Which Large Language Model works best?
In recent years, the increasing popularity of e-commerce platforms has seen more and more products launched and advertised online. Unfortunately, this development has been accompanied by an increase in corporate activities that undermine consumer rights. Everything from misleading product descriptions to dishonest advertising claims that both endanger consumers’ rights and disrupt fair competition can be observed today. For example, according to a 2014 report, about 37% of internet commerce players in the EU disregard the Union’s consumer laws. This results in significant financial damage to consumers:inside, worth an estimated €770 million annually.
It is the responsibility of both consumers themselves and consumer protection agencies, including consumer protection centers and central competition centers, to identify these violations. Once identified, consumer centers typically obtain a legally binding cease-and-desist letter - a contract in which the offending company promises to terminate the illegal behavior and refrain from future violations. However, current processes for monitoring compliance with applicable laws are time-consuming, costly, and labor intensive, as they typically must be performed manually. As a result, actual compliance with the cease-and-desist letter often goes unchecked, allowing many companies to continue violating consumer rights.
In collaboration with our partners vunk and Law & Audrey, we have launched the KIVEDU research project, which addresses this problem using AI technologies, in particular large language models. By identifying consumer rights violations using AI, we strengthen the rights of consumers, facilitate work of consumer protection centers, and thus contribute to fair competition. In addition to automated detection of violations, the system is also used to collect and archive evidence of violations that can be used in court.
„The KIVEDU system developed in our project uses AI to identify and document violations of cease-and-desist agreements. An effective support for consumer protection!“
To kick off the KIVEDU project, we first conducted a comprehensive requirements analysis to understand the needs of consumer protection centers and define the requirements for the system. This included interviews with various consumer protection centers and commercial enterprises, which helped to shed light on current processes for monitoring compliance with cease-and-desist letters. Based on this, we created a requirements document that outlined the necessary features of the system and served as a foundation for further development.
We also addressed the technical and legal challenges of developing AI systems for consumer enforcement. This included an extensive literature review and interviews with a wide range of experts.
We summarized our findings in a scientific article that will be published later ths year as part of the 15th conference “PLAIS EuroSymposium on Digital Transformation”.
The development of an AI system for consumer law enforcement involves a number of challenges. These can be divided into two categories: legal and technical.
General: Legal requirements are fundamental to the development and application of a new system. Non-compliance can lead to prohibitions, rejections, or legal liabilities. We therefore focus on “by-design” lifecycle compliance.
EU GDPR (DSGVO): Our software architecture is mainly based on processing data, so data protection laws are at the heart of our legal assessment. We need to ensure compliance and consider the rights of data subjects.
EU AI regulation: In addition to data protection, we also pay attention to legal developments regulating AI systems. The AI Regulation aims to ensure that AI systems are transparent, reliable and secure.
Intellectual property: When developing AI software, developers should understand and protect their intellectual property rights. This includes, for example, using non-disclosure agreements and respecting the rights of vendors when using open source applications.
LLM selection: Selecting an appropriate large-scale language model (LLM) is a complex process. Models can differ significantly in several aspects, especially performance, processing speed, cost, licensing model, and privacy guarantees.
Explainability of AI: A key problem lies in the inherent complexity of decisions. Because systems are often based on complicated and high-dimensional data representations, it can be difficult to understand the logic behind their decisions.
Archiving evidence of violation: A significant challenge is not only identifying violations, but also subsequently archiving and preserving credible evidence. It is essential to store the evidence in a way that prevents any form of deletion or tampering.
Current status & outlook
In addition to publishing a scientific article on our project, we have already developed a first prototype of the KIVEDU system. This prototype enables the upload of a cease-and-desist declaration and the automatic check of a manually entered URL for violations of the uploaded declaration. In case of a violation, a screenshot of the web page is automatically created and a static export of the web page is saved. Our prototype achieves an accuracy of about 90%.
We also took a closer look at the evaluation of different LLMs. For this purpose, we created a KIVEDU-specific test data set and evaluated both open source and proprietary LLMs on it. Here we were able to determine that the proprietary LLMs currently still achieve significantly better results than the open source LLMs. However, the proprietary LLMs are also significantly more expensive and offer fewer privacy guarantees. Therefore, in the future, we will try to improve the open source LLMs through finetuning or few-shot prompting to achieve better accuracy.
The KIVEDU project is currently in its infancy, but brings with it the potential for a fundamental reformation of consumer law enforcement in Germany and Europe. We look forward to continuing the collaboration with our partners and are excited to see what the future holds.