After putting earnest effort into writing an assignment for school or a post for your own blog or social media account—or content for a client—the last thing you want is to run it through one of the many AI detectors only to discover that what you’ve written bears a striking similarity to AI-generated content.
When this happens, it’s called an AI detection false positive. In other words, it’s when you know with reasonable certainty that some content has been written by a living, breathing person but is believed to have been created by an AI chatbot, like Claude and ChatGPT.
Though anyone can be hit with a false positive, perhaps students feel these ramifications the most. False positives can lead to failing assignments and being accused of academic dishonesty. For everyone else, having any doubt cast on your hard work is never a good feeling and may impact your reputation and ability to land new jobs if you’re a professional writer.
Naturally, the best way to avoid a false positive is to fully understand how AI tools like AI Detector function and how authentic writing differs from AI writing.
What Is an AI Detection False Positive?
To expand on what we touched on in the introduction, a false positive, as it stands for AI detection, involves the AI model erroneously determining that the text you submitted shows signs of being written by a chatbot like ChatGPT, Claude, Gemini, or Copilot.
Typically, AI models have a distinctive way of writing that’s easy for even eagle-eyed readers to pick up on without needing an AI detector. But AI detectors can be better than humans at identifying AI-generated content. These tools, after all, are built in much the same way that the AI chatbots themselves are built, leveraging the power of neural networks and machine learning.
The end result is a program that’s trained on what AI writing looks like while being able to contrast that with what “real” writing looks like. In a perfect world, these tools would be able to tell the difference each and every time. Unfortunately, they can’t. Though increasingly uncommon—thanks to the advancements in AI—false positives still happen.
It’s also worth briefly mentioning false negatives, which is when AI detectors suggest that machine-generated text has been written by a human. Though not the focus of this article, such instances can indicate when an AI detector that you’re using is most likely outdated. To truly tell how reliable any of these tools are, it can be great to first run something you know to be AI-written through them and ensure that they detect it.
To properly avoid AI false detections, an excellent first step is to ensure you fully understand the nuances that separate real writing from AI-generated writing.
Real Writing vs. AI Writing
A pair of concepts associated with writing are perplexity and burstiness. Perplexity is how random the writing is—in regard to the unpredictability of each word, phrase, and sentence. And burstiness is the overall variance of a document involving sentence structure and length, the fairly random mixture of short and long sentences.
Image credit: GPTZero.
As you might imagine, AI models are low in both perplexity and burstiness. In essence, an AI chatbot uses a method of prediction to make an educated “guess” as to what words or phrases come next in a sentence, depending on what you’ve asked it. Thus, they use a lot of the same words, in the same order, with very little variance in sentence structure and length.
This isn’t a huge concern if you’re just looking at a few sentences or a paragraph, but if you try to analyze a full AI-written article, you can easily spot a lot of these redundancies yourself.
If AI-generated writing is low on burstiness and perplexity, then it would stand to reason that real human writing is much higher in both of those things. And you would be right. Some of that is just the nature of being human. We’re not machines. Some of it is also by design since skilled writers know that a little more perplexity and burstiness makes writing more interesting to read.
AI detectors scan the text to detect the presence or absence of perplexity and burstiness to determine whether the submitted text has likely been written by AI. However, ultimately, each AI detector model can vary slightly in its accuracy depending on which one you use.
Who Will Most Likely Receive False Positives?
Virtually anyone could potentially get flagged for AI-generated writing, and all of us probably have sentences in our writing that bear striking similarities to what a chatbot might generate. However, certain people are far more likely to receive false positives.
Notably, non-native English writers have been proven in a study in Patterns to be more vulnerable to false positives as GPT detectors specifically have shown biases against them.
One study even proved that enhancing the word choice of an essay reduced the instances of false positives, whereas simplifying the text had the opposite effect of making false positives more likely.
Image credit: Cell.
It makes sense if you think about it. Non-native writers have less of a command of the English language, so their content tends to have less of that perplexity and burstiness we expect. When these weaknesses are addressed, it reduces the number of false positives non-native writers receive.
However, another study in Science Direct actually proved that while most publicly available AI detectors tend to give more false positives to non-native writers, tools could be designed that would show far less bias against them. So there’s potentially a light at the end of this tunnel.
A low amount of perplexity and burstiness is not limited to non-native writers. Anyone who’s simply not a great writer could potentially fall victim to false positives.
When the AI Itself Causes False Positives
Outside of situations where the person doing the writing is leaving themselves vulnerable to false positives, sometimes the AI detectors are to blame. Much like the AI chatbots, the detectors are built with a large language model (LLM). We won’t go into too much detail about the inner workings of LLMs and AI here, but feel free to check out more about them if they pique your interest.
The only important thing to note is that LLMs must be trained to perform whatever function you want them to, be it responding accurately and as naturally as possible to your prompts or scanning your documents for signs of AI writing. And this training requires vast quantities of data.
Issues begin to arise when these LLMs aren’t properly trained, either from the onset or as time goes on. The best AI detectors are trained to be effective for all major chatbot models. If a detector is only trained on one or two of the models, issues can arise.
