How can transparency work in practice?

As discussed in our Policy Paper #15, transparency is key to ensuring a fair balance between the interests of creators on the one hand and those of commercial AI developers on the other. A transparency obligation would empower creators, allowing them to assess whether the copyrighted materials used as AI training data have been scraped from lawful sources, as well as whether their decision to opt-out from AI training has been respected. At the same time, such an obligation needs to be fit-for-purpose, proportionate and workable for different kinds of AI developers, including smaller players.

While the European Parliament’s text has taken an important step towards improving transparency, it has been criticised for falling short in two key aspects. First, the proposed text focuses exclusively on training data protected under copyright law which arbitrarily limits the scope of the obligation in a way that may not be technically feasible. Second, the Parliament’s text remains very vague, calling only for a “sufficiently detailed summary” of the training data, which could lead to legal uncertainty for all actors involved, given how opaque the copyright ecosystem itself is.

As such, we are encouraged to see the recent work of the Spanish presidency on the topic of transparency, improving upon the Parliament’s proposed text. The presidency recognises that there is a need for targeted provisions that facilitate the enforcement of copyright rules in the context of foundation models and proposes that providers of foundation models should demonstrate that they have taken adequate measures to ensure compliance with the opt-out mechanism under the Copyright Directive. The Spanish presidency has also proposed that providers of foundation models should make information about their policies to manage copyright-related aspects public.

This proposal marks an important step in the right direction by expanding the scope of transparency beyond copyrighted material. Furthermore, requiring providers to share information about their policies to manage copyright-related aspects could provide important clarity as to the methods of opt-out that are being respected, empowering creators to be certain that their choices to protect works from TDM are being respected.

In search of a middle ground

Unfortunately, while the Spanish presidency has addressed one of our key concerns by removing the limitation to copyrighted material, ambiguity remains. Calling for a sufficiently detailed summary about the content of training data leaves a lot of room for interpretation and may lead to significant legal uncertainty going forward. Having said that, strict and rigid transparency requirements which force developers to list every individual entry inside of a training dataset would not be a workable solution either, due to the unfathomable quantity of data used for training. Furthermore, such a level of detail would provide no additional benefits when it comes to assessing compliance with the opt-out mechanism and the lawful access requirement. So what options do we have left?

First and foremost, the reference to “sufficiently detailed summary” must be replaced with a more concrete requirement. Instead of focussing on the content of training data sets, this obligation should focus on the copyright compliance policies followed during the scraping and training stages. Developers of generative AI systems should be required to provide a detailed explanation of their compliance policy including a list of websites and other sources from which the training data has been reproduced and extracted, and a list of the machine-readable rights reservation protocols/techniques that they have complied with during the data gathering process. In addition, the AI Act should allocate the responsibility to further develop transparency requirements to the to-be-established Artificial Intelligence Board (Council) or Artificial Intelligence Office (Parliament). This new agency, which will be set up as part of the AI Act, must serve as an independent and accountable actor, ensuring consistent implementation of the legislation and providing guidance for its application. On the subject of transparency requirements, an independent AI Board/Office would be able to lay down best-practices for AI developers and define the granularity of information that needs to be provided to meet the transparency requirements set out in the Act.

We understand that the deadline to find an agreement on the AI Act ahead of the next parliamentary term is very tight. However, this should not be an excuse for the co-legislators to rush the process by taking shortcuts through ambiguous language purely to find swift compromises, creating significant legal uncertainty in the long run. In order to achieve its goal to protect Europeans from harmful and dangerous applications of AI while still allowing for development and encouraging innovation in the sector, and to potentially serve as model legislation for the rest of the world, the AI Act must be robust and legally sound. Everything else would be a wasted opportunity.

The post The transparency provision in the AI Act: What needs to happen after the 4th trilogue? appeared first on COMMUNIA Association.

Over the course of 20 minutes, Teresa walks the listeners through some of the key questions when it comes to the training of generative AI models and how it affects the text and data mining (TDM) exception laid down in Articles 3 and 4 of the EU Copyright Directive (see also our Policy Paper #15).

Disclaimer: Playback of the embedded video establishes a connection to YouTube and may lead to data being collected by and shared with third parties. Proceed only if you agree.

While the Directive clearly establishes the right to mine online content and use it to train machine learning algorithms, this right hinges on the possibility for rightsholders to opt out their works if the activity takes place in a commercial context. A key issue we are currently seeing is that there is a lack of transparency around the training of generative AI models, which makes it impossible to tell whether such opt-outs are being respected or not.

In the discussion, Teresa highlighted the need for more transparency regarding opt-outs but also across the copyright ecosystem as a whole. COMMUNIA has long advocated for the creation of a database of copyrighted works, which would contribute to managing the system of opt-outs.

