Truth Nuke

Idle
Badge in corner of screen
Results (1/2)
Results continued (2/2)
settings page

Inspiration

With "high quality" artificially generated content more accessible than ever, internet slop has been able to disguise itself as reliable information on a variety of topics. From financial and legal advice to health and family guidance, using the convenience and anonymity of the internet to search for answers to life's questions has now come at a cost.

What it does

Truth Nuke is a browser extension that dissects web pages and looks for signs of misleading or artificially generated content. Publication dates, author reputation, the website's nature itself, citations, primary sources, secondary sources, as well as common algorithmic/probabilistic language structure used by most large language models are packed together with the page's metadata and analyzed by Anthropic's claude-sonnet-4-20250514 model. Parameters are passed to Claude telling it what to look for and how to read the metadata. Flagged information as well as an influence and credibility score are calculated and shown to the user to provide meaningful, non-invasive insight on content trustworthiness.

How we built it

Like all browser extensions, Javascript handles most of the logical processes and the UI is constructed with HTML. We started with identifying the key factors and elements in a webpage that provide tells when AI content is being utilized and moved towards including in-text citations and a bank of reputable scholastic websites to look for. Metadata on the site like ad frequency, publication dates, and registration information as well as domain names are all passed as variables to help calculate influence and credibility scores.

Challenges we ran into

Originally, we wanted to create our own machine learning model to feed webpage content and reinforce patterns of sketchy, untrustworthy information. However, due to time constraints and available resources, we decided to leverage existing AI models and instead feed parameters to steer judgement and help calculate the scores.

Accomplishments that we're proud of

We're proud of our clean and professional user interface designed with custom assets and an intuitive, lightweight, non-invasive experience.

What we learned

Machine learning model training takes longer than 24 hours.
Browser extensions have limited access to many features on modern web pages requiring workarounds when passing data.