It’s safe to say that ever since ChatGPT was launched in November 2022, the world has changed.
AI is infiltrating nearly every industry and is contributing to the rapid evolution of existing technology.
Using AI-generation tools like ChatGPT to produce text has been a popular way to harness its power.
Whether you are a freelance writer, content creator, or even an academic student, AI has been a powerful tool to add to your arsenal.
Enter AI detectors to spoil the party. We’re joking of course, there does need to be a check in place to ensure that AI is used responsibly and with good intentions.
Site owners, schools, and even Google itself utilize AI detectors to ensure we are not passing that work off as our own.
But can AI detectors be wrong? It should come as no surprise that AI technology on both sides of the aisle is still imperfect.
These detector tools routinely suffer from false positive tests. Here’s how you can learn to bypass AI detectors and avoid the hassle of false positives.
AI detectors are built using natural language models and millions of data points of both AI and human-generated text. Whenever content is screened by an AI detector tool it compares it against these data sets and seeks out predictable patterns in syntax, word choice, and the overall structure of the text.
These detectors are trained to recognize patterns and compare them to both AI and human-generated examples.
The findings of the AI detector are the likelihood that the scanned content is AI-generated, not a guarantee. AI detectors work on the basis of probability without any definitive evidence.
What exactly are these patterns that AI detectors are looking for? Two concepts that are driving forces for AI detectors are what are known as burstiness and perplexity.
Burstiness refers to the length and complexity of sentences. If you’ve ever read AI-generated text, it will sound mechanical. This is because many of the sentences have similar length, structure, and even punctuation.
Perplexity refers to how complex the language is as its name suggests, and how easily human readers will be perplexed. For comparison, AI-generated text is pre-programmed to have a low perplexity to make things easier for readers to understand.
When you scan your content through an AI detector you are provided with an AI detection score.
Most AI detectors will make this score a percentage reading out of 100. Some will provide you with two scores that add up to 100: one being an AI score and one being a human score.
What does an AI detection score mean? It depends on the tool.
Most will provide you with a score of say 85%, which means there is an 85% likelihood that the text was AI-generated. It does not mean that 85% of the text is AI-generated.
Likewise, for dual-AI detection scores, it is the likelihood that it was created by AI vs the likelihood it was created by a human. This will result in a score of 85% AI and 15% human.
Yes, and more often than you might think too. Those who have both used AI detectors and had their content scanned by an AI detector, know that these tools are far from perfect.
Even a small percentage of erroneous scans can lead to false accusations against students or writers.
Remember that AI detectors can only scan for recognizable patterns in the text.
If a human writer has the misfortune of writing like a robot, then there is a high likelihood that it will get flagged. This is what is referred to as a false positive.
A false positive is defined as a positive result that is recorded even when it isn’t the true result.
This can apply to science experiments and even medical tests for illnesses. An AI detection false positive happens when the AI detector scans the content and alleges that it was written by AI even though it was written by a human.
You can imagine how problematic this can be when it comes to things like academic integrity and employment for freelance writers.
Turnitin is an online program that many Universities around the world use to allow students to submit coursework digitally.
This site now has an AI detector available for teachers to use, and it claims an accuracy rate of more than 98% with a lower than 1% rate of false positives. Some users believe that the 98% rate is much lower than the reality.
An AI detection false positive can lead to discipline from schools, employers, and Google itself when it comes to SEO ranking.
This reality is settling in and some companies and schools, like Vanderbilt University, are disabling the AI detector on Turnitin.
So what causes these AI detectors to provide false positives? There are a number of reasons why your content might be flagged.
The AI detector itself does not realize it is providing you with a false positive report. There is either an issue with the training of the detector itself or your content. Here are a few reasons why false positives happen so often in AI detection.
At the heart of AI detectors and most AI software for that matter, is how well they are trained. With ever-evolving models and increasingly complex algorithms, AI detectors can be caught with old technology and out-of-date data sets.
