AI is nothing new—researchers began conceptualizing the technology underpinning AI almost a hundred years ago. But to say that this technology has seen a vast evolution over only a few years would be selling it short. Today, AI has become nothing short of ubiquitous.
Since May 2024, ChatGPT alone has had over 200 million active users, as first reported by Axios, and the technology saw one million users in the first five days of its 2022 launch. It didn’t take long for people to start using ChatGPT for a number of tasks, from research to coding—to just having a conversation with something that makes for a pretty convincing person, as artificial as it is.
Thanks in part to ChatGPT, AI is now all over the internet, in our phones, and making its way into countless smart devices. It’s hard to visit any site, buy any device, or open any app without being inundated with AI in one way or another (and for better or worse).
However, when it first began to make waves around late 2021 and early 2022, AI wasn’t anything you necessarily needed a detector to, well, detect. Where text is concerned, nothing written by a machine could really pass the sniff test, so to speak. Just reading it was enough to tell that it had been machine-generated, with plenty of errors and awkward grammar.
Times, however, have changed.
Today, AI-generated text—not to mention images, video, and audio—is quickly becoming indistinguishable from words written by real people. This rapid technological advancement in all things AI has more than necessitated the need for reliable detectors, but there’s a big difference between an AI detector and an accurate one. So, are the 2024 iterations of these AI tools accurate? Yes, but be sure to read on for more specifics as the why and how are also important.
Why Are AI Detectors Important?
With as pervasive as AI-generated content has become, the need for the ability to determine the authenticity of a written piece of content has only grown. In many ways, trying to pass off AI-generated content as your own is synonymous with plagiarism. You are, in essence, taking what the machine writes and passing it off as your own words.
Outside of that, AI has far-reaching implications for academic dishonesty. Say what you will, but a certain subset of students of all ages throughout the years have tried to cheat their way through school.
First, it was essay-writing services and simple plagiarism—and tools arose to help combat those types of academic dishonesty. Now, would-be cheaters don’t really even have to put forth that much effort to cheat. They can just hit up a few AI chatbots to write them a full-fledged essay in a matter of seconds.
Just a few years ago, this essay wouldn’t have passed the sniff test, considering how painfully obvious AI-generated content was back then. Today, however, all a student needs to do is tweak a paragraph here or there, and in not much time, this essay would go from 100% AI-written to 50–75%. Without the help of an AI detector, it can be exceedingly hard to tell when one of these essays is AI-written.
Similarly, in the ever-competitive struggle to maintain that number-one spot in Google search results, many sites are churning out AI-generated content. While a lot of these articles are still helpful—in their own way—to truly stand out as an expert in a field, a writer must create content that rises above the sea of low-effort AI-generated blandness. Only then will the site hosting the content begin driving organic engagement with its truly well-written and enjoyable content.
Granted, as great as AI detectors have become, they’re still prone to false positives, although the best detectors have also strived to iron out such kinks. And to be fair, if you’re writing an article or essay and a tool out there thinks it’s written by AI when it’s not, that can also be a great indicator that maybe you should add a bit more life to your writing.
Image credit: GPTZero.me.
How AI Detectors Work
We explore the inner workings of AI detectors in a bit more detail in a separate article, but for the time being, let’s have a quick primer on their functioning. Ironically, and probably unsurprisingly to many readers, AI detectors are built similarly to how AI tools are created, leveraging the power of machine learning and language models.
A detector’s number one goal is to tell the difference between content written by a living, breathing person and content created by AI. So, naturally, it has to learn to spot the differences between each type of content. For that, the models must be trained on as much data as the developers of that tool can get their hands on.
The Difference Between AI-Generated and “Real” Writing
At the core of how AI detectors work is an understanding of what separates man from machine—at least where words are concerned.
At the end of the day, ChatGPT is trying to write sentences and paragraphs that are easy to understand, and at the basis of this are two hallmarks of writing: perplexity and burstiness. We pesky humans tend to write with a little more perplexity, and oftentimes, it’s a little harder to understand what we’re saying than what AI is saying. After all, at its core, ChatGPT is just trying to figure out what word comes next whenever you ask it a question.
Image credit: scribbr.co.uk.
Similarly, humans tend to write sentences and paragraphs of varying lengths and complexity. Some sentences will be just a few words long, while others will be a couple dozen, with a structure that resembles a roller coaster. This is burstiness. Content written by a human will be perplexing (in all the right ways) and bursting (but not in the wrong ways).
Comparatively, AI content lacks perplexity and burstiness, no matter how many times you try to instruct ChatGPT on how to write more naturally and realistically. AI sentences and paragraphs read well, but by and large, they’re boring compared to what humans (especially talented wordsmiths) are able to accomplish.
To sum up, AI detectors work by being trained to tell the difference between the perplexing nature of human writing and the more structured content that AI generates.
So, How Accurate Are AI Detectors in 2024?
That brings us to the point of this article: How accurate are AI detectors in 2024, given the rapid advancements in AI? Especially considering how scarily realistic text, images, and audio have become, it’s starting to look like a proverbial arms race out there.
As it stands in 2024, AI detectors are surprisingly accurate. Perhaps the most impressive aspect of their accuracy is that they’ve made some seemingly dramatic improvements over the past year alone. If you’d used any of these tools just a few months ago, there was a good chance that a few of them would have given you a false positive—detecting AI when the writing was 100% human.
Today, however, tools like GPTZero, ZeroGPT, and Originality.ai have become almost foolproof, thanks to diligent developers updating the AI algorithms powering these models, as well as feeding them more and more examples of AI and human writing.
