The paper introduces an AI-based virtual assistant VILEM whose primary aim is to strengthen individual right to informational self-determination on the Internet. VILEM helps users to manage their privacy settings, protect themselves against potentially abusive websites, and saves time of users as it presents relevant information on personal data processing in a comprehensible manner. The paper also presents how VILEM ful lls requirements on Trustworthy AI.
E cient privacy protection is one of the crucial values that we need to
foster in our information based society. At the same time, preserving this value
should not hinder the development of society, science, technology, and provided
services. Unfortunately, processing personal data when accessing various online
services, such as social media, can pose various risks for Internet users { namely
undermining their ability to exercise control over own personal data [
The necessity and importance of processing personal data, however, raises as
new services and applications are being developed. One of currently very popular
trends is personalization. Personalization can be understood \as a process that
changes the functionality, interface, information content, or distinctiveness of a
system to increase its personal relevance to an individual" [
Personalization is typically based on pro ling which can be technically done
namely with help of various kinds of cookies ( rst-party cookies, third-party
cookies, Flash cookies, etc.) or with help of other means, such as IP addresses,
web bugs, URL form data, etc. [
This practice causes problems to Internet users who are often not aware
about placement of cookies or are perplexed with a large number of requests on
granting consent with cookies, with long and incomprehensible privacy policies,
multiplicity of actors, and overall information asymmetry they are facing. In
2016, Eurostat described how Internet users protected their privacy online.
Citizens of the Netherlands, Germany or Finland were very much aware of the fact
that they can be traced by cookies. Despite that Internet users were not very
active with regard to changing their cookie settings in a web browser. In
particular, in the Czech Republic less than 20 % of people changed their \browser
settings to prevent or limit cookies use" [
As individual autonomy is threatened in the online environment where there
are many \little brothers" and individual choices are predetermined based on
ubiquitous tracking and personalization [
This virtual assistant contributes to developing socially responsible AI.
Although socially responsible approach to AI has been mentioned by research in the
past [
Therefore, the aim of this paper is to introduce how we intend to implement and operationalize a socially responsible AI system that would assist Internet users with e cient protection of their online privacy and the right to their informational self-determination by providing them with an easy and freely available tool allowing them to administer their privacy preferences, reduce information asymmetry, inform them in a comprehensible manner, and educate them in the area of law and technology. 2
The tool we propose { our virtual assistant { will be initially available for users from the Czech Republic. Therefore, its operation must be based on and compliant with EU and Czech laws related to personal data protection and cookies.
This legislation is quite robust and additional various explanatory documents
such as opinions of the European Data Protection Board or other bodies need
to be taken into account [19{23, 38]. Protection of personal data on the Internet
is regulated namely by the General Data Protection Regulation (hereinafter
GDPR [
The complexity of the situation increases also when cookies are used. Cookies
play an important role in securing proper functioning of providing online content
and online services. EU law recognizes them as legitimate tools, for instance, for
\analysing the e ectiveness of website design and advertising, and in verifying
the identity of users engaged in on-line transactions" [
Originally, cookies were considered a privacy preserving mechanism [
One of the solutions for achieving legal compliance regarding cookies is use of consent management platforms. As preparatory work for designing an AI-based virtual assistant, we needed to verify the level of use of consent management platforms in the Czech Republic. In the preparatory phase, we crawled a number of websites and tried to automatically identify how many of them are using standard cookie consent managers (like CookieBot or OneTrust) because it is much easier to automatically analyze consents on pages using such managers. As the detection was done with help of a rule-based system (using de ned HTML structure and keywords), the results are only approximate. It is possible that a few pages were using consent managers contrary to the results or that on some pages the managers were not identi ed correctly. However, based on manual evaluation of some pages we can say it happened just in several cases. The results of this experiment are shown in Table 1. The results indicate that 314 pages out of 3649 used one of the tested consent management platforms and 809 pages probably do not mention anything about cookies at all. Description Number OneTrust (CMP) 142 CookieBot (CMP) 16 Cookie Consent (CMP) 112 Funding Choices (CMP) 44 Cookies mentioned in Privacy Policy 2526 Probably no cookies used 809 Pages scanned in total 3649
The experiment shows that use of cookies on the Internet in the Czech Republic is quite inconsistent and, therefore, can be also confusing for Internet users. 3 3.1
As suggested above, problems with online privacy protection and bad cookie practice described above led us to the idea that Internet users need to be much better equipped in order to face the growing information asymmetry, overload of information related to personal data protection on the Internet, and a growing number of requests on providing their consent with personal data processing and placing cookies into their devices. As the principle of granting consent is in line with the legal principle of personal autonomy and the right to information selfdetermination, it needs to be maintained. At the same time, Internet users need to be provided with tools on how to exercise their will and rights in practice not to be paralyzed by practical e ects of this legal requirement.
