GPTZero Detects AI Content in Quora Posts with 99% Accuracy

Quora is one of the world’s most popular question-and-answer sites, and like many online platforms today, it’s seeing an increasing mix of AI-generated content. This prompted us to explore the following question: how well can AI detectors actually tell the difference on Quora posts and comments? 

To find out, we evaluated GPTZero on Quora posts and their top comments. Our results show that GPTZero pulls ahead of the competition with over 99% accuracy on both. 

At GPTZero, we’re constantly expanding our AI detection coverage to the latest models and new writing domains. In the last few months, we’ve been training on more human data, including common crawl (publicly available data that covers a significant portion of the internet), scientific articles, and creative writing. 

Recently, we trained on a new dataset of blog posts, and decided to check our performance on publicly available Quora posts and their top comments. We’re particularly interested in evaluating domains we don’t explicitly train on as a way of testing how our model performs on out-of-domain edge cases. 

When we train on new data from a domain like blog posts, it’s helpful to see whether that performance transfers to similar question-and-answer formats. This is part of our broader evaluation strategy, which helps us identify areas for improvement and validate that our detector works as expected in unseen scenarios.

Dataset

This benchmark includes 500 human-written Quora posts, 500 human-written comments, and corresponding AI-generated pairs. We split this into two sub-datasets with balanced human and AI data, with 1000 examples each. For this benchmark, we used LLMs that are popular for web users and academic API usage, including GPT5, GPT4o, GPT4.1, Gemini 2.5, Gemini 2.0, and Claude Sonnet models. 

Our expectation was that both AI detectors would perform better on posts than comments, since comments tend to be shorter and have a more casual writing style. To provide enough context for proper AI detection, all of the posts and comments in this benchmark are at least 250 characters long. 

Results

Table 1 shows the results of our Quora posts benchmark, with 500 human and 500 AI-generated posts. While both detectors have no false positives, GPTZero wins on recall. 

Table 2 shows the results of our Quora comments benchmark, with 500 human and 500 AI-generated comments. Since this is typically a harder task for AI detectors, both models see decreases in performance. Even so, GPTZero remains 99% accurate while Pangram’s detector struggles to identify AI-generated comments. 

Table 1: AI detection performance on Quora posts

Table 2: AI detection performance on Quora comments

API vs. Dashboard Results

When we tested some of the AI-generated comments that were misclassified by the Pangram API, we found that they were correctly classified as AI on their dashboard. We expect Pangram’s latest dashboard model will have improved performance. 

At GPTZero, our model releases go out to 100% of users immediately, so both dashboard and API users have access to the same model. This keeps our enterprise clients up to date with the latest model, and protects students and writers from disagreements with teachers, professors, and reviewers. 

An example of API usage in academia is when teachers use our API through an LMS (learning management system) for AI detection in assignments, so they can evaluate them in batches. Students often use the dashboard since they only have to check their own work. We make sure that both parties have access to the same model version so they get the same response no matter where the request comes from. 

Detecting AI in Quora and related content

This benchmark shows that GPTZero is the best option for AI detection on Quora content, especially comments. We expect this result to generalize to more blog posts and forum-based writing, making GPTZero a strong option for anyone considering which AI detector to choose for content monitoring.