Truthy

Inspiration

With increasingly powerful LLMs comes increasing potential for misinformation: hostile organisations, state actors, and trolls can spread misinformation more credibly than ever. After all, 28 countries declared that frontier systems may amplify risks such as disinformation at the Global AI Safety Summit this week.

The best way to fight misinformation generated by LLMs is to leverage LLMs & AI systems to spot and highlight misinformation immediately when it happens, without requiring active involvement by the user. We believe that a trustworthy internet is vital to any democratic society, and scaling frontier AI systems safely requires that we can rely on this going forward. We're bringing Twitter/X Community Notes to every website so people can reliably spot wrong, misleading, or biased information right away.

What it does

Truthy is a browser extension that highlights passages of text that are likely to contain misinformation and provides sources to debunk false claims.

How we built it

We used Haystack to query the Claude API with the text to be checked for misinformation. We used embeddings by Cohere to implement RAG and help our tool provide relevant sources for our misinformation claims.

Challenges we ran into

Front-end: highlighting relevant information. Access issues: perceived incompatibilities with certain browsers.

Accomplishments that we're proud of

Our MVP works and we can further improve it through user feedback.

What we learned

There are many ways in which misinformation can occur. It takes effort to prompt language models in a way that perceives more nuanced claims.

What's next for Truthy

Provide political leaning and whole article summaries for news sources. Tracking the number of false claims of specific politicians, authors, influencers Searching through the scientific literature to check claims Check political bias of statements