Therefore, we have designed an AI-based virtual assistant VILEM. Its name refers to the principle of autonomy as the etymological meaning of this word is \my will is my protection." Moreover, VILEM is an acronym that stands for Volition Inspirited by Legal EMpowerment.
The idea of using technology to empower Internet users with regard to privacy
protection online is not new. Apart from various plug-ins that help to block
tracking by third-party cookies, there is, for instance, a solution that utilizes deep
learning and helps Internet users to comprehend privacy policies { Polisis [
The uniqueness of VILEM lies in its ability to return its users the decisionmaking capacity that would not be hindered by the necessity to exploit limited personal resources, such as time to search for relevant information, time to manually set up privacy preferences for each visited website, and biologically limited attention span. The following subchapter describes how VILEM shall function in practice. 3.2
Form and appearance VILEM is designed in the form of a sidebar that appears when an Internet browser is opened. VILEM updates itself automatically once a user enters a website on a new domain or when new cookies are detected. It is accessible all the time and not only on demand. Currently, VILEM is designed only in the Czech language and for Czech users. When completed, VILEM will be available for free as a web plug-in.
Personalized privacy protection We presume that upon installing VILEM,
users will ll in a survey regarding their privacy preferences. Our pilot empirical
study that we conducted in the Czech Republic in December 2020 shows that
some users trust certain websites more and are willing to share more
information with them than with others. Moreover, 56 % of respondents stated that
they prefer to assess each purpose of processing personal data individually [
Providing information to users VILEM informs users about the information that the data controller needs to make available according to the GDPR. This information contains the name and contact details of the controller, the purpose of data processing, categories of processed personal data, and legal grounds for the processing (consent, contractual obligation, legal obligation, protection of vital interests, protection of a public interest, or a legitimate interest of the controller). Additionally, VILEM informs users whether the controller makes the data available to other parties, such as processors. In that case VILEM also provides users with respective contact details. Users should understand from the information provided that their consent is reversible, if already given, or optional, if not. The same goes for the stated legitimate purposes.
VILEM will also ful ll an educational role. It will inform users in simple terms about what provided information means and what can be done in each situation. For those interested in the topic and who will want to go into it in more depth, we will provide links to our website with educational videos (in preparation).
Tool for managing cookies VILEM will enable users to forbid cookies unnecessary for website's functioning and inform them that a website is inaccessible without agreeing to certain types of cookies. In the future, VILEM should provide additional functions, such as enabling users to forbid pop-up noti cations and informing them if a website contains paid promotions. 3.3
Various methods can implement mapping of user preferences with speci c rules o ered by websites providers. With growing computational capabilities and more powerful hardware, automatic but rigorous analysis of textual data becomes more imaginable than before. VILEM can use any modern sequence classi cation approach for checking matches in privacy preferences. Most probably, we will use some BERT like models[bert] in cross attention fashion.
