TDM exceptions and trust in AI
Kluwer Copyright Blog
February 17, 2021
Please refer to this post as:, ‘TDM exceptions and trust in AI’, Kluwer Copyright Blog, February 17 2021, http://copyrightblog.kluweriplaw.com/2021/02/17/tdm-exceptions-and-trust-in-ai/
Will the text and data mining (TDM) exceptions, introduced in arts 3 and 4 of the EU Directive on Copyright in the Digital Single Market (DSM Directive) and currently being implemented by the EU Member States, serve its purpose of promoting the development of AI technologies or will they remain (another) set of meaningless black letter rules? This question came to mind when reading about the recent dispute between the US Federal Trade Commission and a Californian company, Everalbum Inc.
FTC v Everalbum
According to the facts of the case, Everalbum offered its users an app called “Ever” that allowed them to upload photos and videos from their mobile devices, computers, or social media accounts to be stored and organized using the company’s cloud-based storage service. Its feature “Friend” used facial recognition technology to group users’ photos by the faces of the people who appear in them and allowed users to “tag” people by name. While Everalbum promised users that it would not apply facial recognition technology to users’ content unless users affirmatively chose to activate the feature, they breached this promise. They turned on the facial recognition feature for some consumers’ content, without an option to turn it off. Further, Everalbum combined millions of facial images that it extracted from Ever users’ photos with images that it obtained from publicly available datasets to create its own datasets. At least one of these datasets was used in the development of its facial recognition technology, which was then sold to Everalbum customers.
The US Federal Commission and Everalbum settled the case with harsh consequences for Everalbum: it had to delete not only all datasets created from photos uploaded by users, but also facial recognition models and algorithms developed by the company.
Copyright and TDM discussion
While the dispute is based on misleading conduct rather than copyright law, it shows general trends in the AI industry that will have an effect on the TDM exceptions under copyright law, and its actual impact in practice.
As we know, as a result of the TDM exceptions, the use of content in machine learning is supposed to become largely legal. It means that AI developers can use copyright protected content – such as photos – without authorization from authors or other right holders, unless right holders explicitly prevent commercial TDM (art 4(3) of DSM Directive). You would think that AI developers are now fine to take and freely use copyright protected content in their projects.
However, as can be seen from the Everalbum case, there are at least two other obstacles they need to overcome. On the one hand, there are privacy laws that might add another layer of protection and require permissions in some jurisdictions, like the EU.
On the other hand, even in jurisdictions where privacy laws are weaker (e.g., the US) or in situations where they do not apply (e.g., the content is not personal), another issue arises: trust in AI. AI companies are working hard to build trust in AI (see this KPMG Trustworthy AI model as an example) which would lead to a broader acceptance of AI technology in society. In order to develop trust in AI, they seek to meet users’ expectations and might promise more than that required by law, such as permissions to use their content for particular purposes (in the Everalbum case they promised to get an explicit consent, which, it seems, was not required under US laws). Asking users for permission to use their data and content means showing respect and courtesy to users, and, as the Everalbum case demonstrates, is becoming increasingly important in the industry where data is seen as ‘fuel’ in the AI revolution.
This essentially means that while the TDM exceptions allows AI developers to use content freely, the need of AI companies to ensure trust in their technologies might mean that they will still want need permission from users to use their content. With the further development of AI and ongoing discussions on who owns or should own data, such permissions from users (content owners) might become a rule. If this happens, TDM exceptions might lose their practical relevance.