At Rise of AI Summit 2020 in Berlin, Dr. Ursula Feindor-Schmidt, a partner at Lausen Rechtsanwälte, surveys European Union proposals for regulating artificial intelligence. She starts from the European Commission's AI High-Level Expert Group definition of an AI system as software, designed by humans, that perceives its environment, interprets data and acts toward a given goal. She notes AI already falls under existing rules such as the General Product Safety Directive, GDPR and sector standards, but argues lawmakers see gaps around security, data-driven bias, autonomy, opacity and unclear liability chains. She reviews initiatives on the table: the European Commission's White Paper on AI, a European Parliament report on AI, IoT and robotics liability, and the upcoming Digital Services Act, which may include obligations for high-risk applications such as biometric identification, conformity labeling, human-contact points and audits of training data quality. In the closing minutes, drawing on her own intellectual-property practice, she covers whether AI itself can be patented or copyrighted, whether training on copyrighted data infringes text-and-data-mining exceptions, and whether AI-generated output can hold copyright, concluding it currently cannot because only humans can be copyright holders.
Key takeaways
- She uses the EU AI High-Level Expert Group's definition of AI as human-designed software that perceives, reasons and acts toward a goal, covering simple algorithms too.
- She argues existing law such as GDPR and product safety rules already applies to AI, but gaps in liability, opacity and autonomy are driving new rules.
- She expects the Digital Services Act and a risk-based framework to require audits, labeling, human contact points and training-data quality checks for high-risk AI.
- She says liability obligations may fall on whichever actor in the value chain is best placed to address a risk, not only the developer.
- As an IP lawyer, she concludes AI-generated output currently cannot hold copyright because only a human can be a copyright holder.
Detailed notes
Defining AI for regulatory purposes
Feindor-Schmidt opens with the European Commission's AI High-Level Expert Group definition: AI systems are software, and possibly hardware, designed by humans to pursue a complex goal by perceiving their environment through data acquisition, interpreting structured or unstructured data, reasoning on collected knowledge and acting in a physical or digital dimension. She stresses this definition covers not only "strong AI" but also simple algorithms that decide, for example, which advertisement to show a targeted group.
Existing rules and why regulators want more
She lists regulation already covering AI products, including the General Product Safety Directive of 2001, GDPR, the Consumer Rights Directive, sectoral rules like the Machinery Directive, and ISO/IEC software standards. Lawmakers nonetheless see gaps: enhanced security risks from AI's connectivity and openness, data dependency that can cause bias and false results, autonomous decisions not scrutinized step by step by a human, public unease about opacity, and product-liability and fault-based liability rules that do not fit services or autonomous decision-making well.
What's in the pipeline: White Paper, IP report and the Digital Services Act
She reviews the European Commission's White Paper on AI, the European Parliament's report on AI, IoT and robotics safety and reliability, a European Parliament report on intellectual property for AI development, and the Digital Services Act, which was in consultation at the time and expected in a first draft "beginning of December." She also notes WIPO conversations on IP and AI at the international level and Germany's national AI strategy, arguing that purely national strategies make little sense without international alignment since companies sell internationally.
Likely obligations for high-risk AI
She expects EU regulation to lean risk-based rather than purely sector-based, applying stricter rules to high-risk applications such as biometric identification and sectors like healthcare, transport, energy and parts of the public sector. Likely obligations include external regulatory audits, explainability, conformity assessment and labeling, informing authorities and users about AI interaction, human contact points and redress systems, quality and neutrality checks on training data, and new rules on accountability, liability and burden of proof, which could apply to any actor in the value chain best placed to address a risk, not only developers or deployers.
AI and intellectual property
In her closing segment as an IP lawyer, Feindor-Schmidt asks whether AI itself can be protected: a patent requires a technical effect, so pure algorithms as mathematical methods are not patentable, and copyright for software requires human intellectual creativity. On IP as "food" for AI, she notes text-and-data-mining exceptions to copyright are narrow, so training on copyrighted content such as articles or photos can be an infringement. On whether AI can create IP, she says current copyright protection requires a human creator, so AI-generated output is not copyrighted, though article 17 content-identification obligations could let AI help enforce IP on platforms.
Quotes
Regulation always means higher expenses, legal compliance expenses, that's clear, but regulation could also save money because it means clarity on responsibility.
Ursula Feindor-Schmidt
The pipeline is long and the pipeline is packed.
Ursula Feindor-Schmidt
Questions answered in this talk
How does the EU define an AI system for regulatory purposes?
According to Dr. Ursula Feindor-Schmidt at Rise of AI Summit 2020, the European Commission's AI High-Level Expert Group defines it as human-designed software, and possibly hardware, that pursues a given goal by perceiving its environment, interpreting data, reasoning and acting, which covers simple algorithms as well as advanced systems.
What new obligations might high-risk AI face under EU law?
Feindor-Schmidt says likely obligations include external audits, explainability, conformity assessment and labeling, informing authorities and users, human contact and redress points, and checks on the quality and neutrality of training data, especially for applications like biometric identification and sectors such as healthcare and transport.
Can AI-generated content be protected by copyright?
No. Feindor-Schmidt explains at Rise of AI Summit 2020 that current copyright law requires a human creator, so output produced by AI is not eligible for copyright protection, though there is discussion of assigning ownership to whoever prepared and published the AI-assisted work.