PropaLens Launches In-Browser Machine Learning Extension to Detect Propaganda Techniques
An on-device Chrome extension flags manipulative rhetoric while preserving reader privacy completely.
1Client-Side Rhetoric Identification Tool
PropaLens is an experimental Chrome extension designed to highlight sentences matching fourteen recognized propaganda techniques across online articles. By analyzing rhetorical patterns directly within the browser rather than assessing factual accuracy, the system alerts readers to persuasive framing mechanisms such as loaded language or appeals to fear.
2Local Inference Preserves Privacy
The analysis executes entirely on small machine learning models running locally inside the user's browser. Because no document content leaves the device, analyzed web page text is never transmitted to external servers or third parties, eliminating data collection and tracking concerns inherent in remote cloud-based NLP APIs.
3Visual Categorization and Confidence Scores
Flagged sentences receive distinct color-coded underlines indicating the specific propaganda technique detected by the model. Users hovering over underlines access one-line definitions alongside low, medium, or high confidence ratings. An integrated popup menu summarizes technique counts and enables fast navigation across flagged excerpts.
4Precision-Recall Threshold Controls
An adjustable slider allows readers to customize the tradeoff between detection coverage and classification reliability. Higher strictness levels prioritize precision by reducing false alarms at the cost of missed instances, whereas lower thresholds surface broader sentence highlights with a higher likelihood of incorrect technique attributions.
5Corpus Evaluation and Validation
Developers evaluated three bundled models against 2,087 held-out test sentences derived from the public PTC corpus. Span-level annotations from the original dataset were mapped to sentence-level units. Benchmark results confirm higher strictness yields higher true-positive accuracy, though performance variations naturally occur on arbitrary web texts.
6Public Training Dataset Foundations
The classifiers are fine-tuned from two publicly available datasets: the SemEval-2020 Task 11 PTC corpus and the PropXplain collection from EMNLP 2025 findings. Modifications adapted these sources for local inference, operating independently of the original academic researchers under CC BY 4.0 and MIT licensing terms.
7Optional Cloud Explanation Tier
While the on-device extension remains free, the team is evaluating an optional hosted service tentatively priced between $1 and $3 monthly. This cloud tier would leverage larger models to provide contextual explanations detailing why specific sentences were categorized under particular propaganda rhetorical techniques.
8Data Governance and Infrastructure
The extension enforces zero telemetry, storing user configuration preferences entirely in local browser storage without analytics scripts or advertising trackers. Waitlist emails for the cloud service are kept in isolated lists, while infrastructure relies on standard server logs for secure site delivery.