How Do AI Detectors Work and Are They Reliable?

‍First, we create AI. And then we build detectors to catch it.

‍Sure, human nature is complicated, and we’ve long given up on understanding that. But with AI scanners, things are a little simpler.

‍That’s why today, we’ll cover all the endless questions about them you might have had:

• Can you trust these checkers?
• Why do some people get false positives while others seem to have very accurate results with them?
• And, of course, how, in the world, do these AI detectors work?

‍If you’ve been wondering about any of these, welcome…

‍Likely, explaining what an AI checker is would be quite useless. We all understand that those are tools designed to check if the content was generated by artificial intelligence or written by a human author.

‍So, let’s get right into the how. How do those AI scanners work?

‍Some sources credit the first launch of a tool like this to Originality in 2022.

‍Source: AIThority

‍The early models relied on spotting simple text generation patterns. They were mostly looking for matching keywords and strings, comparing them with a large database of other published content.

‍But they have gotten much more advanced these days, relying on more informative metrics, like N-gram, burstiness, perplexity, and linguistic patterns.

‍So, say you want to check if your writer used any gen tools for the blog post. How will it go from here?

‍Basically, the scanner will:

• Analyze the style, tone, and structure of the content to spot AI.
• Compare the text across everything it has in its large database.
• Look for predictable patterns.

‍Kind of ironic to have an artificial intelligence “detect” artificial intelligence, but here we are.

‍Anyway, how do AI checkers work?

‍We’ve already mentioned a bunch of unfamiliar (for many) words: N-gram, burstiness, perplexity, and linguistic patterns. These are the text analysis metrics (aka NLP features) that help modern detectors understand whether they deal with human or algorithm-produced copy.

‍But what are these exactly?

N-Gram Analysis

‍N-gram analysis is the backbone of most detection systems.

‍This one works by checking the word sequences and patterns. So, you know how AI likes to repeat certain words and phrases like “delve deeper into,” “a multifaceted approach,” “realm,” “boast,” and so on?

‍N-gram picks up on these repetitive phrases and uses them to spot computer-generated text. Still, now, this method is more updated and sophisticated than just pattern recognition.

‍So, instead of just flagging a potential robotic word, it looks at the different N-grams (word combinations) within the entire content and analyzes phrase patterns in your context.

‍Source: Funnel.io

‍But surprisingly, soon, this might not be the best way to determine artificial copy, as recent research shows that people start using more GPT’s words in their spoken language.

‍Source: Scientific American

‍Basically, researchers took over a million podcasts and YouTube videos from before and after the release of ChatGPT. And they compared the use of those typical AI words with “synthetic controls” (synonyms typically not used by GPT that often).

‍And guess what?

‍People started using those previously “weird words” much more often (across most categories):

‍Source: Scientific American

Burstiness

‍Human content generally has different variations in structure and length.

‍That’s why, mostly, when you read human copy, you get a mix of short and long, simple and complex sentences. You also see that burstiness (aka variation) in vocabulary, punctuation, and the overall writing style.

‍And all that variety is what makes texts unique and allows you to get so immersed in your reading that you lose track of time.

‍Compare that with content written by AI, which tends to be more uniform. This might sound great in theory, but it just doesn’t generally have the same dynamic flow as a human text.

‍Apart from being one of the ways to spot generated text, it’s also a great technique and practice for all writers.

‍This is how Gary Provost (American writer and writing instructor) once put it:

‍Source: Writing Cooperative

‍This means that the scanners just have to look at that unnatural uniformity to understand that something is off. And that’s another factor that helps see a generated copy.

Perplexity

‍Another metric the AI checkers use to detect AI is perplexity.

‍Basically, it measures the level of predictability in your content. That means it tries to literally predict the word that comes next based on the phrases it has seen so far.

‍Chances are that your blog posts, written and edited by you, will be less predictable compared to those you write with a text generator tool.

‍That’s because you have your own style and voice. You probably have certain phrases, slang, expressions, and so on that make up your own personal style.

