How To Detect AI-Generated Content
The 2023 artificial intelligence boom took the digital world by storm, leaving everyone baffled trying to tell AI-generated content from authentic human-written content.
Today, AI tools are evolving from fun internet gimmicks to powerful, adaptive, and complex tools that shape lifestyles, workplaces, entertainment and business.
But it doesn’t come without a catch.
Enter spammy AI-generated content.
If you saw the viral image of the Pope wearing a white puffy jacket, Kuala Lumpur’s annual Cat Marathon, or roller skating grannies, you’re probably among the people who took a while to figure out the posts were AI-generated.
Not detecting the difference between machine and human content leaves potential issues for SEO, content development, and misinformation.
This is where AI content detection comes in. It’s one way people can uncover when something is AI-generated and make informed decisions as businesses or customers.
Let’s dive deep into popular AI tools like ChatGPT, how you can detect AI-generated content, and where Google stands between the human and artificial intelligence spectrum.
ChatGPT: In a Nutshell
ChatGPT is an AI language model based on GPT-4, which is trained to interact with prompts conversationally. With its current AI model, ChatGPT can respond to follow-up questions, challenge incorrect prompts, admit mistakes, and reject inappropriate commands.
The sentences produced by AI models like ChatGPT look and sound right. But while they use the right words and generate correct answers, the tool doesn’t know what it means.
AI models generate content by predicting the likely next word in a sentence. Over time, AI models improve thanks to complex training and reward systems that serve as a feedback loop, training the tool to follow a desired path or produce a desired output.
It follows syntax, grammar, and other instructions during training, but ChatGPT doesn’t know if the information is correct, false, or true.
Why is it so hard to detect AI-generated content?
AI tools are designed to mimic human behaviour, language, and fluency. These models were trained on mountains of parameters and existing databases (human-written text, images, etc.) to generate plausible content that is difficult to distinguish by the untrained eye.
Moreover, the research found people trust AI-generated content as much as real text, as about 72% of respondents viewed AI-written articles as credible.
A University of Maryland report found that state-of-the-art detectors struggle to spot Large Language Model (LLM) outputs in a typical scenario.
But that’s looking at it from a technical perspective.
Google, like content specialists, looks deeper into content quality, plagiarism, and other intuitive factors that telltale signs something is AI-generated.
How to spot AI-generated content (without using detection tools)
There’s no single rulebook for detecting AI-generated content, but here are things to look out for when reading text or browsing images:
Repetition of words and phrases
AI tools sound fluent and well-versed on all topics, but not as much as human experts. In many cases, ChatGPT and other AI tools repeat keywords about a topic with little-to-no variation.
Lacks topic depth
“Remember your training!”
Artificially generated content only processes information based on their training and knowledge database. That leaves them detached from up-to-date knowledge and lack the ability to develop a unique insight about a topic.
In other words, AI content typically lacks depth and will read more robotic or prescriptive than clever, creative, and engaging. It only works excellently on topics it was trained for, so if you’re curious about a niche topic not covered by ChatGPT’s training, you won’t get a comprehensive response.
Produces outdated information
Most content developed by AI is usually correct. However since it only makes predictions, it can sometimes generate incorrect responses or outdated facts.
ChatGPT is only trained with information up to 2021, while Google knows the latest information. So you can’t make excellent content about the 2022 FIFA World Cup or the latest Gen Z marketing statistics.
Predictable content structure and format
AI content follows human-like structure and syntax when writing content. But if you look at the bigger picture, AI-generated sentences are streamlined, thus making them typically shorter and lacking the complexity, variation, and creativity humans produce.
On the other hand, human-written content is more likely to have minor typos and also use casual language, slang, and informal tone.
Best tools to detect AI-generated content
All AI detection software available online is not foolproof. The sheer complexity of reverse-engineering AI-written content is why most AI content slips through (make sure to also check out how businesses are adopting AI for operation and marketing purposes).
Regardless of how fast AI improves, AI content detectors and AI content detection tools are more reliable and objective in flagging artificially generated content than human intuition.
We reviewed popular AI content detection tools and tested their efficacy with a sample AI-generated passage.
Here’s what ChatGPT gave us when we asked to write a paragraph about coffee:
“Coffee is an absolutely amazing drink that people love to have all the time. It has a really long history and is culturally significant. People just can’t get enough of it! It has many flavors, from earthy to fruity, and can be enjoyed in many different ways. Coffee also gives you a real energy boost because of its caffeine. Plus, it’s a great way to socialise with people. But, you must be careful because too much caffeine can cause problems. In conclusion, coffee is a super awesome drink that people enjoy, but you have to be mindful of how much you have.
Coffee, an aromatic beverage beloved by millions around the globe, has long been a source of inspiration, comfort, and stimulation. Its rich history and cultural significance make it a captivating topic to explore. From its humble origins in the ancient coffee forests of Ethiopia to its widespread cultivation and consumption in various forms today, coffee has woven itself into the fabric of human existence.
With each sip, coffee imparts a symphony of flavors that dance upon the taste buds. The dark, velvety liquid carries a distinct profile that ranges from earthy and nutty to fruity and floral, enchanting the senses with its diversity. Whether enjoyed as a simple black brew, transformed into a frothy cappuccino, or blended with indulgent flavors, coffee possesses a versatility that caters to the unique preferences of individuals across the spectrum.”
