Who is responsible when AI speaks?

Google AI Overviews and the Rising differences in the EU AI Governance landscape

A debate came in for a sharp focus following a recent court order against Google's AI Overviews by a German court. The Munich Regional court said in this reported case that AI-generated summaries when used in Google Search are considered as statements made by Google and thus subject the company to liability in case of false information. At the same time, in what could be a defining ruling in AI liability, the court reportedly said that AI-generated summaries presented in Google Search are treated as Google's own statement, making it a matter of legal liability for the company for any misinformation created by its systems.

The question that is crucial to the debates about AI governance, around the world: Under what circumstances is a generative AI system responsible for the information it generates, if it produces inaccurate or harmful information?

It's not just Google: The decision has far-reaching implications. More fundamentally, this issue widens the range of concerns over the sufficiency of existing European regulatory frameworks. Instruments like the AI Act and the Digital Services Act (DSA) signal the European Union's aim to set itself up as a benchmark for AI governance globally, why then did a national court need to respond to the liability aspects of generative AI?

The answer might also signify a developing shortfall in the ongoing AI governance landscape of Europe.

From Search Engine to Content Generator

For the longest time, search engines were middlemen/intermediaries. They were to act as an index in recording information produced by other people and point users to relevant sources. This Scenario typically involved a publisher and not a platform when it comes to content. This dynamic is revolutionized by Generative AI.

Google's AI Overviews aren't just links to other pages. Rather, they’re capable of compiling knowledge from various resources and creating succinct responses within search results. With rising expectations of users, they are more liable to get a pre-made response before clicking on an external link. This difference is significant.

While traditional search systems allow access to existing resources, generative AI systems generate new content. What this results in is not a text duplication from a site, but a content creation done with probability models learned from a huge amount of data.

The German court reports that AI Overviews are statements that are “new and substantial” and can be classified as independent statements. That is, these summaries are new creations based on materials provided by Google's infrastructure and branded with Google's name.

So far as users are concerned, if they understand the outputs shown on Google as a Google representation of fact, then Google can be held responsible if the claims turn out to be false or misleading, the court said.

It is notable that this is a significant evolution in how courts might view AI-powered platforms as publishers of AI-generated content, rather than merely as intermediaries.

Bringing the Accountability Gap for Generative AI to a close

The decision comes at an opportune time for one of AI's biggest governance issues – the lack of accountability.

Large Language Models are probabilistic systems. They produce answers based on what they think the next words will be, not on whether they are right or wrong. Such systems are highly sophisticated, but still prone to “hallucinations,” confidently presented, but erroneous information.

These flaws are not just minor matter-of-fact mistakes. False statements created by algorithms can harm reputations, spread misinformation or cause economic damages. But proving liability is challenging. If it's a problem with the developers, deployers, users and/or the life of the model?

The principle of responsibility following from control has been emphasized by AI governance scholars. Businesses that develop, deploy and reap the profit from AI systems should bear additional responsibilities to reduce foreseeable harms.

So, the German court's decision seems to be in line with it. The court aims to define a platform (either the AI's or a human's) sole responsibility for the AI's output, avoiding the situation where the flow of harmful content operates without a clear human liability, and possibly even a financial one. The court is looking to prevent one of the platforms from having no responsibility for the harmful output, when it is the entire responsibility of the other platform deploying the content.

It is with this understanding that no-one is ever to be deterred from innovation. Instead, it is used to make sure technological development is not followed by mechanisms that can be used in cases of redress and accountability.

The Power of Consumers and The Authority of Algorithms

Another aspect of this is the perception and protection by consumers.

Google has a distinctive advantageous position in the worldwide info system. If information is presented in the prominent position on Google Search, if it is a summary created by AI, many users will likely consider the answer to be credible.

One area that has been examined in the context of algorithmic authority says that users tend to believe that the results given by automatically generated processes are objective or authoritative, particularly from trusted organizations. AI Summaries might thus unwittingly project an atmosphere of validity beyond its merits.

This poses very serious hazards.

When information is presented in authoritative ways with very little context, users may believe presented data without adequate consideration leading to real-world harm. Under these conditions, assigning a platform responsibility may be helpful to the implementation of stricter measures to protect their platforms, better transparency controls and better-quality checks.

A liability role, then, has two aspects, one corrective, the other preventative: giving incentives for safer use of AI.

An EC regulatory blind spot?

The most interesting part of this decision by the German court may be not the ruling itself, but the questions of regulation it brings out.

The EU has always billed itself as a worldwide leading example of digital governance. Brussels has attempted to create holistic and far-reaching governance frameworks for digital technologies and AI systems with the AI Act and the DSA.

The question is then why does this development show up before a German court and not rather than it could be handled at EU level? The poor design of current regulatory tools is one reason.

The key provisions of the DSA apply to online intermediaries, including transparency requirements, content moderation mechanisms, and systemic risks of very large online platforms. But the intermediaries may find their categories to be complicated by AI generative tools. However, if a platform creates content instead of just host-serving, it would be harder to apply the existing legal categories.

Likewise, the EU AI Act places a strong focus on “ex ante governance”. It sets forth standards on the matters of risk analysis, transparency and survival, documentation and standards before the systems enter the market. However, the Act does not address all the ex-post liability issues arising with AI-induced harm to a civil party. The German case, in this regard, could well have a blind spot in the regulation.

The AI Act establishes guidelines for developing and using AI systems. The DSA will mandate the management of risks online for platforms. However, both of those models do not solve a fundamental question that arises: if a generative AI generates fake data, to whom is responsibility clear under the law?

Thus, in the absence of other new bodies, national courts can be very important in filling that regulatory vacuum.

Innovation Versus Responsibility

Those who oppose a greater side liability for AI express concerns about the potential impact of overregulating it.

Generative AI systems come with set technical limitations. Whilst complete accuracy of the content might not be possible, the concept of strict liability would be counterproductive in encouraging companies to roll out technologies which they benefit from, or it could cause problems for smaller companies who are not equipped with the large compliance sections that larger firms are.

These concerns are important things to consider.

But in history, however, technology historical development and legal responsibility have been developing simultaneously. Carriers, the pharmaceutical industry, and digital platforms have introduced a wide range of new products needing to be understood legally, which means assigning responsibility and risk.

The question is not if, but how, to balance the allocation of responsibility and prevent the potential risk of harm to individuals and encourage innovation in the interest of AI.

The German court seems to be arguing that any entity with "significant influence" over an AI system and who is "enjoying commercial benefit" from it, is also expected to have "responsibilities".

Beyond Google: The Future of AI Governance

This battle over Google AI Overviews is not limited to just one website or court ruling, but it is a struggle between a string of companies, courts, and laws all striving for supremacy in the future of AI.

It's not only the copyright status of the AI-generated content that's at stake here, but the legal identity and future of generative AI. Is the speech of the entities the speech of a machine? The response will impact liability structures, the regulatory framework and innovation strategies for years to come.

Ultimately, the decision of the German court may point to a new principle in the governance of AI: competence and trustworthiness for platforms that display AI-generated content under their brand and name.

With the proliferation of generative AI in search engines, productivity tools, and public services, regulators and courts throughout Europe and elsewhere will be faced with the same tough questions. A challenge for policy makers isn't just in regulating AI but ensuring that responsibility grows in line with what AI can achieve.

With AI technologies advancing at an unprecedented pace, whether anyone can trust AI systems in the future may not only hinge on what they can do, but also who would be held responsible for their failures.