Is GPT-5 Content Detectable?

OpenAI released GPT-5 on Aug 7, 2025 and this study looks to understand if AI detectors remain accurate on GPT-5 generated content.

Additionally, we examine what makes GPT-5 unique in terms of writing. Not just is it still detectable but what are its unique features, does it overuse delve, em-dash etc? 

Can AI Detectors detect GPT-5 Content?

Yes, GPT-5 content is able to be detected by Originality.ai at a 96.5% accuracy rate. Currently, GPT-5 content is less detectable than GPT-4o; however, that gap will close quickly as Originality.ai trains on more content from GPT-5.

99% Accurate for Rewritten Content based on a quick test of 100 samples

94% Accurate for Newly Generated GPT-5 Content based on a quick test of 100 samples

Test #1 - Originality.ai vs GPT-5

Method:

1. We built 2 quick datasets to test the detection accuracy as soon as GPT-5 was available.
   
   1. Dataset 1 - Rewrites of human content - 100 samples
   2. Dataset 2 - Adversarial (prompts meant to make content less detectable) AI generation - 100 samples 
2. Both datasets were tested against our latest AI detection model 1.0.1

Results:

• 99% Accurate for Rewritten Content based on a quick test of 100 samples
• 94% Accurate for Newly Generated GPT-5 Content based on a quick test of 100 samples

Test #2 - Other AI Detectors vs GPT-5

Can other detectors such as GPTZero, TurnitIn, CopyLeaks, Quillbot and Grammarly detect GPT-5 generated content?

Method:

The first 10 samples from the GPT-5 rewritten content are tested against multiple detectors.

‍Results:

• Originality - 10/10 Correct
• GPTZero - 1/10 Correct
• Quillbot - 2/10 Correct
• Grammarly - 2/10 Correct
• Copyleaks - 5/10 Correct

Note - I expect the performance of all AI detectors to increase significantly over the coming weeks/months as models are updated. 

What is New with GPT-5 for Creative Writing?

Open AI says that GPT-5 is the “most capable writing collaborator yet”

Basically, what this means is that OpenAI says GPT-5 can…

• Handle more complex writing tasks
• Be more consistent at following a specified style of writing

How is GPT-5 Writing Different

Does GPT-5 still love the em-dash? Yes! 

Out of the 100 human-AI rewrites, none of the 100 human samples used any em-dash’s. However, the GPT-5 rewritten content included the em-dash in 67 of the articles. 

• Human Content: 0% Em-Dash
• AI Content: 67% Em-Dash

GPT-5 Produces Similar Readability Score to Human Input

Previous models GPT-4o etc often rewrote content and the output of the content often clustered around similar Flesch Kincaid Readability scores. 

GPT-5 is being positioned as a superior writing tool that is able to follow a specific style more consistently. 

If that is the case it should show up in readability scores staying consistent with the human text it is being asked to re-write. 

To test this we looked at dataset of Human content and AI content to see how consistent the readability scores were between the input and output using our Readability Checker.

Results:

GPT-5 mirrored the readability score of the human content

Average Flesch Kincaid Readability Score:

• Human: 49.6
• GPT-5 Rewrite: 46.5

With more spread for AI written content (19.1 vs 14.9 for human written)

There is a clear, moderately strong positive relationship. Higher-readability human originals tend to yield higher-readability AI rewrites. 

Take-Away - GPT-5 Better Mirrors Human Writing

GPT-5 rewrites generally mirror the clarity of their human source material. In our 100-sample test, human passages averaged a Flesch–Kincaid score of 49.6, while GPT-5’s versions landed at 46.5—only about three points lower and still in the same readability band. The two sets of scores are moderately correlated (r ≈ 0.58), so a well-written original usually stays well-written after the AI pass. The main difference is consistency: AI outputs show a wider spread in scores (standard deviation 19 vs 15), meaning they sometimes swing from ultra-concise to slightly dense. Bottom line: GPT-5 tends to keep the readability level you give it.

Conclusion

GPT-5 content is detectable by Originality.ai but for now it makes AI detectors less accurate. We would expect all leading AI detectors to get more accurate in the coming weeks/months on the latest model release from GPT-5. 

Yes - GPT-5 still loves to use the Em-Dash

GPT-5 mirrors the style/readability score of human writing more than previous models.