Algoritmit vastaan propaganda - Koneoppimisen ja tekoälyn monimodaalinen kehityskaari propagandan tunnistuksessa 2011–2024
Roth, Joel (2025)
2025
Politiikan tutkimuksen maisteriohjelma - Master's Programme in Politics
Johtamisen ja talouden tiedekunta - Faculty of Management and Business
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2025-06-02
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202505296346
https://urn.fi/URN:NBN:fi:tuni-202505296346
Tiivistelmä
This thesis systematically investigates how artificial-intelligence and machine-learning methods were harnessed between 2011 and 2024 to automate the detection of political propaganda in the digital public sphere. Its scholarly contribution is threefold.
First, it integrates the rhetorical techniques of propaganda, Bayesian receiver theory and network-based anomaly detection into a unified conceptual–operational framework that accounts for the manipulative features of message content, their impact as Bayesian belief updates, and the irregularities of co-ordinated dissemination.
Secondly, the study is the first to chart long-term trends across seven methodological and ethical dimensions, revealing a strategic imbalance in the field: although deep-learning text models have advanced rapidly, causal impact evaluation, temporal robustness, multilingual coverage and ethical-bias auditing remain uncommon.
Thirdly, the thesis introduces a novel three-stage mixed-methods framework—Exploratory Sequential Mixed-Trend-Tracing (ES-MTT)—in which a bibliometric mapping of 2,916 publications, k-means clustering and a dampened citation metric yield a 64-article core sample. Relevance screening was conducted via a multi-stage pipeline. First, a multi-agent LLM consensus algorithm—comprising three independent GPT-based models—issued binary inclusion/exclusion judgements against preregistered criteria, returning them in a fixed JSON schema. In collaboration with a human reviewer this step eliminated 878 irrelevant records. The remaining documents were encoded with the text-embedding-3-large model, projected with LocalMAP, and stratified by HDBSCAN density clustering, yielding a semantically comprehensive, domain-representative, deterministically replicable and statistically validated core set of 64 articles.
The findings identify three distinct developmental phases. During the early period (2011–2015) research relied chiefly on traditional text classifiers and hand-crafted features; the middle period (2016–2019) mainstreamed benchmark datasets (LIAR, FakeNewsNet) and deep CNN/RNN architectures; in the new period (2020–2024) transformer architectures, multimodal vision–language models and heterogeneous graph neural networks combined text, image, video and network dynamics within a single classification pipeline. Nevertheless, only 14 per cent of studies tested models for temporal durability, 13 per cent operated genuinely across languages and 8 per cent reported ethical-bias metrics—despite the accountability requirements emphasised by EU AI regulation.
The study recommends: (i) expanding open, multilingual and multimodal benchmark datasets to encompass European and Global South languages; (ii) making standard causal and robustness tests a prerequisite for algorithmic deployment; and (iii) mandating bias audits and Model Card reporting to ensure transparency in platform content moderation. In so doing, the research supports policy measures aimed at democratic resilience and the responsible governance of algorithmic agenda-setting.
First, it integrates the rhetorical techniques of propaganda, Bayesian receiver theory and network-based anomaly detection into a unified conceptual–operational framework that accounts for the manipulative features of message content, their impact as Bayesian belief updates, and the irregularities of co-ordinated dissemination.
Secondly, the study is the first to chart long-term trends across seven methodological and ethical dimensions, revealing a strategic imbalance in the field: although deep-learning text models have advanced rapidly, causal impact evaluation, temporal robustness, multilingual coverage and ethical-bias auditing remain uncommon.
Thirdly, the thesis introduces a novel three-stage mixed-methods framework—Exploratory Sequential Mixed-Trend-Tracing (ES-MTT)—in which a bibliometric mapping of 2,916 publications, k-means clustering and a dampened citation metric yield a 64-article core sample. Relevance screening was conducted via a multi-stage pipeline. First, a multi-agent LLM consensus algorithm—comprising three independent GPT-based models—issued binary inclusion/exclusion judgements against preregistered criteria, returning them in a fixed JSON schema. In collaboration with a human reviewer this step eliminated 878 irrelevant records. The remaining documents were encoded with the text-embedding-3-large model, projected with LocalMAP, and stratified by HDBSCAN density clustering, yielding a semantically comprehensive, domain-representative, deterministically replicable and statistically validated core set of 64 articles.
The findings identify three distinct developmental phases. During the early period (2011–2015) research relied chiefly on traditional text classifiers and hand-crafted features; the middle period (2016–2019) mainstreamed benchmark datasets (LIAR, FakeNewsNet) and deep CNN/RNN architectures; in the new period (2020–2024) transformer architectures, multimodal vision–language models and heterogeneous graph neural networks combined text, image, video and network dynamics within a single classification pipeline. Nevertheless, only 14 per cent of studies tested models for temporal durability, 13 per cent operated genuinely across languages and 8 per cent reported ethical-bias metrics—despite the accountability requirements emphasised by EU AI regulation.
The study recommends: (i) expanding open, multilingual and multimodal benchmark datasets to encompass European and Global South languages; (ii) making standard causal and robustness tests a prerequisite for algorithmic deployment; and (iii) mandating bias audits and Model Card reporting to ensure transparency in platform content moderation. In so doing, the research supports policy measures aimed at democratic resilience and the responsible governance of algorithmic agenda-setting.