Churning out high volumes of content is a good way to bring more traffic to your website.
It can also be a one-way ticket to internet limbo if you’re not careful.
Since the release of ChatGPT, generating content has become virtually effortless. You feed it instructions, hit submit, and wait for the magic to happen.
But is it a feasible strategy to grow your visibility online? As the Google March 2024 Core Update showed us, not really.
In large part due to this update, more and more people started looking for ways to detect AI-generated content. Why? Because Google started clamping down on sites that abused AI tools to flood the web with mediocre articles.
The conclusion many people have arrived at is that avoiding AI-written content is the way to go. As a result of all this, AI content detectors grew in popularity.
Unfortunately, these tools aren’t perfect either.
In this piece, I argue that humans are better at identifying AI content than machines — at least, in ways that truly matter for SEO.
What does it mean? Stay tuned to learn the answer.
Why Do We Care if Content Is AI-Generated?
We care because Google cares. That’s basically it.
But here’s the catch — Google only cares about the quality of content, which is often lackluster when you put content creation on autopilot.
It doesn’t mind the content being created with the help of AI. Big or small — AI contributions don’t matter as long as the end result is good.
So, what constitutes a good end result? Here are a few guidelines worth following:
- Create People-First Content: When you’re focused on tracking your results, it’s easy to lose perspective. Don’t miss the forest for the trees. Remember that you should create content for others to enjoy and benefit from, not just to make the line go up.
- Demonstrate E-E-A-T Qualities: E-E-A-T is a group of signals that Google uses to determine the quality of content. The more right signals you send, the higher your chances of landing a spot on the first page of search results.
- Be Straightforward About the Use of AI: As one of the elements of creating helpful, reliable, people-first content, Google lists self-evaluating your content in terms of “Who, How, and Why.” When you publish your content online, you should make it crystal clear who the author is, how the content was created, and why it was even written in the first place. So, if you used AI during any step of the process, it’s best not to beat around the bush and admit that upfront.
- Double-Check Everything You Publish: While it may sound trivial, having more than one pair of eyes look at your content before posting it online is a good practice. The temptation to skip a few steps and publish your piece as soon as possible will always be there, but it’s a fool’s game.
- Familiarize Yourself With the Quality Rater Guidelines: In its Search Quality Rater Guidelines, Google shares some critical insights that, frankly, every person who wants to rank on Google should know. Give them a read if you haven’t already.
AI Content Doesn’t Equal Bad Content
Based on our current knowledge and what Google says publicly, there’s no reason to think that Google penalizes websites with AI-generated content per se.
It’s just that a lot of poor-quality content happens to be created by Large Language Models (LLMs). And honestly, that’s not surprising.
Since they’re designed to fulfill a wide range of potential applications, LLMs are trained to provide general answers.
This is the reason why almost every article about health that you generate with ChatGPT will include at least one sentence saying something along the lines of “Please consult your healthcare provider for more information.”
It’s not a bug. It’s a feature.
What we can do, as copywriters, is to find trustworthy sources and present them to the language model as examples. Or, better yet, actually write some content ourselves and heavily edit whatever the machine provides, using it only as a helpful outline.
There’s also the issue of writing concise, straightforward prompts that can nudge the language model in the right direction. Creating such a prompt is considered by many a form of art.
Although it’s not easy, doing it right can significantly increase the quality of the output you receive from language models.
So, we’re not entirely helpless when it comes to making the AI content work for our intended purposes.
Given this context, saying that all AI-generated content is bad content is an overly simplistic description of this problem.
We don’t need to eliminate AI content. We need to eliminate bad AI content.
Should We Bother With AI Detection?
In most cases, determining whether we are dealing with AI-generated content is a smart move. After all, the more information we possess, the better choices we can make — at least in theory.
Still, before we can eliminate bad AI content, we have to identify it first. And that’s the area where humans prevail over AI content detectors.
Yes, an AI content detection tool will likely post better results when it comes to identifying whether we’re dealing with AI or human-written content.
However, it won’t be able to differentiate between high-quality and bad content.
The internet is filled with posts where people ask about reliable AI content detectors, such as this one:
Source:
https://www.reddit.com/r/ChatGPT/comments/13yo3xr/which_ai_content_detector_is_reliable/
Usually, there are multiple answers posted by other users, each recommending a different tool.
