Global Conference on AI and Human Rights

Maja Bogataj Jančič opened the panel with a lecture on whether copyright can bring generative AI to its knees. In a short presentation, she showed how legal regulations around the world regulate the issue of whether it is allowed to train AI machines on big data or not. The differences between the systems of the United States of America (USA), Japan, the European Union (EU) and Slovenia governing how copyrighted works can be used to teach artificial intelligence were presented.

When an AI algorithm is trained, massive data input is required, including pure data and facts as well as personal data and copyrighted content.

In the US, training machines on huge amounts of data is carried out under the assumption that such use is legal without specific permission from the rights holders, because it meets the conditions of the so-called fair use, the standard or defense created by the courts. A landmark case is the NY Times v. OpenAI and Microsoft, which assesses whether the application of this principle is still appropriate. In the EU, we approach the challenge differently, we do not have a concept of fair use, which is assessed on a case-by-case basis, but exceptions and limitations in Articles 3 and 4 of the Directive on copyright and related rights in the Digital Single Market (DSM Directive). Slovenia implemented the articles in 2022 in a progressive manner. Exemptions on text and data mining (TDM) allow digitization of analog materials for TDM purposes and sharing of results. The exemption has teeth, as rights holders need to ensure that the beneficiaries of exemptions can effectively implement TDM and must provide access within 72 hours in the event of technological barriers or face sanctions. The weakness of the Slovenian provision is that the definition of legal access, as newly defined for TDM by the DSM Directive, has not been implemented. Thus, the old definition of legal access also applies to TDM, which does not include access to works that are freely available on the open web, which will cause great problems for researchers and others, as TDM will have essential data available to them.

The weakness of the Slovenian provision is that the definition of legal access, as newly defined for TDM by the DSM directive, has not been implemented. Thus, the old definition of legal access also applies to TDM, which does not include access to works that are freely available on the open web, which will cause great problems for researchers and others, as significantly less data will be available to them for TDM.

Maja concluded the lecture with the statement that we need good data governance for good AI, and for that we first need good governance to adopt good balanced legislation.

Dr. Zachary Cooper followed, with his lecture focused on the question of authorship of works produced by AI, enlivening it with a song created by one of the AI systems. Dr. Matija Damjan presented the problem of violating both moral and economic rights of performers when AI uses their video or voice recording to create new synthetic performances. The last lecturer was Dr. Žiga Škorjanc, who spoke about data for training AI from the point of view of collective rights management.

The panel was anchored by Assoc. Prof. Dr. Klemen Podobnik, who did not prove himself this time: he was impolitely late for the start of the panel, introduced the panelists superficially and stifled the discussion with too long of his own interventions. Despite these obstacles, the audience was enthusiastic and an excellent debate was sparked.