The discussion ended with a call upon the European Commission to provide guidance and lead technical discussions towards establishing a clear, reliable and transparent framework for opt-outs for TDM (see also Open Future’s recent policy brief on this issue). Only through dialogue between the concerned stakeholders, led by an independent third party, will we be able to establish best practices that uphold Articles 3 and 4 of the Copyright Directive while providing a fair and balanced framework for the training of machine learning models.

The post We need to talk about AI and transparency! appeared first on COMMUNIA Association.

In their new Policy brief on defining best practices for opting out of ML training published today, our member Open Future takes a closer look at a key element of the text and data mining exception: The rights reservation mechanism foreseen in Article 4(3) of the Directive, which allows authors and other rights holders to opt out of their works being used to train (generative) ML models.

The policy brief highlights that there are still many open questions relating to the implementation of the opt out mechanism. That is, how the machine-readable reservation of rights under Article 4 will work in practice. One of the key issues in this context is the fact that currently there are no generally accepted standards or protocols for the machine-readable expression of the reservation. The authors of the policy brief provide an overview of existing initiatives to provide standardized opt-outs which include initiatives by Adobe and a publisher-led W3C community working group, as well as the artist-led project Spawning, which provides an API that aggregates various opt-out systems. In addition, they highlight a number of proprietary initiatives from model developers, including Google and OpenAI.

Lack of a technical standard

According to the policy brief, one of the key problems facing creators and other rights holders who wish to opt out of ML training is that it is unclear whether and how their intentions to opt-out will be respected by ML model developers. According to the authors of the policy brief, this is deeply problematic and risks undermining the legal framework put in place by the 2019 Copyright Directive:

Continued lack of clarity on how to make use of the opt-out from Article 4 of the CDSM Directive creates the risk that the balanced regulatory approach adopted by the EU in 2019 might fail in practice, which would likely lead to a reopening of substantive copyright legislation during the next mandate. Given the length of EU legislative processes, this would prolong the status quo and, as a result, fail to provide protection for creators and other rightholders in the immediate future.

It seems clear that this scenario should be avoided, both in the interest of creators, who need to have meaningful tools to enforce their rights vis-à-vis commercial ML companies, and in order to preserve the hard-won compromise reflected in the TDM exceptions.

In line with this, the Open Future policy brief calls on the European Commission to “provide guidance on how to express machine-readable rights reservations”. According to the brief, the Commission needs to step in and “publicly identify data sources, protocols and standards that allow authors and rightholders to express a machine-readable rights reservation in accordance with Article 4(3) CDSM”. This guidance would provide important clarity about the availability of freely usable methods of reservation and certainty as to their functionality.

According to the authors, such an intervention would allow the Commission to support creators and other rightholders seeking means to opt out of ML training, while at the same time providing “more certainty to ML developers seeking to understand what constitutes best efforts to comply with their obligations under Article 4(3) of the CDSM Directive.

Time to act

The Open Future policy identifies an important shortcoming in the existing EU approach to the use of copyrighted works for ML training. Without clear guidelines for standardized machine-readable rights reservations, the opt-out mechanism foreseen in Article 4 is unlikely to work in practice. While there are a number of existing standards, the fragmentation of this system causes tremendous uncertainty for creators.

As Open Future points out, it is up to the Commission, which is responsible for ensuring the proper implementation of the Directive’s provisions, to intervene in this area and provide initial clarity to all stakeholders. In the longer term, it would be ideal to see the emergence of an open standard that is maintained independently of any direct stakeholders. Such an effort should not be limited to aggregating opt-outs, but should also be designed to ensure that works in the public domain or made available under licenses that allow and/or encourage reuse are clearly identified as such (In line with our Policy Recommendation #20).

The post Defining best practices for opting out of ML training – time to act appeared first on COMMUNIA Association.

One of the key issues for the discussions will most likely be how to deal with the rather recent phenomenon of generative AI systems (also referred to as foundational models) which are capable of producing various content ranging from complex text to images, sound computer code and much more with very limited human input.

The rise of generative AI

Within less than a year, generative AI technology went from having a select few, rather niche applications to becoming a global phenomenon. Perhaps no application represents this development like ChatGPT. Originally released in November 2022, ChatGPT broke all records by reaching one million users within just five days of its release with the closest competitors for this title, namely Instagram, Spotify, Dropbox and Facebook, taking several months to reach the same stage. Fast forward to today, approximately half a year later, and ChatGPT reportedly counts more than 100 million users.

One of the reasons for this “boom” of generative AI systems is that they are more than just a novelty. Some systems have established themselves as considerable competitors for human creators for certain types of creative expressions, being able to write background music or produce stock images that would take humans many more hours to create. In fact, the quality of the output of some systems is already so high while the cost of production is so low that they pose an existential risk to specific categories of creators, as well as the industries behind them.

