products,” OpenAI’s “can generate output that recites Times content verbatim, closely summarizes it, and mimics its expressive style, as demonstrated by scores of examples.”
The Exchange explores startups, markets and money.
Read it every morning on TechCrunch+ or get The Exchange newsletter every Saturday.
The Times added in its suit that it “objected after it discovered that Defendants were using Times content without permission to develop their models and tools,” and that “negotiations have not led to a resolution” with OpenAI.
Generative AI: Transforming education into a personalized, addictive learning experience
How to balance the need to respect copyright and ensure that AI development doesn’t grind to a halt will not be answered quickly. But the agreements and more fractious disputes between creators and the AI companies that want to ingest and use that work to build artificial intelligence models create an unhappy moment for both sides of the conflict. Tech companies are busy baking new generative AI models trained on data that includes copyright-protected material into their software products; Microsoft is a leader in that particular work, it’s worth noting. And media companies that have spent massively over time to build up a corpus of reported and otherwise created materials are incensed that their efforts are being subsumed into machines that give nothing back to the folks who provided their training data.
You can easily see the argument from either perspective. Tech companies crawl the internet already and have a history of collecting and parsing information for the sake of helping individuals navigate that data. Search engines, in other words. So why is AI training data any different? Media folks on the other hand have seen their own industry decline in recent years — most especially in the realm of journalism, where the Times is a heavyweight — and are loath to see another generation of tech products that depend on their work collect huge revenues while the folks who did the original work receive comparatively little, or in the case of AI training, often nothing.
We don’t need to pick a side here, though I am sure that both you and I have our own views that we could debate. Instead, this morning let’s take a look at some of the critical arguments in play in the AI data-training debate that are shaping how folks consider the issue. It’s going to be a critical issue in 2024. This will be educational for us both, and I think fun as well. To work!
The Times’ argument
The lawsuit is here, and is worth reading in its entirety. Clearly given its length, complete summary is impossible. But I want to highlight a few key points that matter.
The Times states that creating high-quality journalism is very expensive. That’s true. The Times also argues that copyright is critical for the protection of its work, and the functioning of its business model. Again, true.
Continuing, the Times notes that it has a history of licensing its materials to others. You can use its journalism, in other words, but you have to pay for that right from its perspective. The publication separates those arrangements from how its agreements with search engines function, writing: “While The Times, like virtually all online publishers, permits search engines to access its content for the limited purpose of surfacing it in traditional search results, The Times has never given permission to any entity, including Defendants, to use its content for GenAI purposes.”
Clear enough so far, right? Sure, but if LLMs are trained on oceans of data then why does it matter where any particular scrap came from? Can the Times point out clearly that its material was used in such a manner that it is being leaned on heavily to build a commercial product that others are selling without paying it for its inputs to that work?
The paper certainly thinks so. In its suit the Times notes that the “training dataset for GPT-2 includes an internal corpus OpenAI built called ‘WebText,’ which
includes ‘the text contents of 45 million links posted by users of the ‘Reddit’ social network.’” The Times is one of the leading sources used in that particular dataset. Why does that matter? Because OpenAI wrote that WebText was built to emphasize quality of material, per the suit. Put another way, OpenAI said that use of Times material in WebText and GPT-2 was to help make it better.
The Times then turns to WebText2, used in GPT-3, which was “weighted 22% in the training mix for GPT-3 despite constituting less than 4% of the total tokens in the training mix.” And in WebText2, “Times content—a total of 209,707 unique URLs—accounts for 1.23% of all sources listed in OpenWebText2, an open-source re-creation of the WebText2 dataset used in training GPT-3.”
Again, the Times is highlighting that even OpenAI agrees that its work was important to the creation of some of its popular models.
And Times material is well-represented in the CommonCrawl dataset, what the paper describes as the “most highly weighted dataset in GPT-3”. How much Times material is included in CommonCrawl? “The domain www.nytimes.com is the most highly represented proprietary source (and the third overall behind only Wikipedia and a database of U.S. patent documents) represented in a filtered English-language subset of a 2019 snapshot of Common Crawl, accounting for 100 million tokens,” it wrote.
