Examining generative image models amidst privacy regulations

1 Background

Stability AI released Stable Diffusion, a text to image model that allows users to prompt an algorithm to create a work emulating a particular artist or including certain visual elements. Stable Diffusion, as well as other emerging generative image models such as Midjourney and Dalle2, were trained on a segment of LAION 5B, a dataset of text image pairs compiled through web crawling across the internet (What is the Super Huge Data Set “LAION-5B” That Made a Great Contribution to the Development of Image Generation AI “Stable Diffusion”? 2022). LAION has evaded privacy regulations by being not a dataset of images itself, but rather, an index of links to images that can be used for model training purposes. As a result, LAION is able to push the burden of accessing data to whichever entity is seeking to web-scrape; in this instance, these entities are image generative models. Though the recent LinkedIn v. hiQ ruling has established precedent for legalizing web-scraping publicly available data within a certain proprietary context, there has not been precedent established for web-scraping such a large public domain to be monetized later down the pipeline. This claim, that the data scraped through web crawling, is truly public information is brought to question in this paper (“HiQ Labs v. LinkedIn” 2022). The ambiguity has prompted a slew of lawsuits, with both individual artists and companies, such as Getty Images, outraged over the use of their images to train these models, and calls for reimagining regulations in light of the new technologies that have arisen.

2 Methodology

This paper is a meta analysis review of existing literature, legal sources, and empirical case studies. The primary case study this paper focuses on is the Stable Diffusion data scraping scandal, which was chosen given that its case inspired the questions at the heart of this paper. The topics explored (public/private distinction, legal bases/rights, monopolization concerns) flowed as a natural logical extension of the primary case study. Auxiliary case studies chosen had to speak to the issue of personal/public distinction, an inconsistency with the legality of its justification for use (legal bases), issues with legal rights not preserved, or monopolization concerns that could result from such models. I chose case studies the most relevant case studies that most clearly articulated each relevant section or highlighted a point of contention with another case study.

The literature referenced were selected based on their contributions towards different conceptions of data privacy and competition policy. I chose literature that succinctly exemplified the tension between the two incentives, and fit within the discussion of generative image models. Literature also included news articles of the primary case study (the LAION 5B data scraping/Stable Diffusion incident), which helped explain the fundamental perspectives of different stakeholders. Legal sources were used to explain the existing privacy regulations, such as the CCPA, GDPR, and a number of non-legislative informal regulations made through FTC rulings.

3 Public or personal data?

Stable Diffusion has claimed their images are open source, which implies these images are not personal data. Their claim is not yet supported or opposed by regulatory decision makers or legal precedent. Several privacy frameworks are available for analyzing the legality of gathering imaging data. These frameworks consider a range of factors, including whether this data qualifies as artistic intellectual property or falls under the category of proprietary data (such as trade secrets), which might have the potential to compromise the business interests of companies. This paper will focus on the public/personal distinction as conceptualized under the California Consumer Protection Act (CCPA) and the EU’s General Data Protection Regulation (GDPR), since companies like Stability AI are conducting their business as if the data they scraped fell under public domain within these schemas. en

Under GDPR, personal data is any data that is “related to an identified or identifiable natural person”(“Art. 4 GDPR – Definitions,” 2016). This means that data is considered personal if it can be used to identify an individual, even if there is not a name attached to it. Different pieces of an individual’s seemingly unrelated data are considered personal if, in conglomeration, they can be used to identify said individual. Similarly, under the CCPA, personal information is data that “identifies, relates to, describes, is reasonably capable of being associated with, or could reasonably be linked, directly or indirectly, with a particular consumer or household” (California Consumer Privacy Act 2020).

The input dataset for Stable Diffusion’s model is the dataset from LAION 5B, which contains 5.8 billion image-text pairs and can be considered personal data under CCPA and GDPR. The LAION 5B dataset focuses on photos it ranks as “aesthetic” based on an aesthetic predictive score (What Is the Super Huge Data Set “LAION-5B” 2022). Frequent source domains of these aesthetic images include Pinterest, Shopify, Redbubble, Etsy, Wikimedia, DeviantArt, and stock image sites. The text-image pairs have enabled an indexing into an unprecedented quantity of images, and the ease with which these images can be accessed raises concerns given that there is not yet conclusive legal precedent regarding linked or indexed data. Images linked within the dataset arguably fall under GDPR and CCPA provisions as personally identifiable: Pinterest images and pictures of models from clothing websites undoubtedly contain faces, and an artist’s work on their online store may have their signature or be in a style so distinct that it is identifiably them (Baio 2022).

For the text-image pairs in the LAION 5B dataset, the potential for identifiability exists within a spectrum, illustrating the limitations of binary data privacy classification. Some artists are blatantly identifiable with their names attached in the text, as has been the case for artists Karla Ortiz and Greg Rutkowski (Hill 2023), who have joined the class action lawsuit against Stability AI. For artists whose identifiability is their style, the potential for identifiability is more ambiguous.

