VIGIL: An Extensible System for Real-Time Detection and Mitigation of Cognitive Bias Triggers

This paper introduces VIGIL, the first open-source browser extension designed to provide real-time detection and mitigation of cognitive bias triggers in online information through features like in-situ analysis, reversible LLM-powered reformulation, and privacy-flexible inference.

The Gist

Imagine your brain is a busy highway. Most of the time, you drive carefully, checking your mirrors and reading signs (this is your slow, logical thinking). But sometimes, the road is designed with tricky curves, flashing neon signs, and sudden bumps that make you swerve without even realizing it (this is your fast, instinctive thinking).

In the digital world, bad actors and manipulative content use these "tricky curves" to trick you. They use specific words, repetition, or emotional language to hijack your instincts. These are called Cognitive Bias Triggers.

Until now, tools to help you online have mostly been like traffic signs that tell you who built the road (e.g., "This news site is political") or what is written on the sign (e.g., "This claim is false"). But they haven't told you if the road itself is designed to make you crash.

Enter "Vigil."

Think of Vigil as a smart co-pilot for your web browser that doesn't just watch the road; it watches how the road is built to trick your brain.

How Vigil Works (The Simple Version)

1. The "Spotter" (Real-Time Detection)
Imagine you are scrolling through Twitter or reading news. As you scroll, Vigil acts like a high-tech spotlight. It instantly scans the text you are looking at. If it spots a "tricky curve"—like a word designed to make you angry instantly or a sentence repeated to make it sound true—it highlights that specific phrase.

• The Analogy: It's like a spellchecker, but instead of fixing typos, it highlights "brain traps."

2. The "Translator" (Mitigation)
Once Vigil spots a trap, it doesn't just yell "Danger!" It offers a magic eraser. You can click a button to "Rewrite" the text. Vigil uses advanced AI to take that manipulative, emotional sentence and rewrite it into a neutral, factual version.

• The Analogy: If a news headline screams, "SHOCKING DISASTER!" (designed to panic you), Vigil rewrites it to: "A significant event occurred." It keeps the meaning but removes the emotional hook.
• The Safety Net: You can always click "Restore" to see the original text again. It's like having a "Undo" button for reality.

3. The "Privacy Shield" (Security)
Some tools send your reading habits to a giant server in the cloud, which feels like a spy watching you. Vigil is built differently. It offers four levels of privacy:

• Level 1 (The Fortress): It can run entirely inside your computer using your own graphics card. No data leaves your machine. It's like having a private detective who lives in your house and never calls anyone.
• Level 2-4: If you need more powerful analysis, it can use local software or cloud services, but you choose how much you trust them.

4. The "Lego System" (Extensibility)
Vigil is built like a Lego set. Developers can snap on new "plugins" (little tools) to detect new types of tricks. If someone invents a new way to manipulate people, a developer can build a new plugin to catch it, and Vigil will learn to spot it immediately.

Why This Matters

• Current Tools: Tell you who is talking (e.g., "This is a conservative site") or what they are saying (e.g., "This fact is wrong").
• Vigil: Tells you how they are saying it to trick your brain (e.g., "They are using scary words to make you skip thinking").

The Bottom Line

Vigil is the first tool that acts as a real-time bodyguard for your attention. It helps you slow down, spot the emotional tricks being played on you, and gives you the power to strip away the manipulation so you can see the truth clearly. It's not just about fact-checking; it's about mind-checking.

Technical Summary

1. Problem Statement

The rise of generative AI has exacerbated risks to online information integrity, primarily through mis- and disinformation. While existing media-literacy tools address ideology (e.g., NewsGuard, which rates source political leaning) and factuality (e.g., ClaimBuster, which scores claim validity), they fail to address a subtler but equally corrosive threat: rhetorical manipulation.

Current tools do not detect how information is presented to exploit human cognitive limitations. Drawing on Dual-Process Theory, the authors argue that online content routinely exploits "System 1" (fast, heuristic-driven) cognition through cognitive bias triggers (e.g., loaded language triggering the affect heuristic, or repetition triggering the illusory truth effect). There is a lack of tools capable of detecting these triggers in real-time within the user's browsing experience and offering immediate mitigation.

