It’s becoming increasingly difficult to differentiate between Human-written and AI-generated content.

As a content manager or a blogger, you may find yourself in a tricky position.

Is the article you received from your freelancer genuinely written by a human?

Are you concerned that Google may penalize your AI-generated content?

This blog post aims to demystify the process of detecting AI-written content and provides practical tips to make your content undetectable by AI.

In my opinion, AI content detection is often overhyped. For example, big tech companies like Google have the capability to detect AI content.

Relying on third-party models with unclear documentation and thinking they can outsmart Google when it comes to detection is dumb.

I’ll be discussing how to responsibly detect AI-written content.

So, let’s dive in and explore!

Understanding AI-written content

Before I discuss detecting AI-written content, it’s crucial to understand how AI creates content.

At the heart of most AI content generators is a model called GPT, or Generative Pretrained Transformer.

Think of GPT as autocomplete on steroids.

It’s a prediction machine, trying to guess what should come next, one word at a time, based on what it has learned about the world through text.

In simple words, GPT is like your phone’s autocomplete feature (on steriods!).

The latest iterations, GPT-3 and GPT-4, are incredibly sophisticated but have limitations. For instance, they’re trained on historical data, which means they consistently miss writing about recent developments.

Now, let’s look at some common features of AI-written content:

1. Predictable language: AI tends to use more common phrases and avoids less common ones.
2. Lack of personal experience: AI can’t share personal anecdotes or experiences because it doesn’t have any unless you have it hallucinate giving it a plot to build on. Like say – “It can travel to Zambia on behalf of you, come back and write an essay on it sharing the experience while you’re sipping coffee”
3. Inconsistent details: AI might provide inconsistent details over a long piece of text.
4. Overuse of certain phrases: AI might overuse certain phrases or words, leading to a lack of variety.
5. Lack of deep understanding: While AI can mimic understanding of complex topics, it doesn’t truly understand them.

Based on these, I have shortlisted some proven techniques to help detect AI-written content.

Manual techniques to detect AI-written content

Lack of recent knowledge

AI models, like GPT-3 and GPT-4, are trained on historical data.

This means it consistently misses writing about recent developments.

Even if you prompt the AI with the latest developments, it may still end up relating it back to historical events. So, if an article seems stuck in the past, it might be AI-generated.

For example, as of writing this post, GPT-4 has a knowledge cut-off, which is in September 2021. This means that it generally lacks knowledge of events that have occurred after this date.

Example: An AI-written article on cryptocurrency might say, “Bitcoin, created by the mysterious Satoshi Nakamoto, has revolutionized the financial world.” But it might fail to mention recent developments like Ethereum 2.0 or the rise of NFTs.

Unusual creativity

This one might seem counterintuitive.

After all, isn’t creativity a human trait?

Yes, but AI’s creativity is a bit different. It’s consistent, almost too consistent.

If you ask an AI to make a paragraph witty, it will add wittiness consistently throughout.

But as humans?

Our creativity ebbs and flows based on our energy levels, mood, and that second cup of coffee we just had.

So, if a piece of content maintains the same level of creativity throughout, it might be the work of an AI.

Analysis of overly coherent sentences

Humans are subject to mood swings, and our writing reflects that.

One moment we’re writing long, elaborate sentences, and the next, we’re keeping it short and sweet.

But AI?

It maintains a consistent tone of voice. It doesn’t have mood swings or energy slumps. By default it doesn’t use sentence fragments, and there are no ups and downs in sentence length.

So, if you’re reading a piece of content that’s almost too perfect, with no pattern interrupts, it might be AI-generated.

Example: A human might write, “I love ice cream. It’s just the best, you know? But the calories, oh boy!” An AI, maintaining a consistent tone, might say, “Ice cream is a delightful dessert that many people enjoy, although it is high in calories.”

Pattern repetition and generic phrases

AI is programmed to predict and generate text based on patterns it has learned.

However, this often leads to the repetition of patterns and the use of generic phrases, which can negatively impact the quality and uniqueness of the content.

Also, AI-generated text tends to have low-value density, meaning it contains a lot of filler phrases and redundant information. Human-written text, on the other hand, tends to have higher value density (of course, based on the writer’s experience), meaning it contains more relevant and useful information.

For instance, an AI-generated article on a topic like “how to make a sandwich” might include irrelevant details about the history of bread-making or the science of taste buds (unless AI is primed well with strategic prompts)!

Lack of personal experience or emotional depth

AI lacks the human element and emotional intelligence.

It can’t share personal anecdotes or experiences because it doesn’t have any.

It can mimic emotions based on the data it’s been trained on, but it doesn’t truly understand them. It doesn’t need to. Mimicking suffices here. What it needs to do is to convey the feeling and the rest is taken care of by our god-gifted emotional intelligence.

Lack of self-references

AI doesn’t possess self-awareness or personal consciousness.

It can’t refer back to something it said earlier in the text because it doesn’t have a memory of what it wrote.

Example: A human might write, “As I mentioned earlier, SEO is crucial for your blog’s success.” An AI lacking self-awareness might not make such references.

Error propagation

If an AI makes a mistake early in a text, it may continue to propagate that error because it lacks the ability to retrospectively correct errors.

Humans, however, are more likely to correct an earlier mistake when they notice it later on.

Example: If an AI incorrectly identifies a cat as a dog early in a text, it might continue to refer to the cat as a dog throughout the piece.

Perfect grammar, style, and punctuation

AI-written content often has perfect grammar and style. It uses semicolons, hyphens, dashes, and commas correctly. While this might seem like a good thing, it’s almost too good to be true.

