AG Szpunar opines on interplay between AI and press publishers’ rights in Like Company v Google

Advocate General Szpunar handed down his highly anticipated opinion in Like Company v Google on 3 November. This is the first Advocate General opinion considering the interplay between AI and IP rights (here, press publishers’ rights).

The Advocate General advised the CJEU to declare the referred questions relating to LLM training inadmissible, but nevertheless went on to address them, finding that the text and data mining (TDM) exception does apply in principle to copies made for the purposes of LLM training provided certain conditions are met.

Where, however, an LLM driven chatbot retrieves a publication from a press publisher’s website, reuses that publication to generate an answer, and provides that answer to the user, the press publishers’ right may be infringed in certain circumstances.

Background

This case relates to the training and use of Google’s Gemini chatbot and came up in the context of press publishers’ rights under Article 15 of the Directive on Copyright in the Digital Single Market (DSM Directive), rather than copyright per se.

Under Article 15(1) of that Directive, press publishers are given certain exclusive rights to reproduce their publications and make them available to the public. The scope of those rights is defined by reference to the equivalent rights of copyright holders to control reproduction and communication to the public of their works. However, press publishers’ rights are strictly limited to the ability to control “online use” of such publications by “information society service providers”.

Like Company issued proceedings against Google before the Budapest Regional Court, alleging that Google had infringed its press publishers’ rights by using its articles to train the Gemini model, to “ground” answers provided to users of the Gemini chatbot through retrieval-augmented generation (i.e. accessing information available on the internet in real time to anchor the response), and in the answers generated themselves.

In that context, the Budapest Regional Court referred four questions to the CJEU, which can be broadly summarised as follows:

1. Is the display, in the responses of an LLM chatbot, of text that is partially identical to protected content found on a publisher’s website an act of communication to the public? If so, is it relevant that the response provided by the chatbot is the result of a process in which the chatbot merely predicts the next word based on observed patterns?
2. Does the process of training an LLM-based chatbot engage the right of reproduction? 
3. If it does, does such reproduction of lawfully accessible works fall within the TDM exception in Article 4 of the DSM Directive?
4. Where a user’s prompt contains or refers to a press publication and the chatbot responds by generating all or part of that publication, does the resulting output constitute an act of reproduction attributable to the chatbot service provider?

Advocate General Szpunar’s opinion

LLM training

Unfortunately for those monitoring this case, AG Szpunar advised the CJEU to declare the questions relating to LLM training (questions 2 and 3) inadmissible. This was because he found them to be purely hypothetical for two reasons. Firstly, he found that there was nothing to support the assumption that Like Company’s articles had in fact been used to train the Gemini LLMs. Secondly, he found that the use of a press publication for the purposes of LLM training did not qualify as an “online use” of a press publication and therefore fell outside the scope of the press publishers’ right.

Thankfully, he still went on to address those questions “for the sake of completeness”, drawing attention to their complexity and advising the CJEU to proceed with caution. 

Perhaps unsurprisingly, and setting aside the question of territoriality (which, given its complexity and importance, the Advocate General ultimately suggested would be better addressed by the legislature), Advocate General Szpunar found that the use of protected works in LLM training does involve the making of copies of those works within the meaning of the InfoSoc Directive. However, he tended towards the view that the TDM exception would apply in principle to such copies, provided that: 

1. the works in question were lawfully accessed (e.g. taken from lawful sources and without circumventing technical measures such as paywalls); 
2. the relevant rightsholder had not opted their rights out of the TDM exception; and 
3. the three-step test under Article 5(5) of the InfoSoc Directive (which provides that exceptions to copyright and related rights must be restricted to (a) certain special cases which (b) do not conflict with a normal exploitation of the work or other subject matter used and (c) do not unreasonably prejudice the legitimate interests of the rightsholder) is satisfied on the facts of the relevant case.

Chatbot use, grounding and outputs

Turning to use of the chatbot, Advocate General Szpunar chose to address questions 1 and 4 together, broadly advising the CJEU to rule that when an LLM driven chatbot retrieves a publication from a press publisher’s website, reuses that publication to generate an answer, and provides that answer to the user, the press publishers’ right may be infringed in certain circumstances. 

With respect to reproduction (relevant to question 4), he suggested that whilst copies might be made during the grounding process, those copies may be covered by the temporary copies exception if they “enable a lawful use” within the meaning of Article 15(1) of the DSM Directive. Whether or not they do so will depend on two cumulative factors: 

(i) The way in which the LLM accessed and retrieved the original press publication. Was this done lawfully (and not, for example, by circumventing technical measures that may have been put in place)? 

(ii) The form of the answer provided to the user. Answers which resemble the form of expression of the relevant publication too closely may themselves qualify as copies falling within the press publishers’ right and thereby take the copies made during the grounding process outside the scope of the temporary copies exception.

Separately, the answers provided by the chatbot to the user may themselves be copies that might infringe the press publishers’ right. This will be the case where the protected elements of the relevant press publication are recognisable in the answer to such an extent that the user does not need to access the original publication (i.e. they are a “direct substitute”).   

As for whether the answers provided by the chatbot might infringe the right of communication to the public (relevant to question 1), the Advocate General cast doubt on this. In his view, a publication could only be regarded as being made available to the public through the chatbot if all or part of the publication were “systematically” recognisable (to the point of direct substitution) in the answers provided to related user queries – something he found was unlikely (but not impossible) given the non-deterministic technology involved.

Who bears responsibility for infringing outputs?

Finally, the Advocate General considered where liability would sit if infringements were found, landing on a nuanced answer. Where the user directs the chatbot to carry out a copyright-relevant act, primary liability should generally sit with the user, he said. Where, however, the chatbot provider “deliberately intervenes” to facilitate infringing acts, those acts should be regarded as also being carried out by the provider. This would be the case, Advocate General Szpunar said, where, for example, such acts are inherent to the provider’s business model or where the provider has failed to put in place “appropriate measures that can be expected from a reasonably diligent operator in its situation in order credibly and effectively to counter copyright infringements being committed by way of the chatbot”.

Comment

There is a lot of (thoughtful) ground covered in the Advocate General’s opinion and this blog only scratches the surface of what he said.

Whilst views will no doubt differ on the answers he has given, it is fascinating to see where he has landed.

In particular, it’s interesting that Advocate General Szpunar’s infringement analysis leads to a different place according to whether a model is answering from its trained parameters or where third party content is being retrieved at inference level. 

His views on responsibility for infringing outputs are also worthy of note. Whilst it is perhaps not surprising that he has suggested that who bears responsibility will ultimately depend on the facts of the case, his approach does appear to move the analysis away from a technical question of who carries out the restricted act towards an assessment of the provider’s behaviour and diligence.

One final point worth flagging relates to the Advocate General’s comments on remuneration schemes to compensate creators for uses falling within the scope of the TDM exception. Although a statutory remuneration scheme along these lines does not currently exist at EU level, the Advocate General said there is nothing to prevent Member States from introducing such a compensation system. It remains to be seen if the CJEU will echo that view.

It will, of course, be for the Hungarian national court to ultimately determine whether there has been infringement in this case, but the Advocate General’s provisional view is that Like Company’s press publishers’ rights were not infringed.

With the CJEU’s decision to follow, as well as two further referrals for the CJEU to tackle in Manchester United Fanatics Klub Magyarorszag v Google and Mobilguru v Google, we hope to get a clearer indication of the direction of travel over the coming months.