The Times goes on to argue that similar uses of its material was likely in later GPT models built by OpenAI. Usage of Times material, and giving that used material extra weight thanks to its quality without paying for it, is what OpenAI will have to defend under fair use rules.
The Times argument I think boils down to “hey, you took our stuff to make your thing better, and now you are making tons of money off of it and that means you should pay us for what you took, used, and are still using today.” (This riff doesn’t include the Times argument that certain products that make use of AI models that were trained on its data are also cannibalizing its revenue streams by competing with its own, original work; as that argument is downstream from the model creation point, I consider it subsidiary to the above.)
The tech perspective
There was a discussion held by the U.S. Copyright Office last April that included representatives from the venture capital and technology industries, as well as rights holders. You can read a transcript here, which I heartily recommend.
Well-known venture firm a16z took part, arguing that “the overwhelming majority of the time, the output of a generative AI service is not ‘substantially similar’ in the copyright sense to any particular copyrighted work that was used to train the model.”
In the same block of remarks, a16z added that “the data needed [for AI model creation] is so massive that even collective licensing really can’t work. What we’re talking about in the context of these large language models is training on a corpus that is essentially the entire volume of the written word.” As we saw from the above noted Times arguments, it’s true that LLMs do ingest lots of stuff, but does not give it all equal weight. How that will impact the venture argument remains to be seen.
In an October comment again to the U.S. Copyright Office, the same venture firm argued that when “copies of copyrighted works are created for use in the development of a productive technology with non-infringing outputs, our copyright law has long endorsed and enabled those productive uses through the fair use doctrine,” without which search engines and online book search would not work. “[E]ach of these technologies involves the wholesale copying of one or many copyrighted works. The reason they do not infringe copyright is that this copying is in service of a non-exploitive purpose: to extract information from the works and put that information to use” to extend what it could originally do.
To a16z, AI model training is the same: “For the very same reason, the use of copyrighted works en masse to train an AI model—by allowing it to isolate statistical patterns and non-expressive information from those works—does not infringe copyright either. If the U.S. decides to impose “the cost of actual or potential copyright liability on the creators of AI models” it will “either kill or significantly hamper their development.”
Of course, this is an investor talking its book. But in the realm of tech advancement, sometimes a VC talking their book and arguing in favor of rapid technological innovation are one and the same. Summarizing the tech argument, it goes something like “there’s precedent for ingesting lots of data, included copyright-protected data, into tech products without paying for it and this is just that in a new suit.”
Another way to think about it
There’s an interesting question of scale afoot here. Tech thinker Benedict Evans, a former a16z partner it’s worth noting, dug into the thorny issues above, adding the following bit of cud for us to chew:
[O]ne way to think about this might be that AI makes practical at a massive scale things that were previously only possible on a small scale. This might be the difference between the police carrying wanted pictures in their pockets and the police putting face recognition cameras on every street corner – a difference in scale can be a difference in principle. What outcomes do we want? What do we want the law to be? What can it be? The law can change.
The Times and the tech industry are arguing current law. Evans points out that the scale of data ingestion for AI model creation could create a scenario when existing law might not fit what we want to have happen as a society. And that the law can change — provided that the nation’s elected officials can, in fact, still pass laws.
Summing, the Times argues with the receipts that its data was used more than other data in training certain OpenAI models because it was good. And since that material is copyrighted and used in particular, it should get paid. OpenAI and its backers and defenders are hoping that existing precedent and fair use legal protections are enough to keep their legal and financial liabilities low while they make lots of money with their new technologies. Finally, it’s also possible that we need new laws to handle situations like this, as what we have might not have the right scale in mind to handle what’s going on.
From where I sit, I don’t expect any OpenAI money to come to me for whatever it has ingested of my own writing. But I also don’t own most of it — my employers both current and historical do, and they have a lot more total material, and far greater legal resources to bring to bear along with the very same profit motive that the Times and OpenAI have. Perhaps I too will get dragged into this by proxy. That will make reporting on it all the more touchy. And hey, maybe that reporting itself will help future AI models explain to other people why they don’t have to pay for it.
