Public opinion on EU regulations on artificial intelligence - An unobtrusive NLP approach
Tue-01
Presented by: Veronika Batzdorfer
Background. Consulting public stakeholders like citizens, research, or business institutions regarding their attitudes towards AI technologies is important, in order to tailor policy regulations and mitigate risks. Particularly, considerations on the ethics of AI have gained traction in recent years, as is instantiated by the risk for biases in large language models or profiling for automated decision making. Yet, unobtrusive measures of attitudes that consider multiple stakeholders, multiple stages of a decision-making processes and cross-country perspectives are rarely accessible.
Objectives. The main goal of this study is to compare stances (i.e., relating to perceived chances, risks, and regulations) towards AI on an EU-level for multiple stakeholders, across countries, based on text data.
Research questions. The main research questions investigate whether differences in polarity towards AI and topics discussed across stakeholders exist and whether these can be explained by organization-level and country-level background variables.
Method/Approach. Public data on three EU consultation rounds (06/2020-08/2021) on AI regulations have been obtained by means of dynamic web crawling of PDF opinion pieces and website metadata. Furthermore, multi-lingual text has been detected and translated into English, before obtaining polarity measures with different sentiment analysis techniques (e.g., vader and syuzhet). Lastly, we used BERTopic to combine transformer models with topic models to identify and compare clusters of topics over time and across stakeholders.
Results. On a country-level particularly Belgium have most frequently contributed, among them particularly non-governmental institutions. Research organization predominantly showed a negative stance toward AI and a focus on risks and negative consequences relating to discrimination and misuse.
Conclusions and implications. Gaining knowledge on risk perception and regulation needs of public stakeholders can be fruitful to predict future initiatives and cast a view on policy making efforts across countries. Although the data are not a probability-based sample, leveraging new sources of open data, whilst acknowledging potential biases can enrich other public opinion measurements.
Funding. This research was funded by ITAS/ KIT
Objectives. The main goal of this study is to compare stances (i.e., relating to perceived chances, risks, and regulations) towards AI on an EU-level for multiple stakeholders, across countries, based on text data.
Research questions. The main research questions investigate whether differences in polarity towards AI and topics discussed across stakeholders exist and whether these can be explained by organization-level and country-level background variables.
Method/Approach. Public data on three EU consultation rounds (06/2020-08/2021) on AI regulations have been obtained by means of dynamic web crawling of PDF opinion pieces and website metadata. Furthermore, multi-lingual text has been detected and translated into English, before obtaining polarity measures with different sentiment analysis techniques (e.g., vader and syuzhet). Lastly, we used BERTopic to combine transformer models with topic models to identify and compare clusters of topics over time and across stakeholders.
Results. On a country-level particularly Belgium have most frequently contributed, among them particularly non-governmental institutions. Research organization predominantly showed a negative stance toward AI and a focus on risks and negative consequences relating to discrimination and misuse.
Conclusions and implications. Gaining knowledge on risk perception and regulation needs of public stakeholders can be fruitful to predict future initiatives and cast a view on policy making efforts across countries. Although the data are not a probability-based sample, leveraging new sources of open data, whilst acknowledging potential biases can enrich other public opinion measurements.
Funding. This research was funded by ITAS/ KIT