The current trend in natural language processing is to use large neural
network models pre-trained on huge data sets. Such data do not have to be manually
labelled; we can use automatically generated datasets and design arti cial tasks {
so-called self-supervised learning { to extract knowledge about human language
and wisdom about the world. These models are mainly intended for the English
language. However, researchers released models trained on a couple of languages
simultaneously [
In exponentially growing data production and a rapidly changing legal environment, modern society often searches for an easy and systematic way of solving law issues. Complex but easy-to-use shelf-product solutions are often the rst-choice tool for website operators who want to satisfy complicated law. Projects like OneTrust aim to deal with the changing environment. However, a non-negligible amount of web service providers still does not a systematic solution or do not even implement law obligations. In this regard, VILEM will be able to easily analyse and solve users matches on sites with mainstream systematic solutions due to the known structure of forms and will be able to focus only on textual content and its semantics. In non-systematic solutions, there will be one extra step { to identify the form and parse the statements with their controls. In the next step, VILEM will be able to analyse the texts, highlight match or mismatch in each statement and pre ll the form for users. It will be also able to recognise and send information about potential violation of law to the respective public authority.
As VILEM will use techniques of natural language processing, we need to
take extra care when preparing datasets for its learning. The preparation of the
data is done manually by people with knowledge of personal data processing
and its legal limits as well as the obligations imposed by the data protection
legislation [
There are some speci cities related to annotating in the Czech language and within the Czech legal culture. The rst speci city lies in the workings of the language itself. The structure di ers signi cantly from English, German and other languages used in areas where annotation of legal texts is more common than in the Czech Republic. Therefore, existing conventions from other countries are not usable for our work. The second speci city concerns the legal culture in the Czech Republic, which scarcely uses standardized templates. The governmental bodies, such as ministries and specialized public authorities (e.g., the O ce for Personal Data Protection), do not provide them either. It is customary for the administrative bodies to provide guidelines on the creation of necessary documents instead. As a result, when annotating a Czech legal text one must expect a high level of variability in its structure as well as in the terminology.
The eld of annotating Czech legal texts is not explored thoroughly.
Nevertheless, more than one research project on this matter was completed in previous
years. For data extraction, we follow the best practice set by the team on the
Faculty of Law at Masaryk University [
The challenges of texts containing terms and conditions for personal data
processing are 1) variations in used legal terminology; 2) creativity in the phrasing
of information obligations towards the user; 3) inconsistent level of detail of
provided information; 4) purposeful omission of certain information. The obstacles
we are facing di er from the ones the team of Masaryk University had to solve.
For example, one of the problems they had was an unclear structure of case-law
decisions that required the addition of functionality to their tool, enabling
automated segmentation of the text [
To prevent mistakes in the manual annotation that would have a negative in uence on the functionalities of VILEM, the annotation is done by professionals in the area of personal data protection. This way, they are familiar with the terminology and it will be quicker for them to search for relevant data in the texts. Furthermore, their practical experience shall ensure that they can identify possible hidden information in the text. Such as unlawful limitations of the data subject's rights. These are written intentionally in a way that would confuse an an unprepared reader without professional expertise. In the future, we would like to enable the users of VILEM to send feedback in case they encounter an error made by VILEM, such as the inability to nd all the information on personal data processing in the terms and conditions. This way, VILEM can improve, which is one of the upsides of using AI algorithms. That said, it is necessary to supervise the learning of VILEM on the data feedback, since it might be incorrect. It is bound to happen that the users will not be able to correctly identify the relevant legal meaning behind the phrasing of the text. It can be quite confusing for a consumer. Even so, the texts can be confusing to legal professionals as well.
Therefore, we have implemented the practice of multiple annotations of the
same text. To make the manual annotations as precise as possible, we also created
an annotation manual. The annotation manual is inspired by the one used for the
annotation of Czech judicial decisions [
Any user experience feedback is precious and can improve functioning of a system by adjusting the user interface or enlarging the training corpora. However, in every application, collecting user feedback can be tricky. Active feedback can be time-consuming and annoying for users. Moreover, recording users' behaviour may not be well accepted in a project dealing with privacy issues. However, VILEM can overcome these issues. We will let the users consider the bene t of making the system better and let them decise whether they would accept or deny sending anonymous automatic feedback upon installation of VILEM. meanwhile, we will place feedback buttons in the applications for those who want to give active feedback.
