Originality.ai is the Most Accurate AI Detector According to an Extensive Study “RAID”

Researchers at UPenn, University College London, King’s College London, and Carnegie Mellon University recently completed the most complete study yet to evaluate AI detector efficacy - RAID: A Shared Benchmark for Robust Evaluation of Machine-Generated Text Detectors (study).

Originality.ai’s AI detector was the most accurate detector in the study! 

The study looked at 12 AI detectors, 11 different text-generating LLMs (like ChatGPT), and 11 types of adversarial attacks (like paraphrasing), resulting in a dataset of over 6 million text records. This is the most robust evaluation of AI text detectors to date, and our own AI detection accuracy study is likely the second most in-depth study. 

Below is a summary of the study, including the key findings related to Originality.ai’s industry-leading performance and a couple of our weaknesses identified in the study that we are excited to address. 

Study Details:

1. The ambitious study looked at…some text
   1. 12 AI Detectors (Originaltiy.ai, GPTZero, Winston, ZeroGPT)
   2. 11 Text Generation Models (LLM’s such as ChatGPT, GPT-4, Llama etc)
   3. 8 Domains of Text (different types/categories of text)
   4. 11 Types of Adversarial attacks (strategies to make text undetectable)
2. DataSet Sizesome text
   1. 6,287,820 texts!
   2. Dataset: https://github.com/liamdugan/raid
3. The study used a 5% False Positive Threshold for all tests (more on what this means at the bottom of this blog post).

Study: https://arxiv.org/abs/2405.07940

Originality.ai’s AI Detector Results: 

1. Most Accurate AI Detector on Base Dataset: Originality.ai was the most accurate AI detector on the base dataset with 85% accuracy vs the closest competitor at only 80%.
   ‍
2. Most Accurate AI Detector on Adversarial Datasets: Originality.ai performed the best on the majority of the adversarial techniques. Out of the 12 detectors, Originality.ai placed 1st in 9 of the 11 tests, 2nd in 1 test, and performed poorly on 2 rarely used bypassing techniques which we will be addressing shortly.
   ‍
3. The Most Accurate AI Detector Across All Domains: Across all types of content (domains), Originality.ai’s leading performance held. Out of the 12 detectors, it placed first in 5 of the 8 domains and second in the remaining 3 domains.
   ‍
4. Exceptional Performance on Paraphrased Content: Originality.ai’s ability to identify paraphrased content was exceptional with 96.7% accuracy compared to the next closest competitor at 80% and the average of all other detectors at only 59%! 

It is important to note these accuracy scores reflect a 5% false positive threshold, more on what that means below.

Originality.ai’s model 2.0 Standard was used for these results, however we would expect model 3.0 Turbo (which was released 1 month after the authors used Originlaity.ai) to outperform 2.0.

Finding #1 - Originality.ai is The Most Accurate AI Detector Across 11 AI Models

11 AI models were used to generate a non-adversarial dataset and it was evaluated against all AI detectors. These included the most popular AI models such as ChatGPT, Llama, Mistral and GPT-4. 

Originality.ai was the most accurate across the test achieving 98.2% accuracy on ChatGPT content and an average of 85% across all 11 AI models.

Finding #2 - Originality.ai Is The Most Accurate Detector on Adversarial Techniques

Of the 11 adversarial (adversarial means trying to make undetectable AI content) techniques used Originality.ai was the most accurate of the 12 detectors in 9 of the 11 tests, 2nd in 1, and performed poorly in 2 rarely used bypassing techniques.

“Originality.ai Rank” shows how we performed vs the 12 AI detectors for each of the Adversarial Bypassing Techniques.

Clearly, Originality.ai’s AI Detection needs to improve on the Homoglyph and Zero-Width AI bypassing strategies, but was the leading AI detector on the other more common bypassing techniques.

Finding #3 - Originality.ai is The Most Accurate Across Different Types of Content

The authors looked at how AI detectors performed on different types of content (news articles vs poems etc). 

Originality.ai was the most accurate across 5 of the 8 types of content “domains” and the 2nd most accurate in the other 3 domains.

Important to note is that the study did not include the domain/type of content we at Originality.ai focus most on, i.e. web content/marketing content.

Finding #4 - By Far the Most Accurate Paraphrase AI Detection

The most common method used in trying to make AI content undetectable is “Paraphrase Plagiarism”. The strategy uses a paraphrasing AI like Quillbot to change words in an attempt to bypass AI detection tools. This is the strategy used by common tools such as Undetectable.AI.

Originality.ai performed uniquely well on this common adversarial bypassing strategy achieving 96.7% accuracy while the average accuracy from other detectors was 59%.

What is a 5% False Positive Threshold?

In the context of text detection, a false positive occurs when a detector incorrectly labels a piece of human-written text as machine-generated. False positives can occur for several reasons listed here. The False Positive Rate (FPR) is the percentage of these incorrect labels out of all the human-written texts evaluated. For instance, if a detector examines 100 human-written texts and incorrectly labels 5 of them as machine-generated, the FPR would be 5%.

