When Machines Read for Profit: Rethinking AI Training as Functional Exploitation 

When an AI system is trained on copyrighted books, images or music, developers often describe the process as machine “reading”. But a system that copies a book to help users find information in it is doing something different from a system that uses the book to develop a reusable computational capability such as the ability to generate, classify or transform new content. That distinction matters for how copyright law evaluates AI training.  

Beyond the Reading Analogy  

What does it mean for a machine to “use” a copyrighted work?  

One way of approaching the question is through the idea of reading. Humans read books, acquire information from them and later build on what they have learned. Copyright does not ordinarily give authors control over every idea or capability that another person acquires by encountering a work. A novelist can read hundreds of novels before writing another. A researcher can study existing scholarship before developing a new theory.  

This intuition is reflected in the idea of non-expressive use. In his influential article “Copyright and Copy–Reliant Technology”, Matthew Sag argues that technologies can copy copyrighted works for purposes such as search, indexing and analysis without ultimately communicating the protected expression to a human audience. Such a use may therefore be treated differently from copying that communicates the work’s protected expression.  

Courts have also recognised that copying may serve functions other than communicating a work’s expressive content. In Authors Guild v Google, the Second Circuit held that Google’s digitisation of books to create a searchable database constituted fair use. The court focused in part on the different function performed by the copies: they enabled users to search the books rather than providing substitutes for them. Similar reasoning appeared in Authors Guild v HathiTrust, which involved full-text digitisation for search and accessibility.  

The reading analogy is therefore useful, but it leaves an important question open. What happens when the function of copying is not retrieval or access, but the development of a reusable computational capability?   

By “reusable computational capability”, I mean a capability formed through training that can subsequently be applied to new inputs beyond the individual works used to develop it.  

A Search Index Is Not a Training Corpus  

Consider a copyrighted book incorporated into a search index and the same book incorporated into a model-training corpus.  

In the first case, the book helps the system locate information. The resulting index allows a user to identify a relevant book or retrieve information associated with the collection. The work remains functionally connected to retrieval.  

In the second case, the book becomes part of a training process through which a computational system develops or modifies a reusable computational capability. Foundation models illustrate this relationship: they are trained on broad data at scale to develop capabilities that can subsequently be adapted to a range of downstream tasks. Bommasani et al explain this model of broad training and adaptation, while Henderson et al examine the copyright and fair use questions raised by the development of foundation models.  

Both search indexing and model training may involve copying without communicating the original work to a human reader. Their technological functions, however, are different. The absence of expressive communication does not determine the function served by the copyrighted works. A copy may serve retrieval in one architecture and reusable capability formation in another.   

What Does the Work Do?  

This leads to what I describe as functional exploitation. AI training names a technological process; functional exploitation identifies the role that copyrighted works perform within that process.   

Functional exploitation occurs where copyrighted works serve as productive computational inputs in a process through which a reusable computational capability is formed or modified, and that capability is subsequently applied beyond the retrieval, display or communication of the individual works.  

The focus therefore shifts from what happens to a copyrighted work after it enters the system to the role it performs within the production process.  

That shift is important because AI development involves several different operations. A system may include:  

1. corpus construction;  

2. pre-training;  

3. fine-tuning;  

4. retrieval; and  

5. deployment.  

These stages can involve different relationships with copyrighted material.  

A search system may use copyrighted material to facilitate retrieval. A plagiarism-detection system may use it for verification. A screen reader may process it for accessibility. A model-training process may incorporate it into the formation of a capability that later operates on entirely new inputs.  

The same work can therefore occupy different functional positions at different stages of an AI system.  

Why This Matters for Copyright  

Recognising functional exploitation does not require existing copyright doctrines to be discarded. It provides a way of asking whether the precedents being invoked actually concern the same technological function.  

EU copyright law already provides separate exceptions for text and data mining. Articles 3 and 4 of the Digital Single Market Directive establish these exceptions subject to different conditions. Lawful access is relevant under both provisions, while Article 4(3) specifically permits rights holders to reserve the use of their works for text and data mining.  

In the United States, the legal treatment of AI training remains contested, with fair use at the centre of much of the debate. The US Copyright Office’s report on generative–AI training treats training, model development, outputs and licensing as related but distinct issues.  

The point of functional exploitation is not to predetermine how these legal questions should be answered. It is to ask whether legal reasoning developed in cases involving search, analysis or retrieval adequately describes the technological function of copyrighted works when they become inputs into systems that develop reusable computational capabilities.  

That question can affect how the legal analysis is framed across different legal frameworks. In the United States, for example, it may bear on the assessment of the purpose and character of the copying and its effects on relevant markets under fair use. In the EU, it may instead concern the applicability and scope of specific copyright exceptions, including those governing text and data mining.  

The broader point is that the technological function of copyrighted works may affect which precedents are genuinely analogous, which markets are relevant to the analysis, and whether existing copyright exceptions adequately address the activity.  

Training and Output Effects Are Different  

The distinction also helps separate the function of training from the effects of AI outputs.  

A model may be trained using copies of copyrighted works without the resulting model or its outputs reproducing any particular work used during training. Its outputs may nevertheless compete with human creators or creative services. Conversely, a model may reproduce protected material in its outputs without producing broad market substitution.  

The function of the input and the effects of the output are therefore related questions, but they are not the same one.  

Functional exploitation does not, by itself, establish infringement. Nor does it necessarily demonstrate market harm. It identifies the role played by copyrighted works within the system.  

Whether that role has legal or economic consequences requires a separate analysis.  

Rethinking AI Training  

The language of machines “reading” copyrighted works captures something important. AI systems can process information from works without communicating the works’ expressive content to users or experiencing them as expressive works in the way a human reader does. That insight remains relevant to copyright analysis.  

It does not, however, settle the question of what happens during training.  

When copyrighted works are used as inputs in training to develop or modify a reusable capability, they become part of the process through which that capability is formed. The concept of functional exploitation identifies this relationship. Whether that use is lawful, how applicable copyright exceptions should be interpreted, and whether creators should receive remuneration are questions that follow from identifying this functional relationship, but the concept does not, by itself, answer them.    

Functional exploitation contributes to copyright analysis by making the function performed by copyrighted works within AI training more visible. It distinguishes the use of works in capability-forming processes from uses in which copyrighted works remain tied to retrieval, display or communication, providing a basis for examining the legal and economic consequences of that distinction.  

 

Heading image by Brian Penny from Pixabay.