There’s also a comment or two where the user mentions that no tool is reliable.
Besides providing anecdotal evidence, they sometimes go further and share screenshots where the text undoubtedly written by a human is marked as AI-generated or vice versa.
Source: https://www.reddit.com/r/ChatGPT/comments/13yo3xr/comment/l09888o/
But is this discussion relevant in any way? Why do we have such an obsession with finding the best tool to detect AI content?
I believe that it has to do with our tendency to look for the simplest solutions to complex problems.
Many website owners see that sites with low-quality AI-generated content do poorly on Search and come to a similar conclusion: Since figuring out whether AI-generated content is good or not takes time and effort, why don’t we just ban it?
This approach may seem valid on the surface, but I consider it to be just another case of throwing the baby out with the bathwater.
Similar to how mindlessly publishing content that was generated by an AI is irresponsible, banning all AI content altogether is equally bad.
I won’t get into the details and list all the pros of using automation in the content creation process. Instead, let’s revisit the claim I made at the beginning of this article.
Are Humans Better at Detecting Bad AI Content?
Most people aren’t familiar enough with AI-generated content to correctly identify it in a set of options. That said, they can get better at it — provided they have some examples to use as training data.
This is even more true when it comes to detecting bad AI content.
Since AI detectors evaluate only whether the text was written by artificial intelligence, we cannot expect them to provide us with answers regarding the quality of the text.
That’s why I propose to treat AI detectors as useful tools and not the final obstacles that writers must overcome to get their articles published on a site.
Let’s get it out of the way — humans are not perfect.
According to worldmetrics.org, human error:
- causes about 90% of all data breaches
- is responsible for around 80% of workplace accidents
- causes up to 400,000 deaths in the healthcare sector every year in the U.S.
These statistics don’t paint a pretty picture.
The situation gets even worse when we look at the course description posted on the official website of the University of Texas at Austin, where we learn that:
“humans make at least three mistakes (usually 5-7) every hour that they are awake, increasing to 11-15 per hour under extreme stress or fatigue.”
Ouch.
Okay, so humans make a lot of mistakes. But do we mistakenly mark AI-generated texts as human-written? And if so, how often does that happen?
What Do the Studies Have To Say About This?
A 2021 study by the University of Washington found that human evaluators without training distinguished between GPT-3- and human-authored text at random chance level (they were correct around 50%).
At the same time, their confidence in providing the correct answer remained fairly high (around 50%) across all the given tasks.
| In short, evaluators underestimated the capabilities of modern language models and overestimated their ability to distinguish GPT-3 and human-written texts. |
Here are some of the explanations they gave to prove their point:
Source: https://arxiv.org/pdf/2107.00061
As you can see, we’re dealing here with two logical interpretations of the same text that lead to widely different conclusions.
Only after training on multiple examples were the evaluators able to significantly improve their scores.
Does it mean that the average Joe has no chance of figuring out the truth? No, but guessing what is and what isn’t AI-generated without training is just that — guessing.
It’s more like flipping a coin rather than doing any sort of technical evaluation or analysis.
What Can We Do To Improve These Numbers?
First of all, we have to familiarize ourselves with AI-generated content.
Many of us do it already — albeit involuntarily.
As I mentioned at the beginning of this article, the internet is becoming filled with AI-generated slop at an alarming rate. Google — quite reasonably — decided to take action against it, and now everyone is suddenly avoiding AI-generated texts like the plague.
But Google isn’t perfect, and bad AI-generated content sometimes manages to climb its way to the top of search results.
These are the moments when we have a chance to come across it and learn from the experience.
Case in point — BNN Breaking.
BNN Breaking was a news site that quickly grew in popularity. Over the first two years of its existence, it managed to build quite a reputation.
Well-known websites like The Washington Post and The Guardian were linking to it. It was also promoted by Google News and MSN, a web portal owned by Microsoft.
The website was taken down after an incident where the photo of a prominent Irish talk-show host, Dave Fanning, was mistakenly added by an AI to an article covering the case of sexual misconduct.
After careful examination, it turned out that the site was filled with AI-generated content. The New York Times covered the case in great detail in this article.
Why do I bring this up? Because it’s the perfect example of using AI in an irresponsible manner that nonetheless worked and brought 10 million monthly visitors.