‍Think about it.

‍You rarely know how people around you will finish their sentences or how they will talk about even the simplest things.

‍So, measuring predictability makes sense when it comes to detecting artificial copy.

‍Now, let’s check one example to understand how it works. It’s simplified, but it will give you a better idea.

‍Source: Medium

‍In this sentence, the language model is trying to predict what the next word could be based on all the previous words and their order. So, if you were to add “ice cream” to the sentence, it would be less predictable for any scanner, making your copy more human.

‍Again, it is very simplified. And in reality, your AI score won’t depend on one word. But this is just a general idea of what perplexity really is.

Linguistic Patterns

‍Most AI detection software works by paying attention to several linguistic patterns to figure out who wrote the copy, a human or a program.

‍First of all, tech-produced writing is often more formal and quite too perfect, like in the example below:

‍Source: Upper Elementary Snapshots

‍Of course, you can ask your gen software to be more informal, but it will still sound mostly robotic. Why?

• The grammar might be too polished, even though no one talks like that in real life.
• The structure might seem way too repetitive and lack any depth.
• Even if you ask artificial intelligence to be funny, in most cases, it will be pathetic (no offence, but come on…).

‍If you go on and generate dozens of blog posts with ChatGPT or any other tool, you’ll definitely see how similar they are. You’ll see those similarities, even if you use different topics and ask for different tones.

‍Often, it’s this tiny thing you can’t even define, but it just smells artificial.

‍And well, sometimes, it’s more obvious than that.

‍What makes that generated copy sound like that are those linguistic patterns. And that’s why they can help scanners spot non-human content pretty easily.

‍Now that you know different text analysis metrics, you understand that AI detectors don’t just spot AI based on some random patterns. They have several quite complex mechanisms that help them do that.

‍Besides, keep in mind that these checkers are trained on a large database of both human and generated content. That’s why it’s much easier for them to see those clear features.

‍If you go to any detector’s website, each will claim extreme (almost perfect) precision.

‍Copyleaks says they have “over 99% accuracy.”

‍Source: Copyleaks

‍Now, guess what accuracy Originality supposedly has? Exactly!

‍Source: Originality

‍And Winston claims to have a “99.98% accuracy rate.”

‍Source: Winston

‍Cool. But wait a moment. Why do we get all those false positives, then? Why does generated content sometimes pass the scanners while the human is flagged as AI? Is this software reliable after all?

‍These are really legit questions.

‍Let’s answer them one by one. First, can you really trust them? Are AI scanners reliable? Nothing can answer this question better than a real-world example.

‍So, let’s take this blog post by Semrush, for instance, and run it through a few of the popular checkers like Originality, Winston, Copyleaks, and Grammarly to see what we get.

‍According to the Originality, the text is 100% human-written. This is, obviously, a great score.

‍Source: Originality

‍What if we run it through the Winston next? Would we get the same result? Well, it’s almost the same with just one point difference, which looks good so far.

‍Source: Winston

‍But what happens when we also do the same with Copyleaks?

‍Shockingly, this tool claims that the same blog post is 100% AI-generated. And that’s without any edits done whatsoever to the original article that was tested in the previous programs.

‍Source: Copyleaks

‍Finally, Grammarly claims the text is only 7% AI-generated and 93% human, which is still within the high range, signaling that it’s likely not robotic.

‍Source: Grammarly

‍Phew! What a ride! So, do AI detectors work? Yes.

‍But it’s clear that the results can be complex. So, you are likely better off using the paid programs to detect AI writing, as these tend to be more accurate and stay within the same range. Plus, always double-check with another solution.

‍But why can one blog post give these polar opposite results (Originality – 100% human; Copyleaks – 100% AI)?

‍To be fair, we just found an interesting example, and in most cases, the difference won’t be that drastic.

‍Yet, still, why?!

Why Can You Get False Positives?

‍So, let’s face it: sometimes, checkers perform below expectations. And running the same text through different software could give you widely varying results.