Example of a 5% FPR Threshold Being Applied:

Why a 5% FPR Threshold?

A 5% FPR threshold is a benchmark used to ensure a balanced evaluation of text detectors. It was selected by the authors of this study. 

Here’s why it matters for this study:

1. Consistency in Comparison: By setting a common FPR threshold, different detectors performances can be fairly compared.
   ‍
2. Balancing Precision and Recall: A lower FPR means fewer human texts are mislabeled as machine-generated, which is important for practical applications where such errors could have significant consequences. However, setting the FPR too low could prevent the detection of actual machine-generated texts. The 5% threshold strikes a balance and is a common threshold set when evaluating classifiers within the field of Machine Learning.  

Comments for the Study’s Authors

This is the best study by far that looks to address if AI detectors work and what their limitations are. 

I would have several comments for the authors if the study is continued moving forward or for future studies …

Cover More Relevant Domains:

Most users of AI detectors fall into 2 types of content… Web Content and Academia. 

The domains that were selected for this study included essentially no web content (Abstracts, Books, News, Poetry, Recipes, Reddit, Reviews and Wiki). Our detector is specifically trained on web content and it would have been great to see that domain included. The societal importance of being able to detect AI-generated poems is lower than the importance of being able to detect AI-generated news.

False Positive Rate Threshold Makes Sense But Incomplete for Users:

This study standardizes on a FPR threshold of 5%. However, there is use cases where a VERY low FPR rate is required and use cases where a higher false positive rate is acceptable if it means very very little AI can get past the detector. Ultimately the users of AI detectors will be the ones who want to make a decision on what the right trade-off between correctly identifying AI as AI and correctly identifying human as human is. We ask to consider other methods of displaying detector efficacy instead of just accuracy with an FPR threshold including confusion matrix and corresponding efficacy scores (F1, TPR, FPR, etc). 

AI Edited Dataset:

One dataset that is missing that will become more prevalent is AI-edited content (whether that be lightly AI-edited or heavily AI edited). With AI editing tools increasing the amount of AI they inject into an edited article, the ability to identify lightly edited AI content is important. For a small analysis here is the impact of Grammarly editing on AI detection.

Our Learnings

2 Odd Bypassing Strategies to Address:

Our AI detector performed poorly on 2 bypassing strategies. Homoglphyh and Zero-Width-Space attacks. These two strategies work by creating text that seems readable to humans but is less readable to machines. We strongly recommend against using this strategy to try and trick Google’s AI detection. 

• Homoglpyh Bypassing: This is when Cyrillic script is used to replace ASCII characters that appear the same to the human eye but reads like gibberish to a computer. This would not be a wise strategy for web publishing (Google would not read the text correctly) and will be an easy bypassing strategy for us to close. 

• Zero-Width Space: This strategy similarly makes the text readable to humans but not machine-readable with the addition of a unicode character. From the paper “The Unicode zero-width space “U+200B” is a character that exists in text encoding but is not visible to human readers in most scenarios. Thus, for this attack, we insert this character at every possible opportunity (before and after each visible character in the generation).” This will also be straightforward for our team to address in our models. 

Originality.ai focuses on providing an accurate AI detector for web publishers, so focusing on Homoglyph bypassing and Zero-Width Space had not been a priority for us (these aren’t machine-readable and therefore not a good strategy for web publishing). However, in order to ensure we are the most accurate across all bypassing strategies, we look forward to having our AI research team close these 2 smaller weaknesses in our detector! 

A Sub 1% False Positive Solution is Needed - Our “Lite” Model is being trained

We will shortly have 3 models for users to choose from depending on their desired tradeoff between false positives (calling human text AI) and false negatives (calling AI content human). 

• Turbo - When you want to be sure NO AI touched the text but are okay with being sensitive to AI editing like Grammarly and having a slightly higher false positive rate.
  ‍
• Standard - Best for most users who want as accurate AI detection as possible, accept a very small amount of AI editing and want to minimize false positives to the low single digits <5%.
  ‍
• Lite (coming soon) - When you want to have the lowest chance of calling human written text AI as possible but are okay with a lower accuracy rate. 

This is well represented on the graph Figure 4 from the study and overlaying our target ranges for our 3 AI detection models.

Our model showed leading performance for a given false positive rate until it got near 1%. Our Lite model will have industry leading low false positive rates. 

Final Thoughts

It is great to see a high-quality efficacy study after some flawed early studies provided misinformation about AI detectors and their efficacy. Thank you, researchers - this study will help people! 

It is most exciting to see our approach with our Blue Team (building detectors) and Red Team (trying to beat our detector) is working. Our entire AI research team should be very proud of the industry leading results they are producing. 

If anyone is interested in running a study that includes Originality.ai, we are happy to make credits available and have our own open source tool available for evaluating detector efficacy for free.