That said, let’s explore how well the best AI content detection tools perform with the coffee exerpt above!
Originality AI is a flagship AI-generated content and plagiarism detector designed with content marketing and web publishing in mind. In a benchmark of 1200 data samples, the platform recorded a 95.93% true positive score.
But how does it perform in our coffee example?
It did not even stutter!
Originality.ai rated our AI-written content as 100% machine-generated—on the dot. The platform also runs plagiarism and readability tests along with AI detection.
Here’s what it has to say about ChatGPT’s coffee 101 masterpiece:
One segment in our sample detected a perfect match with existing content. Tendencies like this occur since AI language models scour their database and come up with pieces of information interwoven together to look original—sometimes not paraphrasing content at all.
In terms of readability, our sample scored 48.6 under the Flesch-Kincaid Reading Ease rating. For content to be easy to read and digest online, writers must aim to score 60+ minimum.
Anything below that will be scanned and skimmed by readers with short attention spans.
It’s safe to say their claims hold for this test. But what do the other tools have to say?
Writer AI is a free content detector tool that roughly estimates how much content is produced by AI. It has a paid version that holds the same 300-word (1500-character) limit, and both give a predicted percentage of how much of your text is genuine and written by a human.
It scored the first paragraph of our sample content as 3% human-written content. A pretty accurate score at that!
But the tides took a different turn as we plugged in the entire passage.
ChatGPT’s article scored 99% human-generated content! You be the judge.
Nevertheless, this tool is excellent for B2B and content developers analysing written content before publishing.
Microsoft-backed research firm OpenAI released AI Text Classifier as a platform to detect AI-written text. Think of it as ChatGPT’s counterpart.
OpenAI feeds AI Text Classifier with a plethora of pre-labelled text written by humans and machines until the platform can tell between the two. It requires at least 1,000 characters before you can run a check and detect content written by ChatGPT and ones from Google Bard.
Simply paste your text and submit it. Unlike other tools in this post, the platform will simplify or hedge its output. Based on its confidence, it will respond with labels such as:
- Likely AI-generated
- Unclear if it is
- Very unlikely
GPTZero is a tailored AI text detector designed by a Princeton grad for professors to check if essays turned in by students are AI-written. It’s a clever and useful tool to help preserve academic integrity in schools without barring students from leveraging AI-generated content.
The platform evaluates user entries on two main parameters:
- Familiarity: If GPTZero sees familiar indicators in the submission from machine-generated content it was trained on.
- Uniformity: GPTZero checks how uniform the content is. Machine-generated text is usually robotic and uniform, while humans develop content with diversity and variance in syntax, text length, etc.
At this point, it’s clear that using multiple tools narrows down your results. Undetectable AI Multi-Detection Tool is fine-tuned from multiple AI detectors like GPTZero, OpenAI, Writer, Crossplag, ZeroGPT, etc.
To use it, paste your written content on the text box, and specify the following parameters:
- Readability: (High school, University, Doctorate, Journalist, Marketing)
- Purpose: (General writing, Essay, Article, Marketing, Story, Cover Letter, Report, Business, or Legal Material)
Like previous tools, Undetectable AI checks the likelihood of machine learning to produce a data-driven result based on all content it tested in training. Since it’s trained with eight varied AI detectors, you’re essentially checking your entry over eight AI detectors at once.
Its results are inconclusive because they were verified on several machine learning algorithms. Nevertheless, it’s an excellent tool for contextualising content structure, flow, and creativity—all requiring human touch and expertise.
Google’s guidance for AI-generated SEO content
On April 1, 2022, Google Senior Search Analyst John Mueller said the webmaster guidelines considered machine-generated content as spam. But after several months, Google shifted its stance to only against spammy AI content.
Moreover, Google believes AI can be leveraged to deliver helpful information better and faster than ever.
E-E-A-T is Key
Google will reward content however it is produced. That said, ranking high means producing high-quality and rich content demonstrating expertise, experience, authoritativeness, and trustworthiness (E-E-A-T).
Besides, not all AI-written content is bad, and not every spammy online content is AI-generated. Well-researched, well-developed, and highly engaging content will remain valuable for SEO and your readers, no matter how they’re produced.
The shift towards a certain threshold on AI content is, in retrospect, to webspam practices about ten years ago.
Back then, Google was combatting a rise in mass-produced human-generated thin content. Yet it doesn’t make sense to ban all human-generated content to put an end to webspam practices.
The same holds in today’s AI content, in which Google claims it will keep improving its algorithm to reward high-quality content.
In fact, Google released helpful content update to ensure users get search results created for people and not for search ranking only.
Final thoughts for content creators considering AI
Artificial intelligence’s potential and capabilities can make or break your SEO content campaign. As AI tools become more robust and accurate, AI detection will become more important in determining content legitimacy.
Besides, Google will look into producing original and high-quality content designed for people by people. To demonstrate E-E-A-T in AI-generated content, webmasters and content developers should take the extra mile and verify, improve, and frame AI content to create reliable people-first content.
Contact us today if you want to learn more about how to incorporate E-E-A-T into your content. Our SEO specialists in Sydney will take you through the process and help you stay on course toward your business goals, whether with AI content or not.