More complex data sets and robust training might better equip AI detectors to eliminate false positives. But this will both raise the cost and computational capacity of these programs.
AI detectors have a strong tendency to flag non-native Engish writing as AI-generated. A research study by Stanford suggests that most AI detectors have a bias against non-native English speakers.
Why? It’s not racism, it’s because most non-native English speakers score a low mark in perplexity. As in, their sentences and prose are generally very simple and easy to understand.
AI detectors also follow very well-defined rules and patterns. The machine-learning algorithms that are used to scan content do not allow AI detectors to understand the content itself.
Instead of seeking factual accuracy or readability, the AI detector scans for syntax and structure. AI detectors will not comprehend any linguistic nuances, a change in tone or even writing that is colloquial or sarcastic, and thus will often flag these as AI-generated.
Being falsely accused of using AI-written content as your own is no joke. It will likely lead to academic discipline for students or loss of employment for writers.
What’s worse is that you have no control over your content being scanned. It is your word against the AI detector’s and unfortunately, far too many people will trust the AI detector over the human. Here are a couple of things you can do if your text is recognized as AI-written.
One of the first things you can do is bring to light the inaccuracy of AI detectors. False positives are a well-documented issue for AI detectors, so it shouldn’t be difficult to find.
A good place to start? ChatGPT’s creator OpenAI even shut down its own AI detector platform because it was returning too many inaccuracies.
You can also prove the originality of your work by bringing up previous versions in word processors like Google Docs or Microsoft Word.
If you know your content will get scanned by an AI detector you can even take drastic measures like screen recording yourself writing the content or documenting it with time-stamped photos.
If it seems like a lot of work to prove that your work is your own, then perhaps it is easier to take steps to avoid AI detection altogether. There are things you can do before you submit your work to mitigate the risks of being flagged as written by AI.
Without a doubt, the least time-consuming way to prevent your work from being flagged as AI is to use a tool like Undetectable.ai. What is Undetectable.ai? It is an AI scrambling tool that re-writes your content into text that will have no issues passing any AI detector.
This tool is designed to bypass popular AI detectors like Copyleaks, GPTZero, and Sapling.
Simply paste your text into the app, select the voice you would like to use, and click the Humanize button. Your output will be close to your original text but now altered to receive a human-written score from any AI detector.
Undetectable.ai saves you time and effort that would otherwise be put into hand-editing your work to avoid AI detection. Plans start at just $9.99 per month for 10,000 words or $60.00 for the whole year.
What is the point of a tool like Undetectable.ai? To edit your work or AI-generated work so that they are not flagged as AI-written. If you don’t want to pay the minimal cost for Undetectable.ai, you can just do the manual editing yourself.
The easiest way to ensure you bypass AI detectors is to use your own unique tone and voice. Another way is to ensure your syntax and sentence structure vary throughout your text. Nothing is a more telling sign of AI-written content than having each sentence sound rhythmically the same.
Of course, once you have done all your edits, you will want to scan your final work one more time for AI detection. The best way to ensure you do not get flagged is to pre-scan the work yourself before submitting it.
Make sure you keep a record of the AI detection score in case a different AI detector returns another score. Whether your text is AI or human-written, it only takes a few seconds to scan your final work for AI detection and potentially save yourself a lot of grief.
So, can AI detectors be wrong? Of course. Are AI detectors flawed? Absolutely. Do AI detectors work? Most of the time they do. It’s not uncommon for people to only think in dualities but an AI detector can both be flawed and serve its purpose.
When an AI detector is used properly it can provide a defense against the unethical usage of AI for academia and content creation. But when AI begins to provide false positives and inaccurately flag text as AI-written, then we have a problem.
We have to remember how early we still are on the generative AI roadmap. The technology we have now will look ancient in five or ten years. As AI technology continues to improve, so too will the accuracy of AI detectors.
Until then, we will have to learn to deal with the fact that these detectors are far from perfect and that the results and AI detection score should never be used as definitive evidence of cheating or academic dishonesty.