TruthLens

Inspiration

The inspiration for TruthLens emerged from a growing sense of urgency and frustration. Every day, we witness how the boundary between real and synthetic online content continues to blur. Whether it’s convincingly fake AI-generated news articles, deepfake videos of public figures, or floods of manipulated product reviews influencing consumer decisions digital trust is eroding rapidly.

We were frustrated by the helplessness that many people (including ourselves) feel when consuming information online. It’s an asymmetric battle: the creation of disinformation has become exponentially easier, faster, and cheaper with AI, while verification remains slow, manual, and tedious.

We no longer wanted to stand by as misinformation influences elections, enables e-commerce fraud, and poisons public discourse. The inspiration was the desire to create a tool that not only warns but also educates a "reality filter" that returns control and critical thinking to users, right where they consume information: in the browser.

What it does

TruthLens acts as an intelligent co-pilot for browsing the web. As a lightweight browser extension (for Chrome, Firefox, and others), it continuously analyzes the web page a user is currently visiting — in the background.

It operates on three levels:

1. Text Analysis:
   Scans text sections in real time to detect subtle stylistic patterns typical of AI text generators (LLMs). It highlights passages that are likely not human-written or may contain “hallucinated” facts.

2. Image & Video Analysis:
   Checks embedded media for signs of manipulation, deepfake artifacts, or inconsistencies in metadata (provenance).

3. Source Analysis:
   Assesses the credibility of the domain itself, based on its history, network, and verified fact-check databases.

Instead of blocking the page, TruthLens displays a small, color-coded badge (e.g., green, yellow, red) next to the browser’s address bar. A single click opens a simple panel explaining why certain content was flagged (e.g., “This paragraph shows features typical of AI-generated text” or “This source has been frequently corrected by fact-checkers”).
It also provides evidence and links to alternative, trustworthy sources, allowing users to make informed decisions.

How we built it

To bring TruthLens to life, we follow a hybrid “privacy-first” approach that combines on-device processing with powerful cloud-based AI analysis.

Frontend (Browser Extension)

Built with JavaScript/TypeScript, the extension runs lightweight models (e.g., distilled Transformer models) and fast heuristics directly on the user’s device (“edge-first”).
This enables real-time text analysis while ensuring that sensitive browser data (like your history) never leaves the device.

Backend (Cloud Infrastructure)

For compute-intensive tasks that cannot run locally (e.g., deepfake video analysis or large-scale fact-checking), the extension sends targeted, anonymized requests to our secure servers.

The AI Modules

• NLP Module:
  A fine-tuned NLP model (e.g., from the RoBERTa or DeBERTa family), trained to recognize stylistic “fingerprints” of generative AI models.

• Vision Module:
  Combines CLIP-based models for image similarity and provenance checking with specialized detectors for visual artifacts (e.g., deepfakes).

• Fact & Source Module:
  A retrieval pipeline that normalizes claims found on a page and cross-references them with curated databases (e.g., Correctiv, Politifact, Snopes) and a graph of trusted news outlets.

Challenges we ran into

Developing TruthLens brought several complex challenges:

1. The AI Arms Race:
   The biggest challenge is the constant cat-and-mouse game. As soon as a detector works well, a new generation of AI models (e.g., GPT-5, Claude 4) emerges, designed to evade detection. Our models must be continuously retrained and adapted.

2. Avoiding False Positives:
   User trust is our most valuable asset. It’s crucial not to falsely flag legitimate content (e.g., creative writing that sounds like AI or authentic journalism). A single mistake here undermines our entire mission.

3. Performance & Scalability:
   Analysis must occur in milliseconds to avoid slowing down browsing. Finding the right balance between depth and speed deep analysis versus a seamless user experience is technically demanding.

4. Explainability:
   A raw AI score (e.g., “75% AI-generated”) is meaningless to users. Translating probabilistic outputs into clear, human-readable explanations (“Why was this flagged?”) is a key UI/UX design challenge.

Built With

• javascript
• mongodb
• python