Image credit: Zdnet.com
Though it wasn’t included in this ZDNet examination, we can also attest to the accuracy of AI Detector, which has been meticulously crafted to detect a wide range of AI tools, from Claude to ChatGPT, with over 95% accuracy.
Despite how great AI detectors have gotten, it’s also important to remember that no tool is 100% infallible. So, if you’re worried about whether a piece of writing has been AI-generated, getting a second or third opinion is also a good idea. And as AI models continue to evolve—ChatGPT 5 is on the horizon—the detectors, too, will need to up their game to maintain their high levels of accuracy. So, with each new AI model release, it’s an excellent idea to check whether the detection tool you’re using has been updated recently.
What Detracts From AI Detector Accuracy
At the end of the day, as accurate as AI detectors have become, there are still some considerations to keep in mind before diving into the technology—some of the things that can impact accuracy in the most noticeable ways.
False Positives (and Negatives)
No AI detector is completely foolproof. For starters, no two writers are the same. I would wager that somewhere out there is at least one person who just happens to write similarly to that of a machine—not to throw any shade. Perhaps the rise of AI-generated articles, at the very least, has spurred this one person to write a bit differently. But if not, then there’s a good chance AI detectors would have difficulty determining the authenticity of this person’s writing.
AI detectors have also shown a bias against non-native English speakers. One study in the journal Patterns showed that TOEFL (Test of English as a Foreign Language) essays written by Chinese students showed an average false-positive rate of 61%. It turns out that students who are just learning English showed near-AI levels of perplexity and burstiness—similar to that of AI writing tools.
Image credit: cell.com.
Fortunately, however, there’s light at the end of this tunnel, so to speak. Another study in Science Direct proved that you could create a detector that wasn’t biased, even if most of the publicly available ones still struggle.
The opposite can also be an issue—false negatives—when AI detectors mistakenly let AI-generated content slip through the cracks. This can occur just by happenstance; maybe ChatGPT just really nailed that one essay. But it can also happen when the person goes through the efforts of specifically instructing the AI chatbot on how exactly to write. Many AI detectors will still catch these efforts—but not all. Similarly, people can also use AI humanizers to rewrite machine-written content so that it appears more like it was written by a real person.
The Fast-Paced Nature of AI
As we touched on just a little while ago, ChatGPT 5 is on the horizon. More recently, OpenAI (the company that developed ChatGPT) rolled out 4o, an updated version of its previous model. Even more recently, Anthropic updated its AI model, Claude, to version 3.5. Each new version improves the output’s quality, making it more natural, with more perplexity and burstiness—the two traits that separate “real” content from AI content.
This rapid advancement only throws a wrench into the gears of AI detectors. In many ways, it really is a veritable arms race. With each advancement, detection tools must play catch-up in order to do their job. By this logic, AI detectors will be their least accurate just after a new AI model has been launched, but of course, the best tools out there will be able to adapt quite quickly to this new paradigm.
It’s All About the Data
In a similar vein, the better the data an AI detector is trained on, the more accurate its results will be. For example, if the model is only trained on two-year-old AI-written articles, it’ll have a harder time determining if more recent articles are real.
On the other hand, it’s also vital for an AI model to be trained on a wide variety of articles written by real humans—and not just a small subset or even just one person. No two people write exactly the same, so if the model really only knows how one person writes, it’ll be harder for it to detect if anyone else’s writing is AI-generated.
Who Can Benefit the Most From AI Detectors?
At the end of the day, anyone can benefit from being able to tell whether a written piece of content has been generated by AI. However, as previously indicated, AI detectors are highly useful in the education industry, especially since modern AI chatbots have gotten so good that often, teachers and students have a hard time telling the real assignments from the AI-generated ones, especially if they’re not familiar with a student’s writing style—or if the student is non-native.
Diligent students can also use these tools to check their own work to ensure that they don’t get unexpectedly slapped with a false positive, despite not having used an AI-writing tool. From there, the student can revise the trouble points. Though, with the current advancements in AI, these sorts of false positive instances will probably be rare—but they do still happen. That’s why, although AI detectors can be of great benefit to both teachers and students, these tools should not be the end-all-be-all.
Outside the educational sector, with the rampant rise of misinformation and disinformation all over the internet, it can behoove most people to double-check whether what they’re reading was written by a real person or generated by AI as a quick means of spreading bad information. That doesn’t prevent real people from spreading lies, but it can certainly help separate some fact from fiction.
Copywriters, too, can benefit from AI detection tools, primarily as a means of protecting themselves against false accusations. With as much AI-generated content as there is on the internet, knowing that what you’ve written won’t be mistakenly called out can bring great peace of mind. It’ll also help ensure that what you’ve written will rise above the sea of content written by chatbots. Similarly, websites and publishers can ensure that the content they’re receiving from writers is authentic.
Ultimately, as accurate as AI detectors currently are, their applications extend far and wide to virtually any form of written word.
The Bottom Line
In many ways, the accuracy of AI detectors is something akin to playing Whac-A-Mole as AI models are forever evolving and advancing. However, as it stands in 2024, the good news is that the best detectors are currently surprisingly accurate and able to tell almost without fail whether something has been AI-generated.
The bad news, of course, is that it’s only a matter of time before something new and shiny—like ChatGPT 5—comes around. These new models will undoubtedly be able to write better and more naturally, and of course, AI detectors will have to adapt to continue detecting as well as they do today.
Fear not, though; AI detectors have grown quite accustomed to this game of cat and mouse, with many new models and updates to existing models coming out over the past two years to counter the stream of updates to AI content generators.
So, we can rest assured that the best AI detectors, like AI Detector, will adapt and continue identifying AI writing with the impressive accuracy that they have had throughout 2024.
