P-SEAN: A Political sentiment analyzing language model

An online language model with internet access has become popular for its ability to search real-time data and interpret sentiments. Originally, an AI company developed the model for content creators on social media platforms. Unlike traditional data analysis tools, SEAN (Sentiment Analyzer), as a so-called webcrawler, can not only analyze the number of interactions to, say, a video or photo post, but also interpret the content of the comments (Deng et al. 2023 show that LLMs can analyze the sentiment in the financial market baked on Reddit posts. This is why it is conceivable that LLM’s can label the sentiment in posts about other topics as well as being positive or negative, etc.) This enables content creators to produce targeted content and respond to viewers’ wishes or criticism.
The German government asks the company to develop SEAN into P-SEAN. P-SEAN is to be fine-tuned with political content and opinions in order to be able to analyze political sentiments with temporal precision. For example, P-SEAN constantly analyzes posts on social media platforms that are tagged with certain hashtags (Törnebrg (2023) shows the possibility of this). In this way, radicalization in society can be quickly identified and prevented through appropriate measures.

Best case:

In the run-up to a national election, a team of political scientists uses P-SEAN to measure the current public sentiment. The precise analyses provide the parties with a clear picture of current political trends and enable them to address the electorate in a targeted manner. Especially during interviews or TV debates it is possible to analyze exactly how the voters evaluate the candidates. Because P-SEAN collects and analyzes data in real time, parties can respond quickly to changing sentiments and adjust their campaigns accordingly. Voters also benefit from accurate predictions, allowing them to make informed decisions and evaluate potential party coalitions. The elections are a success in terms of citizen participation and transparency.

Worst case:

Opponents of a small, not-yet-established party on the left end of the political spectrum use a second LLM to generate mass social media posts from fake accounts. However, since the texts appear fully human, they cannot be sufficiently deleted (Goldstein et al. (2023) and Kreps & Kriner (2023) tested that LLMs are able to write political statements convincingly, that are almost unrecognizable from human-written texts). The content of the posts is consistent with the small party’s platform. P-SEAN predicts an electoral success for the left-wing party. Since other voters fear the success of this fringe party and its extreme left-wing policies, they change their voting behavior in the opposite direction, supporting parties at the other end of the spectrum. Surprisingly, this leads to electoral success for right-wing parties, shifting the country’s politics toward a small group of voters.

What could be done better:

AI-generated texts need to be better filtered and deleted on social media networks. This could be done by another neural network (Chavan & Kane (2022) show that typical propaganda techniques can be accurately labeled by LLMs. This could then be transferred to model-written propaganda). In addition, it must be ensured that accounts can actually be traced back to real people. It is also necessary to verify the model’s predictions with traditional election analysis techniques. Finally, the model should be periodically checked for political neutrality to ensure that no training bias is introduced by the data.

References:

Chavan, T., & Kane, A. M. (2022, December). ChavanKane at WANLP 2022 Shared Task: Large Language Models for Multi-label Propaganda Detection. In Proceedings of the The Seventh Arabic Natural Language Processing Workshop (WANLP) (pp. 515-519).

Deng, X., Bashlovkina, V., Han, F., Baumgartner, S. & Bendersky, M. (2023). LLMs to the Moon? Reddit Market Sentiment Analysis with Large Language Models. Association for Computing Machinery. https://doi.org/10.1145/3543873.3587605
Goldstein, J. A., Chao, J., Grossman, S., Stamos, A., & Tomz, M. (2023). Can AI Write Persuasive Propaganda?
Kreps, S., & Kriner, D. L. (2023). The potential impact of emerging technologies on democratic representation: Evidence from a field experiment. New Media & Society, 0(0). https://doi.org/10.1177/14614448231160526
Törnberg, P. (2023). ChatGPT-4 outperforms experts and crowd workers in annotating political Twitter messages with Zero-Shot learning. arXiv (Cornell University). https://doi.org/10.48550/arxiv.2304.06588