With insufficient training data, AI detectors won’t be able to properly tell the authentic writing from the AI-generated writing, which will lead to inaccurate results. While, typically, this might also make false negatives more likely, when the AI model doesn’t know the types of content that different chatbots can generate, it can also lead to false positives.
Image credit: OpenAI
Similarly, AI models like Claude, ChatGPT, and Gemini are often updated multiple times per year. Each new version can significantly improve how the chatbot responds to your prompts. Naturally, this can make it more difficult for AI detectors that are trained only on older versions to recognize the AI-generated text, potentially causing false positives.
AI detectors will be most accurate when they’re trained on vast amounts of varied data—both AI-written articles and authentic pieces of human-written content. Only then will they be reliable.
How To Avoid False Positives
Although we’ve talked about these points already, here’s a consolidated list of the steps you can take to help prevent false positives:
Vary your writing techniques
Variety is the spice of not only life but also your writing. So, add as much variety as you can without going overboard. Remember to vary your sentence structure, overall sentence length, and word choice. Also, adding personal anecdotes and your own subjective context—where appropriate—is another way to avoid sounding like an AI chatbot.
Add your own voice
AI doesn’t have much in the way of a personality, so be sure to inject yours where it makes sense. Add some humor if possible, use an atypical style if appropriate, and throw in a few opinions for good measure. As for tone, chatbots are understandably robotic and painfully neutral, so be sure to double-check that your tone matches your intended audience.
Incorporate rhetorical questions or metaphors
This is a great tip because chatbots typically avoid both of these devices.
Talk to a chatbot
If you want to avoid sounding like a machine, a great idea is to simply have a conversation with one. Ask it a few questions on a variety of topics, then note how it responds.
With any luck, these tips will help you avoid running into any false positives—and many of these tips will help enhance your writing, which has its own benefits.
What To Do If You’ve Been Hit With a False Positive
If you’ve been unlucky enough to have encountered a false positive, for starters, don’t panic. These things happen, and most of the time, they’re not a big deal. Fortunately, there are a few easy things you can do if you do happen to fall into this situation:
Get a Second Opinion
Although most of the best AI detectors are created quite similarly, they’re not entirely equal. So, if you’ve gotten a false positive from one tool, a great idea is to run your document through a different tool and see if you receive similar results.
Since the AI models themselves, as well as the training data that these models are built with, can all vary from detector to detector, you might very well have different results from detector to detector. So shop around and get a second opinion.
Do Some Editing
Although it might be met with a little chagrin from some writers, you might just need to perform a little revising using the methods we’ve explained earlier in this article. This option can be especially nifty if you’ve found that your submitted text is only slightly suspect and isn’t flagged with 100% AI-generated writing.
Usually, most detectors will flag which sentences that they suspect might be written by AI. So by zeroing in on those sentences and rewriting them to be just a tad more unpredictable, you can usually pass the AI detectors. Sometimes, all it takes is a different word here or a variation in sentence structure there.
Responding to Accusations
While a great rule of thumb is to scan your documents before you submit them to your boss or your teacher—or upload them to your own blog or social media presence—understandably, sometimes this will slip the mind.
So what do you do once you’ve already finished something, sent it to where it needed to go, and then received an accusatory response—maybe a threat of a failing grade? Well, you’re down, for sure, but you’re not quite out.
For starters, you can try to find some evidence that you did, in fact, write the piece of text. Both Google Docs and Microsoft Word maintain the full version history of the file, with timestamps dating back to where you began working on the file.
So, take a few screenshots—or record some footage with a program like Loom—showing that you didn’t just paste 1,000 words from an outside source into the document. If you’ve been working on that file for a few days, you likely have multiple versions available to prove that you are the sole author.
Next to that, if you like to take notes or keep ideas for your writing on a separate file, that can be great to share. Basically, any file that you can share that shows that you spent time working on any given piece of writing can help back you up in the event that you’re slapped with a false positive.
If you do a lot of writing, you can also provide some alternate writing samples to show similarities in your tone, voice, and writing style.
Legal and Ethical Considerations
As commonplace as AI has become, especially in a classroom setting, it’s raised some important considerations about the legal and ethical ramifications of using the software. Plagiarism and academic dishonesty are nothing new, but they’ve become far more accessible with as easy as these AI tools are to use.
That’s why it’s become important for schools and businesses to ensure due process. No institution should rely solely on AI detectors or any other AI tools. They’re not an end-all-be-all option when weighing someone’s guilt over academic dishonesty or plagiarism as there are plenty of other bits of evidence that could prove or disprove any dishonesty in question.
The Bottom Line
At the end of the day, false positives from AI detectors are a rare occurrence, but they will sometimes happen. So there’s always a chance that you could one day run into one yourself.
Fortunately, as we touched on in this article, there are plenty of preventive measures you can take to try to avoid these instances from happening. And if you’ve been unlucky enough to have received an accusation that you submitted AI-generated writing, there are a few things you can do. Ultimately, these kinds of detection tools, like AI Detector, can help give you peace of mind before you submit anything you’ve written, helping you avoid any potential allegations of either academic dishonesty from your teachers or professors or any professional repercussions from your boss or workplace.
