The problem Tweet Detective solves

As the use of artificial intelligence continues to grow, the internet has seen a surge in AI-generated content. While AI-generated content can be useful in some situations, it poses a significant problem for platforms that rely on high-quality content from human creators. The increasing prevalence of AI-generated content on social media platforms and other online media has resulted in spam and automation that reduces data reliability. This, in turn, is affecting the quality of the overall user experience, and platforms are seeking solutions to address the issue.

Our product is designed to solve this growing problem by detecting and weeding out AI-generated content on Twitter. Our team has developed two custom fine-tuned GPT3 models trained on a dataset of over 100,000 tweets. These models are designed to detect AI-generated content with high accuracy on Twitter. Our AI text detection model boasts an accuracy rate of 98%, while our Anomaly Detection model has an accuracy rate of 87%.

We ensured the tweets were pre-GPT3 to avoid any adulteration of the dataset.

The product we have developed is a simple chrome extension that can be installed on any web browser. Once installed, the extension activates in the background when Twitter is opened and detects whether AI has written the tweet.

This provides a much-needed layer of protection for platforms that rely on high-quality content from human creators. With the ability to detect AI-generated content, our product can help platforms maintain the integrity of their content and ensure that their users have a reliable and trustworthy experience.

Challenges we ran into

Fetching 50k+ tweets through Twitter API of 675+ users over the past year and fine-tuning GPT3 on OpenAI:

• This took a lot of time and resources to get the desired accuracy for the anomaly detection model, which generates a persona of the Twitter handle and then compares it with other tweets to see if the tweet is an anomaly.

Generating AI-written Tweets through GPT3 API:

• We took a unique approach for the AI detector model; we fed GPT3 tweets and asked it to extract the tweet's context and generate its own tweet for the same context.
• This helped us create a custom dataset of AI and Human-generated tweets for the same context.

Creating a chrome extension:

• Since this was our first time developing a chrome extension, we had to learn a lot about its working, how the background processes work, how they communicate with UI and how it can be used to inject additional data on top of the UI