Generative AI detection

Generative AI detection

Detect AI generated content

Here’s a demo of a generative AI detector. This tool only covers very short documents, but if you’d like to run a more thorough analysis, please get in touch.

This tool attempts to detect signs that a document has been written by a generative AI model such as ChatGPT, GPT-3, BARD or similar.




Results appear below. Green means that a word is likely not AI generated, red means it is probably AI generated, and orange means that the word could have been AI generated but it's not the top choice of our AI in that context.

Results

[enter a text in the window above]

Generative AI and stylometry analysis

Analyse documents

Accused of cheating, plagiarism, or using Generative AI or GPT for an assignment? Do you have a legal dispute over authorship?

Is it possible to prove whether a document was written by AI?

It’s not possible to prove unequivocally whether a person used generative models to write a document. However, we can look at the content of the document and check what a generative model such as GPT-4 would predict as the next word given the preceding tokens, and what probability it would assign to the words that actually appear in the document.

We can quantify how surprised a generative model would be if it saw the document under investigation, and use this to derive a probability that the document was authored by an AI. This measurement of the degree of surprise is called perplexity.

Can you prove authorship of a document?

For cases of disputed authorship among humans, where parties dispute who wrote a document, we can perform a stylometric analysis. This field is known as forensic stylometry. You can try a similar demo of a trained forensic stylometry model using the Burrows' Delta algorithm.

Expert advisor and expert witness reports

Thomas Wood undertakes expert witness and expert advisor engagements for clients in the area of text analysis. Recent engagements include a civil dispute over authorship of documents in the legal sphere by a multinational accountancy firm.

Detecting generative AI texts using Faststylometry library

In 2024, Stefan Björk-Olsén experimented with our open source Faststylometry Python library on detecting AI-generated texts:

What we can do for you

Transform Unstructured Data into Actionable Insights

Contact us