‍There are a couple of reasons why this happens:

1. First of all, people can also use those weird wordies (like “delve”) and write using formal structures, especially if they’re just out of college. And these can trigger n-gram, language patterns, or burstiness.
2. Besides, if a writer doesn’t really research the topic deeply, they will inevitably have issues with perplexity because they will create superficial and predictable copy.
3. And let’s not forget that the scores you get are a probability. For example, a score of 90% and above means that your text is likely written by a human, but a score that is less than 30-40% is most likely an AI creation (at least partially).

‍So, there are several factors that can influence your score. But you, as a writer, can also make sure that your copy doesn’t sound robotic. We’ll explain how further below.

‍Short answer? Yes. There are two most popular ways people use to bypass detection when using artificial writing.

‍ShortNote: We don’t recommend using non-human copy on your website, as it generally lacks deep meaning and lowers your chances of both ranking well and converting. If anyone can generate your content in seconds with ChatGPT, what value does it have?

#1 Humanize the Text Manually

‍This is the option that requires more time, but can actually help you make your posts at least a bit more valuable. Writers who use this approach typically do the following:

• Change the structure so it’s a mix of long and short sentences.
• Add their own personal experiences, stories, examples, and some unique insights.
• Include images and videos to make the content more engaging and relevant.
• Improve the tone and replace some too formal words that sound off.

‍Source: X

#2 Use Paraphrasing Tools

‍The second option of using a paraphrasing tool is obviously faster and less complicated. Yet, it doesn’t really help you make your copy more valuable, as it simply spins an already superficial generated text.

‍These solutions aren’t new. Even before this whole AI thing, there were paraphrasing tools that helped you rephrase content to pass the plagiarism check.

‍Yet, if you ever used them, you understand how bad they were. Most of the stuff they rephrased was pure nonsense.

‍Did they get better, though? Let’s check with a real-life example.

‍So, using ChatGPT, we quickly “created” a blog post on how to make your website SEO-friendly, and this is what we got:

‍Source: ChatGPT

‍Obviously, it was flagged as 100% generated text.

‍Source: Originality

‍Now, let’s try to humanize that GPT’s blog post.

‍Source: Grammarly

‍The copy definitely looks a bit better. There is more flow to it, and not just those abrupt bullet points, ChatGPT loves so much.

‍Still, based on the NLP features we’ve seen above (n-gram, perplexity, burtiness, and linguistic patterns), it seems like the article still feels very AI.

‍But let’s check it.

‍Source: Originality

‍Well, not only is the AI score still 100%, but the plagiarism has actually increased from 16% in the initial scan to 36% for the humanized article.

‍And let’s not forget a couple of huge details:

• When you “humanize” texts with no value, no humanizer will add that usefulness. These tools simply work with the original input.
• You can still get a plagiarized copy, whether you humanize it or not.
• The ownership of your content is still an open question, as you’re not the author, while any gen tool uses materials that are already out there online.

‍So, well, yes, you can fool detectors.

‍But will it help you make your SEO and content marketing successful? Hardly.

‍We’ve already touched on false positives. But if you go through the forums, you’ll see that it’s a real issue for many writers out there.

‍Source: Reddit

‍So, let’s go through each possible factor that can lead to this to help you improve the quality of your copy.

‍First, you have to understand one thing. The way AI detectors work has nothing to do with understanding your writing and its depth. In fact, they have no clue what your writing is about.

‍What they check is patterns (metrics) that we’ve already seen above.

‍So, sometimes the articles you spent hours carefully crafting can be flagged as AI-generated because it has one or several of the following characteristics:

Artificial Structure and Formula

‍You already know what makes AI text detectable. Sometimes, if you follow a structure that has a predictable pattern, you will get a false positive.

‍The same thing happens when you create template-based content. Yes, there are many scenarios where marketers need templates for one reason or another. Still, it’s important to know how to fix your writing so it doesn’t get flagged (more on this later).

‍For example, an SEO blog post with lots of lists, bullet points, or short sections usually follows predictable structures that can trigger false positives.