But how do generative AI systems achieve this and what is the secret behind their ability to produce works that can comfortably compete with works of human creativity? Providing an answer to this question, even at surface level, is extremely difficult since AI systems are notoriously opaque, making it nearly impossible to fully understand their inner workings. Furthermore, developers of these systems have an obvious interest in keeping the code of their algorithm as well as the training data used secret. This being said, one thing is for certain: generative AI systems need data, and lots of it.

The pursuit of data

Creating an AI system is incredibly data intensive. Data is needed to train and test the algorithm throughout its entire lifecycle. Going back to the example of ChatGPT, the system was trained on numerous datasets throughout its iterations containing hundreds of gigabytes of data equating to hundreds of billions of words.

With so much data needed for training alone, this opens up the question how developers get their hands on this amount of information. As is fairly obvious by the sheer numbers, training data for AI systems is usually not collected manually. Instead, developers often rely on two sources for their data: curated databases which contain vast amounts of data and so-called web crawlers which “harvest” the near boundless information and data resources available on the open internet.

The copyright conundrum

Some of the data available in online databases or collected by web scraping tools will inevitably be copyrighted material which raises some questions with regards to the application of copyright in the context of training AI systems. Communia has extensively discussed the interaction between copyright and text and data mining (TDM) in our policy paper #15 but just as a short refresher about the clear framework established in the 2019 Copyright Directive:

Under Article 3, research organizations and cultural heritage institutions may scrape anything that they have legal access to, including content that is freely available online for the purposes of scientific research. Under Article 4, this right is extended to anyone for any purposes but rights holders may reserve their rights and opt out of text and data mining, most often through machine-readable means.

While this framework, in principle, provides appropriate and sufficient legal clarity on the use of copyrighted materials in AI training, the execution still suffers from the previously mentioned opacity of AI systems and the secrecy around training data as there is no real way for a rightsholder to check whether their attempt to opt out of commercial TDM has actually worked. In addition, there’s still a lot of uncertainty about the best technical way to effectively opt out.

Bringing light into the dark

Going back to the EU’s AI Act reveals that the European Parliament recognises this issue as well. The Parliament’s position foresees that providers of generative AI models should document and share a “sufficiently detailed” summary of the use of training data protected under copyright law (Article 28b). This is an encouraging sign and a step in the right direction. The proof is in the pudding, however. More clarity is needed with regards to what “sufficiently detailed” means and how this provision would look in practice.

Policy makers should not forget that the copyright ecosystem itself suffers from a lack of transparency. This means that AI developers will not be able – and therefore should not be required – to detail the author, the owner or even the title of the copyrighted materials that they have used as training data in their AI systems. This information simply does not exist out there for the vast majority of protected works and, unless right holders and those who represent them start releasing adequate information and attaching it to their works, it is impossible for AI developers to provide such detailed information.

AI developers also should not be expected to know which of their training materials are copyrightable. Introducing a specific requirement for this category of data adds legal complexity that is not needed nor advisable. For that and other reasons, we recommend in our policy paper that AI developers be required to be transparent about all of their training data, and not only about the data that is subject to copyright.

The fact that AI developers know so little about each of the materials that is being used to train their models should not, however, be a reason to abandon the transparency requirement.

In our view, those that are using publicly available datasets will probably comply with the transparency requirement simply by referring to the dataset, even if the dataset is lacking detailed information on each work. Those that are willing to submit training data with a data thrust that would ensure the accessibility of the repository for purposes of assessing compliance with the law would probably also ensure a reasonable level of transparency.

The main problem is with those that are not disclosing any information about their training data, such as OpenAI. These need to be forced to make some sort of public documentation and disclosure and at least need to be able to show that they have not used copyrighted works that have an opt-out attached to it. And that begs for the question: how can creators and other right holders effectively reserve their training rights and opt-out of the commercial TDM exception?

Operationalizing the opt-out mechanism

In our recommendations for the national implementation of the TDM exceptions we suggested that the proper technical way to facilitate web mining was by the use of a protocol like robot.txt which creates a binary “mine”/“don’t mine” rule. However, this technical protocol has some significant limitations when it comes to its application in the context of data mining for AI training data.

Therefore, one of the recommendations in our policy paper is for the Commission to lead these technical discussions and provide guidance on how the opt-out is supposed to work in practice to end some of the uncertainty that exists among creators and other rights holders.

In order to encourage a fair and balanced approach to both the opt-out and the transparency issues, the Commission could convene a stakeholder dialogue and include all affected parties, namely AI developers, creators and rights holders as well as representatives of civil society and academia. The outcome of this dialogue should be a way to operationalise the opt-out system itself and the transparency requirements that will uphold such a system without placing a disproportionate burden on AI developers.

Getting this right would provide a middle ground that allows creators and other rights holders to protect their commercial AI training rights over their works while encouraging innovation and the development of generative AI models in the EU.

The post The AI Act and the quest for transparency appeared first on COMMUNIA Association.