If we put in only a bit more effort to fact-check the data and heavily edit the output generated by artificial intelligence, we can use the same tools in a much better and more reliable way.
Getting Familiar With AI Content
You don’t have to rely on random chance to find yourself some examples of AI-written articles. You can use ChatGPT for free and generate tons of texts in just a couple of minutes.
Then, it’s only a matter of putting in effort to familiarize yourself with these articles and search for repeating patterns.
Remember that study done by researchers from the University of Washington? Training with examples improved the overall accuracy of identifying machine- vs human-written texts from 50 to 55%.
The researchers note that “the significant difference in overall performance is mainly contributed by the story domain” and has to do with eliminating the preconception that machines cannot generate “creative” text.
Still, getting more familiar with AI-generated texts has its merits.
Here are some telltale signs that you’re dealing with content generated by a machine:
- Paragraphs are long and very similar in length.
- The article’s structure remains similar, regardless of the subject matter.
- There are many complex, technical terms in the text.
- The text is very predictable, i.e., it has low perplexity.
Of course, this list is far from exhaustive, but it should give you an idea of what to look for while you’re doing your research.
Fixing Bad AI Content
Okay, so you’ve identified weak content created by an AI. How can you fix it?
There are many ways in which you can improve these texts. The entire process is similar to what you would do if you were editing weak human-written content.
Here are a few factors worth paying extra attention to:
- AI-generated titles can be unimaginative. Change the title to give your article more personality.
- As mentioned above, AI tends to stick to one structure whenever possible. To mitigate this problem, either feed the language model your own structure or heavily edit the text, adding, changing, and removing headings as needed.
- Some AI content detectors have an issue with specific words or phrases, which they consider signs of AI use. If your article will be checked by these tools and you want to avoid detection, paraphrasing tools can come in handy. They will replace those problematic words with synonyms without changing the meaning.
- Fact-check the data and do your own research on the topic. If you find credible sources, cite them in your article to improve its quality.
- Refrain from blindly accepting everything the language model gives you. Mix and match the text to your heart’s content to create something truly interesting. Add some tables and lists to make the article easier to read.
- Provide some real-life examples. AI has no real-life experiences that it can describe, but you do! Use it to your advantage and get some bonus points by demonstrating one of the components of E-E-A-T.
- Remove repetitive sections and replace them with something new. Either write those new sections yourself or ask the language model to describe a particular issue that fits well into the overall theme of the article.
Using AI in a Responsible Manner
If I were to condense the message I want to convey in this article into a single thought, it would be this:
There’s nothing wrong with using AI to improve our content generation process. But we must do so responsibly.
It refers to both using a language model to generate the text and using an AI content detector to determine whether it’s worth publishing.
In both cases, we get a chance to work with tools that can do our job for us. But I don’t think that handing our task to an AI and calling it a day is the right way to go about it.
AI tools are not advanced enough to leave them unattended or take what they give us as 100% true. We still need to put in some effort ourselves.
Even editing the text generated by an AI is a step in the right direction.
Folks from Originality.ai still consider such texts AI-generated, but I disagree with this assessment.
Źródło: https://originality.ai/blog/ai-content-detection-accuracy
Like in many other areas of SEO, whether we can still consider the text to be AI-generated or not depends on various factors.
Primarily, it depends on how many changes we made to the article.
At which point or after how many changes can we claim the article is no longer AI-generated but human-generated? It’s hard to say.
Some might claim that after some improvements, we’re no longer dealing with editing but rewriting the article. But again — the line between the two is blurry.
When can we safely assume that what we’re doing is rewriting the article and not editing it? After we changed 50% of it? Less? More?
Who gets to make the call and resolve the matter?
As you can see, there are more questions here than answers.
Final Thoughts
The bottom line is that AI content detectors — like all AI tools — cannot be treated as infallible.
While humans make errors, using artificial intelligence isn’t a magical shortcut that will help us avoid making mistakes.
In the words of JC Denton, the protagonist of Deus Ex (the video game released in the year 2000):
“Human beings may not be perfect, but a computer program with language synthesis is hardly the answer to the world’s problems.”
I feel the same goes for all the AI tools that are so popular nowadays. They’re not the answer to our problems, but they can be a significant help in achieving the desired outcome.
We just have to be smart about how we use them and for what purpose.
