I used the X API and GPTZero’s source finder tool - basically Perplexity on steroids; scans longer text and uses AI to extract claims, analyzes those claims against other online sources, checks THOSE sources for AI-generated content (because it’s an AI detector) and then responds back with a summary of its findings.
The X account is @GPTZeroAI and you can follow then tag @GPTZeroAI with “fact check” on any thread or post that contains an article. I’d love for people to test it out and help me improve it! I work at GPTZero so I can try to improve any part of the tool.
Some technical bits:
Used Redis as rate limiter to adhere to X api rate limits:
450 get tweets / 15 mins / app
900 get tweets / 15 mins / user
100 post tweets / 15 mins / user
100,000 post tweets / 24 hours / app
300 mentions timeline calls / 15 minutes / app
We don’t get webhooks from X api, so we poll the mentions timeline every 4 seconds. Tweets have a 10-12 second delay before becoming available in the api, so we scan a 4 second long sliding window that starts 20 seconds in the past.
Only 100 posts / 15 mins / user = multiple bot accounts that respond. Greedy algorithm to concentrate posting on the main account, then bot 1, then bot 2, etc
Used Inngest queue for pipeline robustness, automatic retries, and observability. Great tool, definitely recommend!
Used BrightData Web Unblocker API to scrape public articles (we do not scrape paywalled sites). This API avoids getting blocked while scraping articles by managing the proxying, req headers, fingerprinting, and CAPTCHAs. It’s good bang for your buck too ($1.50 / 1k requests).
The X account is @GPTZeroAI and you can follow then tag @GPTZeroAI with “fact check” on any thread or post that contains an article. I’d love for people to test it out and help me improve it! I work at GPTZero so I can try to improve any part of the tool.
Some technical bits:
Used Redis as rate limiter to adhere to X api rate limits: 450 get tweets / 15 mins / app 900 get tweets / 15 mins / user 100 post tweets / 15 mins / user 100,000 post tweets / 24 hours / app 300 mentions timeline calls / 15 minutes / app
We don’t get webhooks from X api, so we poll the mentions timeline every 4 seconds. Tweets have a 10-12 second delay before becoming available in the api, so we scan a 4 second long sliding window that starts 20 seconds in the past.
Only 100 posts / 15 mins / user = multiple bot accounts that respond. Greedy algorithm to concentrate posting on the main account, then bot 1, then bot 2, etc
Used Inngest queue for pipeline robustness, automatic retries, and observability. Great tool, definitely recommend!
Used BrightData Web Unblocker API to scrape public articles (we do not scrape paywalled sites). This API avoids getting blocked while scraping articles by managing the proxying, req headers, fingerprinting, and CAPTCHAs. It’s good bang for your buck too ($1.50 / 1k requests).
LLM: combo of o3-mini and 4o-mini.