‍It’s even worse when the sentences are short as well and look very similar to each other.

‍Take a look at this Hostinger article, for example:

‍Source: Hostinger

‍This is a well-written, basic how-to guide for people looking to learn how to create SEO content that performs well. But it also has a structure that triggers checkers with short paragraphs that look similar in form:

‍Source: Hostinger

‍Besides, it might be a bit too basic, without many in-depth explanations. This is absolutely normal for a beginner-friendly guide. But scanners often have an issue with this.

‍That’s why when checked with Originality, we get a result like this:

‍Source: Originality

Academic or Technical style

‍Formal texts tend to trigger AI checkers the most.

‍Most of the time, this is academic and highly technical content. So, it isn’t just essays. It’s also technical documentation, user guides, some detailed step-by-step tutorials, instructions, etc.

‍Take a look at this guide by Apple:

‍Source: Apple

‍And this is what the detector has to say about it:

‍Source: Originality AI

‍This type of copy often triggers checkers because the language is stiff, the style is formal, and there is a lot of uniformity within the structure that looks a bit like robotic content.

Non-Native English Writers

‍Non-native English speakers and writers tend to use repetitive phrases or sentence structures that are quite predictable, and this is easily mistaken for AI writing.

‍For example, most native speakers hardly use those essay-like phrases (“In today’s digital world…,” “shed light on,” “pose a significant challenge,” etc).

‍Native speakers are far more likely to skip some words or use simpler alternatives. And overall, their language, flow, and style will be much more straightforward and conversational compared to non-native speakers.

‍So, if you are a non-native writer getting a lot of false positives, the issue could be your writing style. And the best thing you can do is try to avoid those typical phrases you learned in your English book.

‍Source: English Classes Online

‍While these phrases are much needed to pass your language test, they’re useless when writing anything marketing-related. Or anything you want people to actually read because they want to, and not because they have to grade it.

Document Editing Tools

‍You already know that AI is changing the consumer market, but it is also doing the same to the digital marketing scene.

‍Think of tools like Grammarly and other software designed to help you catch grammatical errors and polish your writing. The idea is to help marketers make their messages clearer and more impactful.

‍They are wonderful, but they also often highlight that easy, colloquial language and suggest more formal replacements.

‍For example, here, in one of our previous sentences, it suggests replacing “people looking” with “individuals seeking.”

‍Now, you tell us – what sounds less human?

‍And guess what? Most of those suggestions are also AI-based. So, of course, those “sterile” alternatives could be flagged as a generated copy.

‍That’s why many scanners make a similar side note:

‍Source: Copyleaks

‍At any rate, it doesn’t matter what the reason is behind those confusing results. Most AI detectors depend on machine learning (ML) and natural language processing (NLP).

‍ML and NLP help checkers see the linguistic nuances and the structural patterns of the text.

‍This is why, if you are using these scanners, it’s important to accept that the software can misidentify writing as AI-generated. So, if possible, try to:

• Use more than one tool.
• Pay more attention to the style of artificial copy so you can spot the difference.
• Know your writer’s style so you can recognize when it’s not their work (if you want to double-check your outsourced or in-house copywriters).

‍Okay, now you know much more about the detection process and finally have an explanation for those annoying false positives you’ve been getting.

‍So, the only thing left for us is to understand how you could make your writing as human as possible, whether you’re doing it for detectors, search engines, or your readers.

‍Once you understand how AI writing is detected, all the following tips will seem pretty logical to you. Still, let’s summarize everything you can do to make your copy better:

Find Your Own Voice

‍One of the reasons people turn to artificial intelligence is the fact that it can replicate any style.

‍Still, let’s face it, it can hardly create one – at least not one that is both original and truly interesting.

‍So, if you really want to stand out from the competition, the best thing is to find your own voice. Think of any brand you admire. Chances are that they have a distinct personality. And most likely, you can see it through every post they publish.

‍For example, check out dbrand. It’s one of the boldest labels out there. Their copy is fun, edgy, and sometimes a bit rude (or maybe even inappropriate to some).

