It shouldn’t be, but ChatGPT is slowly starting to become the norm both in classrooms and the workplace. No one would bat an eye anymore if a junior developer opens ChatGPT to create a RegEx pattern, or if the person sitting next to you stealthily boots up the app to ask why the moon is round. But there’s one problem:
It’s so transparent.
Despite all the advancements, a discerning eye could still pick ChatGPT writing from a lineup. The same goes for AI detectors, which are evolving and becoming better at detecting AI writing and preventing false positives. So, what can you do?
Read along as I discuss the science behind AI writing detectors, how you can tweak your writing to avoid detection, and the best AI bypass tools in the market.
How Does AI Detection Work?
Before we can make ChatGPT writing undetectable, we must first discuss how AI writing detection works. There are three components to this: training, analysis, and feedback loops. Here’s a more in-depth look into each of them, and what makes them so important.
Training
Like ChatGPT, AI detectors are powered by knowledge extracted from thousands, if not millions, of texts and data through learning. The dataset used by detectors contains both human-written and machine-generated text.
During this stage, the model will look for the linguistic features that separate human and AI text. This includes patterns, styles, and word usage. Once these are extracted, the information is ready for analysis.
Analysis
Imagine you’re a computer. You don’t see language the same way as people do. How would you, given the tools you have at your disposal, create a system that can distinguish if a different machine writes a text? The answer is simple– sort of.
Through mathematics and probability, of course.
Here’s how a machine differentiates human writing from machine writing in simple terms.
Perplexity
Perplexity measures the predictability of the next word in a sentence. A really good, but somewhat primitive, example of this is when you’re texting someone, and a pop-up appears above your keyboard with suggestions of what your next word could be.
The difference is that your phone only measures the probability using your own text messages as its dataset. LLMs use the entirety of their dataset, which often includes billions of tokens.
AI models aim for low perplexity because it’s often the expected output by an end user since it yields more accurate and logical answers. However, low perplexity also means higher predictability, which is something that detectors look for in a text. Here’s a good visual representation of this concept:
Burstiness
I have a fun exercise for you. Ask yourself something that you’re familiar with. It could be about your favorite band, the rings of Saturn, or what constitutes a human being. Ask yourself, then write down the answer.
Now, ask ChatGPT the same question and look at the difference.
Chances are, you’d notice the monotony of ChatGPT’s responses. It’s well-written, yes, but it doesn’t sound quite like a person. It’s not how you answered the question. It lacks passion. It lacks chaos. It lacks burstiness.
If you’re a writer, you might’ve already seen this popular passage from Gary Provost:
This is a good example of what differentiates AI writing from human writing. The former doesn’t have a rhythm. The sentences are almost always the same length, the same pattern, and the same words. AI detectors take note of this when they’re determining which are AI and which are not.
Redundancy and Coherence
The reason why I personally don’t use ChatGPT to write for me is redundancy. It doesn’t usually manifest in repeated words or phrases, but rather in redundant concepts. For example, if you ask ChatGPT to write an essay with a long word length, you’ll notice that it’s just repeating the same thought over and over again, just in a different way.
This is a symptom of low coherence. LLMs don’t actually respond to your question; they just output a string of words determined by an algorithm. There’s no logic or reason. It’s just probability.
AI hallucinations exist for this very reason. It’s why ChatGPT, Bard, and Claude sometimes give you incorrect information. This is something that AI detectors check using their own database.
Feedback Loop
Some detectors also use reinforcement learning to improve over time. With this approach, the system asks end users to evaluate its predictions. Receiving this feedback powers the AI model to get better at differentiating between AI-generated and human-written text. Over time, this leads to fewer false positives and better accuracy.
What Makes Your ChatGPT Text More Detectable?
Here are common AI identifiers present in most ChatGPT text:
- Transition Words. Words like “additionally” or “furthermore” are overused by GPT models.
- Lack of Sentence Variation. This goes back to our topic about burstiness a while ago. Most AI texts have the same sentence length and structure. One example of this is starting every paragraph with verbs.
- Generic Language. There are words that ChatGPT uses a lot. Some examples include revolutionize, delve, ensure, and many more.
- Absence of Emotion. More than anything, ChatGPT lacks the emotion and passion behind human writing.
If you’re looking to avoid AI detection, be mindful of these identifiers and remove them from your work. And I can’t stress this enough: don’t actually cheat. If you’re using ChatGPT for your assignments, always add your own thoughts into the mix. This would help you better in the long run.
AI Bypassers: Tools To Make ChatGPT Undetectable
If you have the money but lack the time, you can also consider investing in AI bypassers. Here are some of our favorites:
Undetectable AI
Ask me what my favorite AI bypasser is, and the answer will always be the same: Undetectable AI. Not only does it have a robust AI rewriter, it also has output customization features that ensures your tweaked ChatGPT writing doesn’t steer too far away from its original use case and readability level.
Undetectable AI has a different pricing structure if you’re paying monthly or annually. It starts at $9.99 a month for 10,000 words and scales to $209 a month for 380,000 words. If you’re paying annually, the price starts at $5 a month for 10,000 words and scales to $104.42 for 380,000 words.
We’ve tested Undetectable AI a lot of times before, and let me tell you, it’s really effective. Read more about this bypasser here.
HideMyAI
Next on our list is HideMyAI. Like Undetectable, it also has output parameters for use case and readability level. It also adds multilingual rewriting to its feature set, being capable of tweaking text in English, Spanish, Russian, and French.
For monthly payments, HideMyAI starts at $8.25 per month for 10,000 words and goes to $145.20 per month for 200,000 words. The annual option begins at $5 per month for 10,000 words and stops at $88 per month for 200,000.
If you’re interested, here’s our full review of HideMyAI, along with a comparison to Undetectable AI.
Netus
You can also consider getting NetusAI. This AI bypasser may lack the output customization that other bypassers offer, but it more than makes up for it with a robust system that’s almost as good as Undetectable AI.
Netus AI doesn’t charge monthly or annually. Instead, they use a credit-based system wherein one credit is equal to 10 or so words.
- $19: 10,000 credits.
- $39: 30,000 credits.
- $99: 100,000 credits.
GPTinf
GPTinf is a recent discovery of mine, and I’ve been loving it so far. Unfortunately, it doesn’t offer any feature other than a freeze list which allows you to pick parts of your input that won’t be modified. However, its effectiveness earns it a spot on this list.
GPTinf’s monthly cost starts at $12 for 10,000 words and reaches $349 for 500,000 words. If you choose to pay yearly, the monthly cost shrinks down to $8 for 10,000 words up to $233 for 500,000 words.
The Bottom Line
Let’s go back to the original question: How do you make ChatGPT writing undetectable?
Try to use these tools and scramble your writing. Spend extra time revising what they’ve outputted to you, and you’ll be on your way to trying to scramble AI detectors.
Remember, this isn’t 100% conclusive. Nothing is. The way AI detectors work is fundamentally flawed. To use them to predict AI is completely fine, but when used to determine if someone used AI, that’s another scenario.
Best of luck!
