Large Language Models (LLMs) like ChatGPT, Bard and even open source versions are trained on public Internet content. But there are also indications that popular AIs might also be trained on datasets created from pirated books.

Is Dolly 2.0 Trained on Pirated Content?

Dolly 2.0 is an open source AI that was recently released. The intent behind Dolly is to democratize AI by  making it available to everyone who wants to create something with it, even commercial products.

But there’s also a privacy issue with concentrating AI technology in the hands of three major corporations and trusting them with private data.

Given a choice, many businesses would prefer to not hand off private data to third parties like Google, OpenAI and Meta.

Even Mozilla, the open source browser and app company, is investing in growing the open source AI ecosystem.

The intent behind open source AI is unquestionably good.

But there is  an issue with the data that is used to train these large language models because some of it consists of pirated content.

Open source ChatGPT clone, Dolly 2.0, was created by a company called DataBricks  (learn more about Dolly 2.0)

Dolly 2.0 is based on an Open Source Large Language Model (LLM) called Pythia (which was created by an open source group called, EleutherAI).

EleutherAI created eight versions of LLMs of different sizes within the Pythia family of LLMs.

One version of Pythia, a 12 billion parameter version, is the one used by DataBricks to create Dolly 2.0, as well as with a dataset that DataBricks created themselves (a dataset of questions and answers that was used to train the Dolly 2.0 AI to take instructions)

The thing about the EleutherAI Pythia LLM is that it was trained using a dataset called the Pile.

The Pile dataset is comprised of multiple sets of English language texts, one of which is a dataset called Books3. The Books3 dataset contains the text of books that were pirated and hosted at a pirate site called, bibliotik.

This is what the DataBricks announcement says:

“Dolly 2.0 is a 12B parameter language model based on the EleutherAI pythia model family and fine-tuned exclusively on a new, high-quality human generated instruction following dataset, crowdsourced among Databricks employees.”

Pythia LLM Was Created With the Pile Dataset

The Pythia research paper by EleutherAI that mentions that Pythia was trained using the Pile dataset.

This is a quote from the Pythia research paper:

“We train 8 model sizes each on both the Pile …and the Pile after deduplication, providing 2 copies of the suite which can be compared.”

Deduplication means that they removed redundant data, it’s a process for creating a cleaner dataset.

So what’s in Pile? There’s a Pile research paper that explains what’s in that dataset.

Here’s a quote from the research paper for Pile where it says that they use the Books3 dataset:

“In addition we incorporate several existing highquality datasets: Books3 (Presser, 2020)…”

The Pile dataset research paper links to a tweet by Shawn Presser, that says what is in the Books3 dataset:

“Suppose you wanted to train a world-class GPT model, just like OpenAI. How? You have no data.

Now you do. Now everyone does.

Presenting “books3”, aka “all of bibliotik”

– 196,640 books
– in plain .txt
– reliable, direct download, for years: https://the-eye.eu/public/AI/pile_preliminary_components/books3.tar.gz”

So… the above quote clearly states that the Pile dataset was used to train the Pythia LLM which in turn served as the foundation for the Dolly 2.0 open source AI.

Is Google Bard Trained on Pirated Content?

The Washington Post recently published a review of Google’s Colossal Clean Crawled Corpus dataset (also known as C4 – PDF research paper here) in which they discovered that Google’s dataset also contains pirated content.

The C4 dataset is important because it’s one of the datasets used to train Google’s LaMDA LLM, a version of which is what Bard is based on.

The actual dataset is called Infiniset and the C4 dataset makes up about 12.5% of the total text used to train LaMDA. Citations to those facts about Bard can be found here.

The Washington Post news article published:

“The three biggest sites were patents.google.com No. 1, which contains text from patents issued around the world; wikipedia.org No. 2, the free online encyclopedia; and scribd.com No. 3, a subscription-only digital library.

Also high on the list: b-ok.org No. 190, a notorious market for pirated e-books that has since been seized by the U.S. Justice Department.

At least 27 other sites identified by the U.S. government as markets for piracy and counterfeits were present in the data set.”

The flaw in the Washington Post analysis is that they’re looking at a version of the C4 but not necessarily the one that LaMDA was trained on.

The research paper for the C4 dataset was published in July 2020. Within a year of publication another research paper was published that discovered that the C4 dataset was biased against people of color and the LGBT community.

The research paper is titled, Documenting Large Webtext Corpora: A Case Study on the Colossal Clean Crawled Corpus (PDF research paper here).