‍Source: dbrand

‍And you can see this voice through their communications, whether it’s their Instagram, where they ask “to give them money immediately.”

‍Source: Instagram

‍Or their website, where they sell an X-ray phone case and mention that “medical debt isn’t included.”

‍Source: dbrand

‍Sure, now, you could “feed” any gen tool with their messaging and ask it to create something similar, even for their new launches.

‍But if they were to create their brand messaging together with AI from the start, do you really think that ChatGPT would suggest this in-your-face style for a tech accessory brand?

‍After all, brand voice is a big deal, so maybe you don’t really want to give the algorithm the responsibility of creating one for you.

‍Besides, AI software basically just mixes up what others have already done on the internet. Of course, human creativity is also about working with already existing concepts and packaging them unexpectedly, but for now, no tool can really do it at the same level.

‍That’s why, at best, it will come up with a copy of someone else’s voice, which can create an entirely new issue on its own.

Avoid Words That Trigger AI

‍Alright, it’s a straightforward tip, really. Do your best to avoid phrases that scream, “I’m robotic as hell.”

‍If you aren’t sure what those could be, you can check GPTZero’s AI Vocabulary with the 100 most popular words artificial intelligence uses while humans rarely do:

‍Source: GPTZero

‍In reality, when you use this strategy, you do more than simply avoid AI detection. You make your texts more understandable and simple. Because those words aren’t often used by human writers for a good reason.

‍You want to simplify your sentences as much as possible to create a more casual, natural flow. And, of course, avoid using repetitive phrases.

Share Your Experience

‍Look. We all understand that today, anyone can create any content on any subject in a couple of seconds. So, if your strategy was always to post as much as possible without really caring much about quality, you’ll have to tweak it.

‍Creating a content farm now is useless. You have ChatGPT for this.

‍That’s why the only option these days is to share something that is so personal and special that no one would even think that it’s generated.

‍How can you do that?

• Share your own experience. We all know that theoretical knowledge is great in college, but in real life, things often work differently. So, show that you’re an expert by adding your own real-life knowledge.
• Give clear examples and support your points with data. Again, the generated copy is mostly very superficial. So, make your content research-based and show how people can apply that with examples.
• Tell stories. Storytelling has been one of the most popular and effective marketing tactics for decades (if not longer). We all love stories, so add those to your writing. You know that most big brands use storytelling. And it seems like this strategy never gets old.

‍Source: YouTube

‍As much as possible, you also want to add images, memes, light humor, etc., to make your copy more like a normal, human conversation. How would you talk to your friends, family, or even colleagues?

‍We bet you wouldn’t sound like “in today’s digital world.” 🙂

Interview People

‍The worst thing about AI content is the fact that it’s all fluff – no meat and no substance.

‍So, these days, your readers much prefer your content to give them real insights that they can apply in real life to solve their very real problems.

‍Think about it.

‍Who wants to take driving lessons from someone who has read tons of books but never gotten behind the wheel? Yeah, definitely not us. It’s the same thing with expecting people to take your generated copy as a solution.

‍Artificial intelligence has never had any experience, and will never have.

‍So, you can use this to your advantage. Give your readers clear guidance and be fact-based. One way of doing this is by interviewing people to get their expert knowledge and experiences.

‍Here is an example of how TypeForm and MailChimp teamed up to give even more value to their users.

‍Source: Typeform

‍So, if you never did anything similar, it might be a great time to try.

‍We don’t really know why you were interested in how AI detection tools work.

‍If you were just curious how to improve your writing and make sure it doesn’t give off that robotic impression, good for you.

‍If you simply wanted to know how to bypass those scanners and get away with low-quality content, well, no judging.

‍But in reality, looking at the development of artificial intelligence, it seems like very soon, any detection will be pointless.

‍Algorithms get better, and it’s only a matter of time until they fool any scanner.

‍But they will never fool people.

‍Because it looks like more and more users will only read articles that actually make sense. That have clear value. And offer real solutions.