I recently tested and made a system on how to bypass AI Detection tools. It involves paraphrasing and asking AI to deliberately make misspellings. It worked 90% of the time but is time consuming. Now I just told ChatGPT to "write an article about "XXY" so that 100% of the time it does not contain any machine-generated content." That pretty much solved this problem and bypassed any AI Content Detection tool available so far. Of course, this may change as technology grows. It's just a matter of who is diligent enough to go deeper into the rabbit hole. Cheers!
@@CristiVladZ Unfortunately, its not a panacea. The writer will need to test the prompts if it gives the preferred results. The best method now is back to the manual way. Run the content to Quillbot paraphraser then deliberately make some typographical error and ask ChatGPT again to rewrite word for word but insert 1 grammatical error. This will make the AI detection tool see it as Human generated content. I tried GPTZero and it is more robust. No amount of rephrasing and typo can trick it. The challenge now is how to bypass that.
In summary, the text describes the use of a language model called Roberta-based OpenAI detector to determine if text has been generated by a chatbot called GPT (Generative Pre-training Transformer). The detector was trained by fine-tuning a Roberta-based model with the outputs of GPT2. The text also mentions Hugging Face, a platform for AI-related resources, and its hosted API interface for the detector model. The text then goes on to test the effectiveness of the detector by feeding it text generated by a version of GPT called Chat GPT and text taken from a Google search on implementing encryption in Android apps. The text concludes by stating that the detector is able to accurately identify fake text generated by GPT, but may not work as well for text taken from other sources.
hmm, I think if I implement en detector I will preprocess the input to remove the characters that are not encountered in the vocabulary of the model. but yes, there are many adversarial examples to trick the detector I think.
If the language model can analyze the probability of a word coming next it certainly could be made to identify low-probability breaks in the chain. But its a seesaw of fix and brake.
I played around with the detector and it’s more focused on grammer, I made 3 grammer errors by removing a period and 2 commas and it said it was real, it is very interesting but I wouldn’t trust it 100%, I think it’s a 50/50 depending on who wrote it, how they wrote it, ect.
Has the code been updated since your video? I mean, out of curiosity, I copied some AI text into the detector... and it detected its 99.5% fake... then I tried to add the / but its no longer causing the results to change
I had chatGPT write a RU-vid script on a topic. The output was good but of course, failed the Hugging face test. I've been using Quillbot to paraphrase and that has been working pretty well. So, in the same chattGPT chat for the YT script, I wrote" write it so it won't be AI detected" and sure enough what it produced was 98.37% real when put through Hugging face.
Me: write me a script for RU-vid on how to pass originality checks on chat gpt generated work GPT: Hello and welcome to this tutorial on how to pass originality checks when using OpenAI's GPT-3 language model. So you've used GPT-3 to generate an essay or article, but now you're worried it may not pass originality checks. No need to worry, as there are a few simple steps you can follow to increase your chances of passing these checks. Step 1: Know the software being used for originality checks. Different software will have different methods for detecting machine-generated content, so it's important to know which software will be used for the originality check. This will help you determine the best strategies to use to avoid detection. Step 2: Make edits to the text generated by GPT-3. The most effective way to avoid detection is to make significant changes to the text generated by GPT-3. This includes adding, removing, or rephrasing entire sentences and paragraphs, as well as making small changes like correcting grammar and spelling errors. The more changes you make, the lower the chances of detection. Step 3: Use plagiarism checkers. Another useful tool to help you pass originality checks is a plagiarism checker. These checkers will analyze your text and compare it to other sources on the internet, identifying any instances of plagiarism. This will give you an idea of how original your text is, and help you make any necessary changes. Step 4: Write in your own voice. When making changes to the text generated by GPT-3, try to write in your own voice. This means using your own words and phrasing, rather than simply copying and pasting from the GPT-3 generated text. Step 5: Avoid using overly technical language. Using overly technical language is another common characteristic of machine-generated text, so it's important to avoid this. Instead, try to write in a more conversational tone, using language that is easy to understand. In conclusion, by following these five simple steps, you can greatly increase your chances of passing originality checks when using GPT-3 generated text. Remember, the key is to make significant changes to the text, write in your own voice, and avoid using overly technical language. Good luck!
Interesting, which made me wonder, can you just ask chat gpt to do it for you. Turns out, of course. Original prompt: Write a one paragraph explanation of how SSD works Result: 99% label 0 (fake, written by AI) Solid State Drives (SSDs) are a type of storage device that uses flash memory to store data. Unlike traditional hard disk drives (HDDs) which use spinning disks to access data, SSDs have no moving parts, making them faster and more reliable. They work by using a series of interconnected flash memory chips to store data, which can be quickly accessed and written to using an integrated circuit called a controller. The controller manages the flow of data to and from the memory chips, allowing the SSD to access and write data at much faster speeds than an HDD. Followup prompt: [exact prompt redacted...] Result: 99% label 1 (real, written by a human) Solid State Drives (SSDs) are a type of storage apparatus that employs flash memory to keep data. Unlike conventional hard disk drives (HDDs) which utilize spinning disks to access data, SSDs do not have moving parts, making them quicker and more dependable. They function by utilizing a series of interconnected flash memory chips to keep data, which can be rapidly accessed and written to using an integrated circuit referred to as a controller. The controller oversees the flow of data to and from the memory chips, enabling the SSD to access and write data at much faster speeds than an HDD. As you can see, GPT is quite capable of defeating these checkers in a way that maintains a nearly entirely natural writing style (it doesn't appear clearly manipulated). As long as there are models capable of generating this text, the same models will be capable of defeating any system or watermark attempting to detect it as written by said model.
i also was able to do this and i concur that it works using the right prompts. prompts aren't really special. just think of something that you insert while writing that is typical to what humans do
@@CristiVladZ Yeah, people can't digest long form. It's better for the performance of your video. If you edited it slightly you would get a lot more views.
@@danielprokoshinthose people are not what I'm looking for in my audience. So it's better if they skip. By making long form, the content will stick to those who resonate with it...If you don't like it, you shouldn't be here. I don't need TikTok-like audience...
@@CristiVladZ I get your sentiment but to generalize a 'TikTok-like audience' when there are over a billion monthly active users is a bit silly. Short, concise and well edited content is good content. If you don't want to make good content, that's your prerogative. But just say that you don't want to put in the effort instead of making generalizations.
@@CristiVladZ I get that, I am very against TikTok and informative TikTok-style content as well. However, I am against informative TikTok-style content because it usually brushes over important things and tries to summarize way too much information into a 1-minute video. In my opinion, you could summarize the useful information here into a 3-4 minute video that wouldn't resemble TikTok content, it would just be more concise and get to the point faster.
You dont need to worry about any detection...just stop using these tools and write it yourself. Problem solved. Google will always find out. Humans just stop being lazy!
It’s not even really being lazy. Life is too short to spend on stupid stuff like writing a ten page essay that you will not even remember about in 2 years.
@@ihateteo we never had gpt3 a few years back and we survived very nicely indeed. Gpt3 and beyond including everything associated with it will mainly be used to try and trick google..that is my point. People will always take the path of least resistance. No matter what you do..google will get you. So just go back to pen and paper and use your brain.