We need to keep in mind that some serious problems can arise if users would have the possibility to a ect the decision process of VILEM by sending feedback on incorrect annotation of legal texts. In the rst place, a usual user is not a lawyer, so that their reasoning can be simply wrong. Fortunately, modern models can handle non-systematic noise brought by users well. However, if there was some systematic misunderstanding of the law by the users, the model could drift to this potentially wrong interpretation. We will avoid this unwanted state by searching for a systematic deviation and investigating such singularities by professionals.
Nonetheless, the fact that the user would not be satis ed with the outcome of VILEM will be essential for us. We can solve this issue in several ways: by providing us with information why VILEM marked a statement as it had done in the rst place and let the user share his own opinion on the subject for further processing. If VILEM was wrong, we would add this "outlier" to the training corpora. If VILEM was right, it could be a sign of an unclear understanding of the setting of users preferences in VILEM after installation. 3.5
In November and December 2020, we tested the rst proof of concept and the
user interface of VILEM. In December 2020, we conducted the rst pilot
empirical study and presented a mock-up version of VILEM to 50 Czech Internet
users [
The feedback from respondents was very positive. 92 % of respondents
considered such web plug-in as desirable. 94 % of respondents evaluated VILEM as
useful and 84 % considered it trustworthy. 56 % of respondents expressed that
VILEM would strengthen their control over information and would help them
to make the process of privacy protection more comprehensible to them.
Respondents expressed their expectations that VILEM would help them to protect
themselves from and block harmful websites as well as save their time. Only 4
respondents out of 50 were hesitant or negative about the use of VILEM. The main
reasons were a general concern that VILEM would slow down a computer and a
general distrust to any solution that needed to be installed into a computer [
The pilot study showed us that Czech users would welcome a technical solution that would strengthen their control over personal data, warned and protected them against threats (namely when a websites requires more information than users are willing to provide), and instructed them what to do in certain situations.
The main lesson we took from the study is to design VILEM in such a manner that it shall provide a maximum level of information, choice and control to users an a very comprehensible and easy-to-understand manner. In order to strengthen trustworthiness of VILEM, we need to diligently implement and operationalize requirements on Trustworthy and Responsible AI. The following chapter will illustrate how we plan to do so.
The term of Trustworthy AI was introduced by the High-Level Expert Group on
Arti cial Intelligence [
four ethical principles on Trustworthy AI { respect for human autonomy, prevention of harm, fairness, and explicability. VILEM ful lls the rationale of all of the four principles. Its aim is to strengthen human autonomy by providing Internet users with a free tool for better management of own choices and hereby prevents harm that they could face by careless sharing of personal data. VILEM will not discriminate any user as it will be freely available to all Czech citizens. Moreover, functioning of VILEM will be explained to users in a comprehensible and transparent manner.
with seven requirements. Compliance of VILEM is described below for each of the requirements.
Human agency and oversight As mentioned above, the main function of VILEM is to strengthen personal autonomy. VILEM supports individual decision-making related to protecting own online privacy.
Technical robustness and safety The main risk to technical robustness and safety would come from the side of users who could in uence functioning of VILEM. However, all input and feedback from users on problems related to malfunctioning will be checked manually.
Privacy and data governance VILEM will not collect or process personal data related to its users. All activity will be done only on the side of users. It will be possible to share data with us if a particular user will wish to do so for the purpose of improving the system.
Transparency VILEM shall be completely transparent. We intend to share the code as open source. Moreover, users or any other person will be provided with information which training data was used.
Diversity, non-discrimination, and fairness VILEM is designed as user-centric and will be made available to anyone for free. Societal and environmental well-being By protecting users VILEM will contribute to overall societal well-being.
Accountability With regard to securing safety and robustness, functioning of VILEM will be continuously monitored and improved. 5
Our virtual assistant VILEM is in the process of development. However, even
after it will be nished, we will need to continuously update it, expand the
training corpora and keep analyzing how law and privacy policies as well as
cookie legislation and practice evolves. One of the upcoming challenges we will
need to react to is a change in use of so called third-party cookies [