The most evident ones are the perfect use of semicolons, dashes, and hyphens.

Human writers, even professional ones, occasionally make minor grammar or style mistakes.

Lack of parenthetical statements

Human writers often use parenthetical statements or brackets to provide additional context, clarify a point, or present an alternative viewpoint in the middle of their content.

On the other hand, AI lacks personal consciousness and doesn’t use such statements.

Example: A human might write, “I love ice cream (especially mint chocolate chip), but I hate the calories.” An AI might write, “Ice cream is a popular dessert, although it is high in calories.”

Best AI detection tools

In the ongoing arms race between AI content generation and detection, several tools have emerged that can help you identify AI-written content. Here are some of the best ones:

• Originality.ai (Best value for money): Another tool that can help detect AI-written content. However, like all tools, it can present false positives. It provides the best value for money as the pricing is based on an as-you-go basis.
• Content at Scale AI Detector (Best free tool): Offers a free advanced AI Detector that checks if your content is human or AI-generated from ChatGPT, GPT4, Bard, and other language models. It can check up to 25,000 characters at once. This tool goes deeper than generic AI content detectors, offering more transparent scoring.
• Undetectable.ai (Best detection): Checks content through a fine-tuned model trained on batched documents submitted to each of the AI detectors they feature, such as Originality and GPTZero.
• ZeroGPT: A tool designed to detect content generated by GPT models, including the latest iterations like GPT-3 and GPT-4.
• Writer.com AI Detection Tool: Offers an AI detection tool to help identify AI-written content.
• GLTR: Uses visualizations to help you detect AI-written content. It’s a great tool if you prefer a more visual approach.
• OpenAI Text Classifier: A tool by OpenAI that can help detect AI-written content. It consistently performs poorly compared to other tools in the blog post.

While tools like Originality.ai or ZeroGPT can be preferable, they can also present false positives.

And on the other hand, to make things worse, there are tools like Undetectable.ai that help writers generate content that’s undetectable by these tools. But based on my talks with my content writers, it doesn’t work that effectively.

It’s an ongoing arms race!

How to make content undetectable by AI?

Some savvy writers have learned to make their AI-generated content undetectable.

They might vary sentence structures, break patterns, sprinkle in self-references, or even introduce a few punctuation errors to mimic human imperfections.

Some even take it a step further by paraphrasing AI-generated content using tools like QuillBot or Netus AI Bypasser (works well!), making detection even harder.

But remember, the goal isn’t to fuel the AI arms race.

Instead, focus on creating valuable and engaging content. After all, whether a human or an AI writes it, useful content is what truly matters.

Suppose your tool shows a 100% confidence score that a piece of content is AI-written; who cares? It’s just a score.

Even Google doesn’t care.

AI content is not bad

It is important to debunk the notion that AI-written content is inherently bad!

In fact, Google has stated that what matters is the usefulness of the content, not how it’s generated.

In a 2023 blog post, Google clarified, “Using AI doesn’t give content any special advantage. It’s just content. If the content is useful, helpful, original, and satisfies aspects of E-A-T (Expertise, Authoritativeness, Trustworthiness), it might rank well in search. If it doesn’t, it might not.”

AI can generate ideas, provide structure, and even write coherent and contextually relevant content. However, they should be used as a supplement to human thought and creativity, not as a substitute.

The purpose of AI is to augment human capabilities, not replace them.

They can help us process information faster, generate ideas we might not have thought of, and even write content when short on time or energy.

So, if you’re using AI to structure your ideas and create content, there’s nothing wrong with that.

In fact, generative AI is meant for that.

After all, AI is a tool, and like any tool, it’s how you use it that truly matters. But be transparent about it.

Pro tip (Here’s what I’m doing)

Detecting AI-written content (that too manually) can be tricky.

Because it’s not just about understanding the quirks of AI writing; it’s also about acknowledging our biases and potential errors in judgment.

Added to this, as AI advances (which it is, at a staggering rate!) and prompt designing becomes more sophisticated, these manual detection tips may not always work.

So, the best approach is to foster transparency with your writers. If you have a well-connected team that knows how to use AI responsibly, you’re already sorted!

But what about when you’re hiring freelancers?

Tip #1

One effective strategy is to use Google Docs.

Its version history feature allows you to look for anomalies and patterns in how the content was built up. Did the writer copy and paste large chunks of text? Is there a lack of editing? Then chances are there it’s written by AI!

You can also use tools like Originality.ai which offers a Chrome extension that allows you to “watch a writer write” the article.

The tool screen-records the entire writing process, allowing you to easily determine if a piece of content was written by a human or copied from AI. However, this method has its limitations as a person may choose to manually type the entire content rather than copying and pasting it.

The extension can reverse engineer every edit made in the document to create a real-time or sped-up display of the document’s creation.

If a writer simply copies and pastes from ChatGPT, you’ll see a single block of text getting dropped into Google Docs!

Tip #2

Another good practice is to ask writers to come up with a really detailed outline or draft before using any kind of AI.

This ensures that the core ideas and structure of the content are thoughtfully considered and crafted by a human.

After completing the outline or draft, AI can be used to structure the content and create coherent narratives.

Conclusion

In conclusion, detecting AI-written content is becoming increasingly difficult, but there are still manual techniques that can help.

By understanding the limitations of AI-generated content and looking for specific features such as lack of recent knowledge or unusual creativity, you can identify potential AI content.

However, it’s important to note that AI content is not inherently bad and can be a powerful tool for content creation when used responsibly.

Ultimately, the quality and usefulness of the content should be the main focus, regardless of whether it was written by a human or an AI.