UOKiK, Poland - ARBUZ – AI-powered assistant detecting abusive contract clauses 

UOKiK, like other national consumer protection authorities in the EU, is obliged to ensure that contracts used in everyday consumer transactions such as financial services or online subscriptions do not contain provisions detrimental to the interests of consumers.

The usual way to investigate whether such terms exist has been through the legal teams at UOKiK investigating external complaints of clauses which potentially infringe consumers’ rights. In response, they carry out the time-consuming tasks of reading, analysing and assessing and so must read, analyse and assess standard contracts to identify abusive clauses.

To speed up and streamline this process and support staff in the initial review of contracts, the UOKiK held an open competition to create an artificial intelligence tool to perform the assessment task. The programme they selected is called ARBUZ - which in Polish means 'watermelon' and is a word similar to 'abusive'.

The ARBUZ system uses a class of AI solutions that come from deep learning, called ‘Transformer deep neural networks’. The system was trained on a valuable database that included a register of court judgements that officially recognised clauses as violating consumer interests. Since 2016, it has been the remit of UOKiK to recognise such clauses (with the business having the right to appeal to court), and so the database also included a large number of clauses that the UOKiK found to be abusive.

To add value to this information, the register was manually annotated by around 50 of UOKiK’s legal officers over the course of several months. They tagged the entries with the relevant industry, keywords and the passage from the judgement which stated why the clause was deemed to be unfair.

The input of experienced, legal staff gave ARBUZ the knowledge required to attempt automated analysis. The detailed annotation meant that when abusive clauses are identified, they are accompanied with a justification for the assessment given.

Employees dealing with the detection of prohibited clauses can now log in to ARBUZ and use it to support their daily tasks. As a first step in surveillance, it is equipped with a ‘crawler’ that allows a search of internet domains to assess whether they contain standard contracts.

ARBUZ can also be fed with contracts received along with consumer complaints or obtained from other sources, for example, submitted by the investigated company at the request of UOKiK.

Based on the forms uploaded to the system, ARBUZ reviews the contract selected by an employee of UOKIK. Using the slider, the sensitivity with which the analysis is to be performed can be set - e.g. 5, 10 or 20 illegal provisions are to be found. The system compares the contract with the clauses identified as abusive in the database.

Visually, the newly identified potentially unfair clause is highlighted, next to a second, more detailed screen showing how likely they are to be unfair, according to the system. This is shown by values in percentages based on a comparison of how similar they are to provisions formerly assessed as abusive. This is not just a simple comparison, ARBUZ uses intelligent algorithms to recognize the meaning of complex sentences written in legal language.

UOKiK sees ARBUZ as working at the level of an intelligent assistant - more than just a simple tool but not yet set up to make independent decisions.  UOKiK staff either accept or reject its proposal which is an important stage in ‘supervised learning’ as it learns to better refine results. Management oversight is critical and tools have been built-in to allow directors to check and if necessary override incorrect judgements by officers, which reduces the risk of the program going in the wrong direction.