Trying to build a cgpt to bypass ai detection

Community
1

Hello! I am interested in building a cGPT or a single prompt(even it is lenghty) to achieve a writing style that is not recognisable as ai written. I understand that tools like zerogpt work with machine learning and in order to bypass this we need to create a whole new pattern of sentences or writing perhaps.

I have tried simple prompts like add human like errors (typos, grammatical) keep it kinda unformal, have an unusual pattern to write and change the tempurature. Although these have worked for a single text and with 3 to 4 trys feeding with ai detection scores but it did not stayed consistent.

Long story short I think the way that I can achieve this by having a NEW type or pattern of writing even tho it is not(at least not excepted to as much ) human like.

My question is do you think that this is achievable with gpt 4o or gpt builder?
Or do you have any susggestions to lead me forward?

although I am able to somewhat tune down the scores of tools ( I know that they are not reliable, I don’t care) I can not achieve this in long texts and with certain topics.

I have been trying to create a single prompt or a gpt to achieve this for the last week non-stop and the limit of prompts makes it even harder.

And when I search the web about this, there are not a lot of useful info since they are just student looking to ease their work.

Thank you all of you that who has maken it so far.
I am eager for your inputs.
(English is not my first language I may had some miss wording I will try to clarify what I meant if you ask)
vedo

2

Considering that these tools (as you’ve said) are unreliable you’re better off throwing a dart to determine if it will pass their tests.

Based on that. Considering that luck is involved. The first thought that springs to mind is to create the content, and then run the content through a word-spinning GPT until it passes all of these detection tools. Kind of a brute-force attempt but if there is no logic, what else can one do :person_shrugging:.

You may be able to save all the results that passed and find some common denominator to improve your brute-force algorithm.

3

Thank you for your kind words

Do you think that a document filled with %0 ai scores and a document full of %100 90 texts can be used to educate a gpt?

If so can gpt really understand the form of the sentences or word choice?

you might not know the exact answer I am looking for an educated guess

4

It’d be interesting to see what other people say. It feels like a “yes” - especially if these detection tools use machine learning.

1 Like
5
  • A model like u want to define is of course concerning
    various anomalies which requires some high tech skill set
    to leverage GPT. Though the interpretation u stated is to
    generate the o/p on a single prompt may not be a
    unworkable.
  • From my view, I think it u should make a cornerstone of the
    AI detection algorithms rather on making a single prompt
    GPT.
  • For Ex : If a employee is travelling a business firm he might want to have information about the firm . Only one directional (only meetings, collaborations ,etc ) is not feasible, it must be considered to follow up additional reflection to learn about the firm.
  • What I mean is that creating text that consistently evades AI detection tools while maintaining a natural and coherent flow is a challenging task, especially for long texts and specific topics. In order to conflict this u have to understand AI detection algorithms, fine-tuning, prompt chaining and iterate the testing and feedback part more precisely which may be a possible way.
    Also, have to guide the model so sophisticated in generating text that mimics human writing and evades detection(Use Custom Prompt Engineering).
  • If u are leveraging all this, then u may also use neural networks for boosting the efficiency which reduces the computational cost compared to training a model from scratch.