Human on the Podium

Why

The use of AI in political speeches and official documents often goes undisclosed. Our analysis of over 200 government documents found clear evidence of AI usage, from complete speech generation to factual errors that indicate AI involvement. While using AI is not inherently wrong, people have the right to know when their elected leaders use it. We created this system to enable that transparency.

What we built:

A live AI detector web app that classifies voice as human or AI, shows confidence levels, maintains history and classifies a chunk of sentences.
An vectorized database of hundreds of government documents and video speeches that found 22 percent contained undisclosed AI usage.
A Chrome extension that works in video meetings like Zoom and Google Meetings to detect whether someone is reading from an AI assistant, useful for interviewers.
API connected to an ESP32 hacker badge that uses light signals to indicate AI-generated speech (red light) or human speech (green light).

Technologies and tracks

GPTZero: It's the core of our app - AI usage detection, confidence levels, per chunk reasoning, fact checking.
ElevenLabs: It's for STT transcription with their real-time API and STT transcription with their YouTube video extension. We are utilizing their voice models to generate voices when giving feedback after detecting AI.
elastic.co. We are utilizing them as part of our embedding vector database. I'm hosting all of the documents and speeches, and the vectorized version of them, on their serverless cloud to visualize them on a 2D or 3D graph.
Google Gemini API. We are using Google Realtime API for the real-time talks done during meetings, such as Google Meet.
Rox as we are aiming to build something very meaningful from messy and unstructured data, cleaning it up, and building something very meaningful with AI agents.

How we built it

For the live AI detector: Audio/Video stream > STT with ElevenLabs > chunks of 80 words > GPTZero > Analysis > API hit to hacker badge > Shows live signals if how much AI is being used.

See our very cool demo here :))

For the web extension: Tab audio → offscreen recorder → Gemini transcription → chunking → GPTZero → analysis → WebSockets → side panel + report. WXT Chrome extension + Fastify backend. Local “AI slop” patterns give a fast signal before GPTZero returns.

Cool stuff we’re proud of

Joshua has made the hacker badge work, and we successfully were able to integrate with our next app via an API that Joshua created on Lua. I can directly hit that API and make the hacker badge change the way we want.

Viho has made the Chrome extension work. This can be used in Zoom or Google Meetings. It is helpful for interviewers where people use AI to answer questions, and every 30 seconds, our app can detect if it's AI or not with some confidence level.

Gurkirat is really, really proud of the idea and coming up with a unique and niche area to target. This has come from a place of actual work, at the IJF, where he works, and got inspired with this idea, where the company does investigative journalism. This idea is to live-detect AI in speeches, videos, and documents. We had a dataset of around 200 documents and some of the previous speeches, and analyzing them and creating a vectorized visualization of where AI is being used has been really great.

Harman is proud of his work. He's been keeping his frontend game top notch and the cool stuff he's able to do quickly with DeepSeek models, and providing everything quickly and presenting to the team, really working collaboratively and just enjoying here.