Although there are still considerable ambiguities to be resolved in the public/private distinction of information, the usefulness of this binary distinction on the basis of identification may be limited due to technological advances in de-anonymization methods. Some privacy scholars have claimed that the framework of personal information, in being based on identification in both GDPR and CCPA statutes, appears increasingly inconsequential as computer scientists find that anonymized data is easily de-anonymized, and data in conjunction with other data can be used to easily identify an individual (Tene and Polonetsky 2013). In a study examining the identifiability of anonymous authors, Narayanan et al. (2012) were able to re-identify anonymous authors solely based on the author’s style. Though anonymized data in the past has allowed businesses the ability to pursue ventures while still maintaining some semblance of respecting privacy, it’s become clear as the complexity of data collection has increased, anonymization/de-identification is not a permanent condition (Purtova 2018). In order to make such a point, one privacy scholar, Purtova, conducted a case study including weather information collected from a weather station in the Dutch city of Eindhoven. Because the weather station, which alone collects rainfall, wind speed, wind direction, and temperature, can be tied to other city crime prevention data such as audio sensors and video footage, Purtova explains how individuals can be identified from such data. She argues that weather, a seemingly innocuous piece of information, has become personal. Purtova concluded “irreversible anonymity is no longer possible” (Purtova 2018).

Furthermore, the personal nature of these images may be expanded simply through the added dimension of information that these images present–their city/location, and their interests. However, it is increasingly challenging to distinguish personal data using this binary, as the rather expansive datification of life would mean almost all information could be deemed personal, thereby nullifying public/private distinction’s power. A GDPR provision does claim that identifiability remains dependent on the reasonability of an individual being identified, effectively steering the conversation towards context dependence in determining the public-personal binary. However, “as the data processing technologies advance, and the pool of data which can be combined grows … so does the reasonable likelihood of somebody being able to link any piece of information to a person” (Purtova 2018). As Tene and Polonetsky have explained, a more useful framework for explaining the harm and risk potential of identifying an individual could be in the form of a risk assessment matrix or some other continuous taxonomy (Tene and Polonetsky 2013). In the case of the images in the LAION 5B dataset, a privacy framework with different thresholds depending on the severity of harm if identified could be a useful conceptualization.

4 Legal bases and rights dependent on PII distinction

In spite of the fact that increasing ease of identifiability is blurring the distinction between public and personal data, data protection laws such as GDPR and CCPA that rely on this distinction still act as the primary legal foundation for prosecutors seeking to establish red lines for data controllers. There are different legal bases for data collection and legal rights for data owners under the CCPA and GDPR depending on whether data is classified as personal or public. In the EU, GDPR differentiates between personal and public data for legal bases of data collection, and a company must have a legal basis for processing data if it is categorized as personal. However, under US law, no such legal basis is necessary to process personal data, though there is an expectation that data collection is not “unfair or deceptive” if the data is collected from consumers. Under both the CCPA and GDPR, data deemed personal entitles data subjects to certain legal rights. Though scholars have acknowledged the public/personal distinction between data has become rather dated, it is helpful to examine the legal bases and legal rights still dependent on this classification, especially in the case that the public/personal schema remains. In addition, it’s important to examine where a web scraped data source such as the LAION 5B one would fall amidst existing privacy regulation, as privacy authorities have yet to decide where the data collection falls amidst this framework.

5 The monopolization problem

Whether we allow for open sourcing of images depends on the classification of such data in the aforementioned public/personal categories. If regulatory agencies do side with companies like Stable Diffusion, then we can examine the competition and privacy tradeoffs for this decision. The public/private classification of data as specified by GDPR and CCPA has a significant downstream effect on the competitive landscape and the applicability of privacy regulations. Additionally, the current conception of such a distinction is dependent on decisions that regulatory agencies make as technologies push the boundaries of the meaning of regulations. Thus, it only makes sense that such a decision is not made in a vacuum of outdated black letter criteria, but also within the broader context of privacy norms, expectations, and consequences.

6 Conclusion

It’s important to note that the alternatives in this piece are not exhaustive nor are they necessarily prescriptive. At this stage, it is difficult to imagine a viable alternative given the complexity. However, it’s crucial that the lack of an easily understandable solution does not discourage examination of privacy frameworks as new technologies that do not fit squarely within our frameworks arise.

As image generative models circumvent copyright laws through the fair use clause, I look at how privacy laws could be applied if artists’ works are imagined as personal, rather than public data. Legal bases would be required for such a collection. This could provide artists a protection against having their images used against their wishes, especially if companies choose to use consent as their legal basis. Additionally, the requirement of providing individuals legal rights to their imaging data could mean a renewed dedication to tackling the delearning problem or an increased dedication to ensuring consent was given in the first place so as to mitigate the amount who opt out later.

Accounting for privacy matters does come at a technical cost; inhibiting image datasets from being interpreted as publicly available will likely decrease its size and range, hurting the training capability of these models. However, it should be noted that the quality of artwork artists create already determines the quality of the model. Accounting for privacy concerns of the artists whose works determine the quality of the model may come at a technical cost, but perhaps a necessary one, given the extent that artists are stakeholders.

Lastly, interpreting these datasets as public or not creates monopolization implications. Though privatizing these datasets may appear to further entrench the market position of dominant players and restrict competition, remedies through antitrust law and other state regulation for such instances can be implemented as a preventive measure.