It was discovered by the researchers that the dataset contained negative sentiment against people of Arab identies and excluded documents that were associated with Blacks, Hispanics, and documents that mention sexual orientation.

The researchers wrote:

“Our examination of the excluded data suggests that documents associated with Black and Hispanic authors and documents mentioning sexual orientations are significantly more likely to be excluded by C4.EN’s blocklist filtering, and that many excluded documents contained non-offensive or non-sexual content (e.g., legislative discussions of same-sex marriage, scientific and medical content).

This exclusion is a form of allocational harms …and exacerbates existing (language-based) racial inequality as well as stigmatization of LGBTQ+ identities…

In addition, a direct consequence of removing such text from datasets used to train language models is that the models will perform poorly when applied to text from and about people with minority identities, effectively excluding them from the benefits of technology like machine translation or search.”

It was concluded that the filtering of “bad words” and other attempts to “clean” the dataset was too simplistic and warranted are more nuanced approach.

Those conclusions are important because they show that it was well known that the C4 dataset was flawed.

LaMDA was developed in 2022 (two years after the C4 dataset) and the associated LaMDA research paper says that it was trained with C4.

But that’s just a research paper. What happens in real-life on a production model can be vastly different from what’s in the research paper.

When discussing a research paper it’s important to remember that Google consistently says that what’s in a patent or research paper isn’t necessarily what’s in use in Google’s algorithm.

Google is highly likely to be aware of those conclusions and it’s not unreasonable to assume that Google developed a new version of C4 for the production model, not just to address inequities in the dataset but to bring it up to date.

Google doesn’t say what’s in their algorithm, it’s a black box. So we can’t say with certainty that the technology underlying Google Bard was trained on pirated content.

To make it even clearer, Bard was released in 2023, using a lightweight version of LaMDA. Google has not defined what a lightweight version of LaMDA is.

So there’s no way to know what content was contained within the datasets used to train the lightweight version of LaMDA that powers Bard.

One can only speculate as to what content was used to train Bard.

Does GPT-4 Use Pirated Content?

OpenAI is extremely private about the datasets used to train GPT-4. The last time OpenAI mentioned datasets is in the PDF research paper for GPT-3 published in 2020 and even there it’s somewhat vague and imprecise about what’s in the datasets.

The TowardsDataScience website in 2021 published an interesting review of the available information in which they conclude that indeed some pirated content was used to train early versions of GPT.

They write:

“…we find evidence that BookCorpus directly violated copyright restrictions for hundreds of books that should not have been redistributed through a free dataset.

For example, over 200 books in BookCorpus explicitly state that they “may not be reproduced, copied and distributed for commercial or non-commercial purposes.””

It’s difficult to conclude whether GPT-4 used any pirated content.

Is There A Problem With Using Pirated Content?

One would think that it may be unethical to use pirated content to train a large language model and profit from the use of that content.

But the laws may actually allow this kind of use.

I asked Kenton J. Hutcherson, Internet Attorney at Hutcherson Law what he thought about the use of pirated content in the context of training large language models.

Specifically, I asked if someone uses Dolly 2.0, which may be partially created with pirated books, would commercial entities who create applications with Dolly 2.0 be exposed to copyright infringement claims?

Kenton answered:

“A claim for copyright infringement from the copyright holders of the pirated books would likely fail because of fair use.

Fair use protects transformative uses of copyrighted works.

Here, the pirated books are not being used as books for people to read, but as inputs to an artificial intelligence training dataset.

A similar example came into play with the use of thumbnails on search results pages. The thumbnails are not there to replace the webpages they preview. They serve a completely different function—they preview the page.

That is transformative use.”

Karen J. Bernstein of Bernstein IP offered a similar opinion.

“Is the use of the pirated content a fair use? Fair use is a commonly used defense in these instances.

The concept of the fair use defense only exists under US copyright law.

Fair use is analyzed under a multi-factor analysis that the Supreme Court set forth in a 1994 landmark case.

Under this scenario, there will be questions of how much of the pirated content was taken from the books and what was done to the content (was it “transformative”), and whether such content is taking the market away from the copyright creator.”

AI technology is bounding forward at an unprecedented pace, seemingly evolving on a week to week basis. Perhaps in a reflection of the competition and the financial windfall to be gained from success, Google and OpenAI are becoming increasingly private about how their AI models are trained.

Should they be more open about such information? Can they be trusted that their datasets are fair and non-biased?

The use of pirated content to create these AI models may be legally protected as fair use, but just because one can does that mean one should?

Featured image by Shutterstock/Roman Samborskyi