2. Methodology and System Architecture

Vigil is a Chromium-based browser extension designed to detect and mitigate cognitive bias triggers in-situ. It features a modular, extensible architecture (Fig. 1) comprising four main components communicating via typed messages:

• Content Layer: Handles text extraction, viewport tracking, and rendering. It enables "scroll-synced" analysis, meaning the system automatically tracks the visible portion of the webpage without requiring manual text selection.
• Message Router: Coordinates data flow between components.
• Sidepanel: Provides the user interface for "Analyze" and "Settings," displaying detection results and mitigation options.
• Inference Runtime: Executes the detection models. It supports both in-browser execution (via WebGPU/WebLLM) and server-side execution.

Plugin System

Detection logic is encapsulated in plugins that adhere to a shared Finding contract (defining trigger type, severity, text span, and explanation).

• Browser Plugins:
  • cbt-regex: Uses pattern matching against a 14-type cognitive bias taxonomy. It operates with zero latency and requires no network connection.
  • cbt-llm: Uses an in-browser LLM (Llama 3.2 1B via WebLLM) to detect triggers by mapping SemEval-2020 propaganda techniques to specific cognitive biases.
• Server Plugins:
  • moralization-llm: Detects moralization based on Moral Foundations Theory, supporting English and German.

Privacy-Tiered Inference

Vigil offers four tiers of inference to balance privacy and capability:

1. In-browser Regex: Zero-network, instant screening.
2. In-browser WebGPU LLM: Zero-network, deeper semantic analysis.
3. Local Ollama API: Runs on the user's machine.
4. Cloud Endpoint: Uses remote APIs (e.g., OpenAI-compatible).
   Note: Tiers 1 and 2 are "verifiably zero-network," ensuring no data leaves the user's machine. No data is persisted server-side.

3. Key Contributions

1. First Real-Time Browser Extension for Cognitive Bias: Vigil is the first tool to provide real-time, in-situ detection and mitigation of cognitive bias triggers directly within the browsing experience.
2. Mitigation via Reformulation: Unlike passive detection tools, Vigil allows users to mitigate bias. Users can click "Rewrite" to trigger an LLM to reformulate the biased text into a neutral, semantically equivalent version. This process is fully reversible (users can restore the original text).
3. Extensibility: The system is designed with a plugin architecture, allowing third-party developers to add new trigger types easily.
4. Privacy-First Design: The architecture supports fully offline operation, addressing privacy concerns inherent in AI-driven content analysis.

4. Results and Evaluation

The authors evaluated Vigil using state-of-the-art benchmarks and measured performance in terms of accuracy and latency.

• Detection Performance:
  • Propaganda Detection: On the SemEval-2020 Task 11 benchmark, the production prompt achieved a micro-F1 of 0.533 (Precision = 0.626), deliberately prioritizing precision to reduce false positives.
  • Moralization Detection: On the Moralization Corpus, the plugin achieved a macro-F1 of 0.789, outperforming the corpus authors' best result (0.772).
• Latency:
  • Regex: 0.03 ms (instant).
  • WebGPU (In-browser LLM): 3.4 seconds.
  • Cloud API: 3.9 seconds.
  • Optimization: Caching ensures that repeated views of the same content are instantaneous.
• User Experience: The system successfully integrates with Twitter/X feeds and news websites, providing visual highlights, tooltips, and a scroll-synced "Currently Viewing" card.

5. Significance and Discussion

Vigil represents a paradigm shift in media literacy tools. While previous systems focused on who said something (ideology) or what was said (factuality), Vigil focuses on how it was said (cognitive framing).

• Comparative Advantage: As shown in the paper's comparison table (Table 1), Vigil is the only system that combines In-situ delivery, Real-time detection, and Active Mitigation at the Span (text segment) level.
• Responsible AI: The system is open-source and grounded in established academic taxonomies. It adheres to "privacy-by-design" principles, collecting no telemetry.
• Limitations: The authors acknowledge potential LLM false positives (mitigated by a human-in-the-loop design where users can restore original text) and the reliance on proxy benchmarks.

In conclusion, Vigil provides a critical infrastructure for empowering users to recognize and neutralize manipulative rhetoric in real-time, fostering more rational public discourse in the age of generative AI.