if you want to avoid getting caught using AI, whether for school or for SEO
you should learn how plagiarism & copyright detection works in general
detecting AI-generated content isn't new
"document fingerprinting" has been studied as early as 2003
here's the latest research:
you should learn how plagiarism & copyright detection works in general
detecting AI-generated content isn't new
"document fingerprinting" has been studied as early as 2003
here's the latest research:
1. Watermarking
one of the biggest problems with AI-generated text is that it has a hidden "watermark"
there are words or sentences that will flag your text as AI-generated
there are currently four known ways to bypass watermarking in GPT-generated text:
one of the biggest problems with AI-generated text is that it has a hidden "watermark"
there are words or sentences that will flag your text as AI-generated
there are currently four known ways to bypass watermarking in GPT-generated text:
Bypass #1: Emoji Attack
Ask GPT to put an emoji in between every pair of words
Then, ask it to remove the emojis between those words
This bypasses detection algorithms that rely on analyzing sequences of words, including GPT-Zero
Ask GPT to put an emoji in between every pair of words
Then, ask it to remove the emojis between those words
This bypasses detection algorithms that rely on analyzing sequences of words, including GPT-Zero
Bypass #2: Translation Attack
Ask the model to write in a different language
Then translate the response to the language of choice
Warning: this can significantly degrade the quality of the text depending on how accurate the translation is
Ask the model to write in a different language
Then translate the response to the language of choice
Warning: this can significantly degrade the quality of the text depending on how accurate the translation is
Bypass #3: Paraphrase and Substitution Attack
Get your response from the model
Manually replace some of the words with synonyms, and paraphrase some parts
If you can replace a sufficient amount of the text, it won't be caught by the detection algorithm:
Get your response from the model
Manually replace some of the words with synonyms, and paraphrase some parts
If you can replace a sufficient amount of the text, it won't be caught by the detection algorithm:
Bypass #4: "Just Ask" Attack
You can literally just ask ChatGPT to bypass detection algorithms
Under the hood, it changes its model parameters
You can ask to increase the temperature, frequency penalty, and presence penalty. These factors will often evade detection algorithms.
You can literally just ask ChatGPT to bypass detection algorithms
Under the hood, it changes its model parameters
You can ask to increase the temperature, frequency penalty, and presence penalty. These factors will often evade detection algorithms.
2. Document Fingerprinting
In 2003, a group of researchers created a copy-detection algorithm that checks three properties
These three properties have been used as the basis for plagiarism detection algorithms in schools and search engines:
In 2003, a group of researchers created a copy-detection algorithm that checks three properties
These three properties have been used as the basis for plagiarism detection algorithms in schools and search engines:
Property #1: Whitespace Insensitivity
When matching text, they ignore additional whitespace, whether you press the "spacebar" or "enter" keys
It also does not check capitalization or punctuation differences
When matching text, they ignore additional whitespace, whether you press the "spacebar" or "enter" keys
It also does not check capitalization or punctuation differences
Property #2: Noise Suppression
Chunks of your text will be matched against existing text on the internet
They won't detect common words or idioms, but if you've copied a sentence or two, they'll know
Chunks of your text will be matched against existing text on the internet
They won't detect common words or idioms, but if you've copied a sentence or two, they'll know
Property #3: Position Independence
You can modify the document contents by changing the:
- position of paragraphs
- deleting paragraphs
- adding extra paragraphs
These changes will not affect the detection algorithm–you still still get caught
Make sure you change the contents.
You can modify the document contents by changing the:
- position of paragraphs
- deleting paragraphs
- adding extra paragraphs
These changes will not affect the detection algorithm–you still still get caught
Make sure you change the contents.
Don't get left behind.
Knowing these tricks and understanding prompt injection will help you use AI to its fullest extent.
---
If you're interested in this topic, see the link to my newsletter in my profile in case Twitter dies
I won't spam–I only post infrequent, long essays
Knowing these tricks and understanding prompt injection will help you use AI to its fullest extent.
---
If you're interested in this topic, see the link to my newsletter in my profile in case Twitter dies
I won't spam–I only post infrequent, long essays
Sources:
Undetectable Watermarks for Language Models (2023)
eprint.iacr.org
eprint.iacr.org
Paraphrasing evades detectors of AI-generated text, but retrieval is an effective defense (2023)
arxiv.org
arxiv.org
Winnowing: Local Algorithms for Document Fingerprinting (2003)
theory.stanford.edu
theory.stanford.edu
Loading suggestions...