Is GPT-4o Content Detectable?

With the release of OpenAI’s latest AI model GPT-4o “omnimodel” there is a  need to understand our AI detectors accuracy. 

This quick study looks at 1000 GPT-4o generated text results to answer if GPT-4o is able to be detected.

Is GPT-4o AI Content Detectable?

1. Yes - GPT-4o Text is Detectable with a 96.3% Accuracy for our Model 2.0 Standard and 97.8% accuracy for our 3.0 Turbo model.
   ‍
2. This is a slight drop from our GPT-4 performance of over 99% and is expected to close as we train our AI detector on GPT-4o content

Try our AI Detector here.

Dataset

In order to evaluate the detectability of GPT-4o, we prepared a dataset of 1000 GPT-4o-generated text samples.

AI Generated Text Data

For AI-text generation, we used GPT-4o based on three approaches given below:

1. Rewrite prompts: Generating the content by providing the model a customized prompt along with some articles (probably generated by LLMs) as a reference to rewrite from. (450 Samples)
   ‍
2. Rewrite human-written text: Generating the content considering the provided prompt to bypass the AI Detection tool by rewriting the human-written text which we fetched from an open-source dataset (325 Samples)some text
   1. One-Class Learning for AI-Generated Essay Detectionsome text
      1. Paper: https://www.mdpi.com/2076-3417/13/13/7901
      2. Dataset: https://github.com/rcorizzo/one-class-essay-detection
         ‍
3. Write articles from scratch: Generating the articles from scratch based on the given topics ranging from fictional and non-fictional diverse domains such as history, medicine, mental health, content marketing, social media, literature, robots, future etc. (225 Samples)

Evaluation

To evaluate the efficacy we used the Open Source AI Detection Efficacy tool that we have released:

• AI Detector Used in this Study
• AI Detector Efficacy Testing Article
• Open Source AI Detection Efficacy Testing Tool

Originality.AI has two models namely Model 3.0 Turbo and Model 2.0 Standard for the purpose of AI Text Detection.

• Use Version 3.0 Turbo - If your risk tolerance for AI is ZERO! It is designed to identify any use of AI even light AI
• Version 2.0 Standard - If you are okay with slight use of AI (i.e. AI editing)

The open-source testing tool returns a variety of metrics for each detector you test, each of which reports on a different aspect of that detectors performance, including:

• Sensitivity (True Positive Rate): The percentage of the time the detector identifies AI correctly.
• Specificity (True Negative Rate): The percentage of the time the detector identifies humans correctly.
• Accuracy: The percentage of the detectors predictions that were correct
• F1: The harmonic mean of Specificity and Precision, often used as an agglomerating metric when ranking performance of multiple detectors.

If you'd like a detailed discussion of these metrics, what they mean, how they're calculated, and why we chose them, check out our blog post on AI detector evaluation. For a succinct upshot, though, we think the confusion matrix is an excellent representation of a model's performance.

Below is an evaluation of both the models on the above dataset. 

Confusion Matrix:

Evaluation Metrics:

For this smaller test to be able to identify the ability for Originality.ai’s AI detector to identify GPT-4o content we look at True Positive Rate or the % of the time that the model correctly identified AI text as AI out of a 1000 sample GPT-4o content. 

‍

Model 2.0 Standard:

• Recall (True Positive Rate) = 96.4%

Model 3.0 Turbo:

• Recall (True Positive Rate) = 97.8%