The ethics of using books to train AI

Published books vs online text

Unlike online material, such as social media posts and blogs, books offer a higher quality of data. Books have been edited to a more stringent level than most online material, and usually someone has paid the author to publish them, so the grammar and punctuation are of a higher standard. Furthermore, books can articulate long-form arguments well, unlike online material, which is often shorter and more condensed.

“We’ve talked for so long about ‘garbage in garbage out (GIGO)’, and it remains true,” says Smith. “If AI models were trained on higher quality data, then you wouldn’t need so much of it. I’m not sure that AI models know what good looks like.”

By using books that were printed before 2022 (the year ChatGPT was launched), developers can be certain that generative AI was not used in the writing of the books. AI models lose accuracy after being recursively trained on data produced by other AI systems, as it can lead to a degenerative learning process and ultimately model collapse.

Copyright and licensing challenges

Authors and publishers have experienced multifarious challenges following the release of generative AI systems. There have been questions about the legality of AI tools being trained on copyrighted works and concerns that some AI systems can generate text in the style of specific authors.

It is incredibly easy for AI companies to scan books … at scale. Doing it at scale is the reason the books are destroyed, because they need to cut the spine off so that they can feed it through scanning machines at incredible speed

In the US, there is the fair use doctrine (part of Section 107 of the 1976 Copyright Act), which permits limited use of copyrighted material, such as for converting physical media into digital files, without permission from the copyright holder. In the UK, Section 29 of the Copyright, Designs and Patents Act 1988 permits the conversion of physical media into digital, but only if it is for non-commercial purposes.

“UK copyright law provides specific ‘fair dealing’ exceptions, which are narrower and not interchangeable with US fair use,” says Stevenson. “Whether a particular use qualifies as a ‘fair dealing’ depends on the facts, including whether the amount taken was reasonable and appropriate. A use is unlikely to be fair if it substitutes for the original work and causes the copyright owner to lose revenue.”

It is worth noting that some books exist outside of copyright controls. For example, the copyright will have expired for some works (in the UK, this is usually 70 years following the author’s death) and some will have been published free of copyright limitations.

This has brought up the question of territorialism in copyright enforcement. If an AI system is trained in a country where it is permitted to use copyrighted material, what happens when that AI is deployed in regions that have much stricter copyright controls?

Territorial control in copyright was one of the core elements in the Getty Images vs Stability AI legal case.

“We all thought that this case was going to provide clarification in this area,” says Mark Deem, digital lawyer and founder of Fortiter. “Unfortunately, it didn’t, because on this specific point, they could not easily establish that the copying took place within the territorial limit of the UK, and the question was not finally resolved.”

The aforementioned regional challenges of AI and copyright licence agreements are similar to those experienced by the film industry. Since 1997, the film industry has used digital rights management to control the regionalisation approach to DVD and Blu-ray distribution around the world.

Does AI learning require content licensing regulation?

Given the copyright challenges of AI training, it raises the question of whether there needs to be a global consensus regarding the use of copyrighted material for the purposes of training AI. Just as the General Data Protection Regulation has become the de facto standard for data protection regulation, does the AI sector require a similar approach to regulation?

There is an emerging licensing market, which enables AI companies and rights holders to enter into licensing deals and agree on terms and remuneration about how the work is to be used

Some publishers have already agreed licensing agreements, whereby AI companies that agree to certain conditions can train AI on that publisher’s books, provided they and their authors are appropriately compensated for the use of their material.

“There is an emerging licensing market, which enables AI companies and rights holders to enter into licensing deals and agree on terms and remuneration about how the work is to be used,” says Ambre Morvan, senior public policy manager and contracts adviser for the Society of Authors.

In its drive to compete with other countries, the UK government published a paper outlining four proposals for copyright reform.

Two of the proposals suggested weakening copyright laws by providing exceptions for when works are used for the purpose of training AI systems. However, following a campaign by the creative industries, the government has since indicated it will no longer proceed with any proposals that would weaken the existing copyright framework.

“The law can never prescribe exactly where technology is developing and how it’s developing, but it can provide the appropriate guardrails for the way in which it is developed, and we need to be very careful that what we don’t allow the laxing of those guardrails to be dressed up as an opportunity to try to get better development of AI,” says Deem.

“It’s not the speed that is the issue at the moment, it’s making sure that the guardrails are there to make sure it develops in the right way, respectful of where society is at the moment, the rights people have within society at the moment and the potential damage it could do to the creative industries.”

The future of AI training

As an alternative to large-scale AI, small language models (SLMs) are being deployed for specific tasks, which therefore require far less data to be effective. However, small language models lack the multi-functionality of large language models (LLMs).

“Some organisations actually prefer, rather than large language models, small language models; models that are highly tuned for specific purposes,” says Smith. “They are trained on specific data for particular use cases, so it’s not a case of large language models or nothing.”

Nonetheless, there remains a need for an international regulatory framework for AI systems to ensure that they are trained and deployed appropriately and ethically. However, given current geopolitical tensions, any form of internationally agreed treaty seems unlikely.

Anthropic was approached for comment, but did not respond.