We're exploring plenty of ways to use GPT-3 more effectively in our SEO work. In this article, I wanted to list a few tools for detecting content created with GPT-3, since GPT-3 is the model available on the market as I write this.
Before diving in, even the rumor of 100 trillion parameters in the image below has me excited for GPT-4: (Image source)
Let me also summarize a few highlights from papers written about GPT-3:
- Understanding human behavior: AI can generate large amounts of content in seconds, but it often falls short on emotional connection, which can make it hard to hit the tone you're going for.
- Cutting-edge ideas: AI generates content based on data it's already collected, so it may not fully capture the newest ideas out there. It works from what it's gathered from the internet and whatever information it's given.
Keep in mind that there's always a margin of error when testing tools like these, so it wouldn't be wise to trust any single result completely. These tools can behave quite differently across languages too. All the tests below were run in English.
You'll find a list of AI content detection tools below. I haven't included every tool I've come across, just the ones I think are genuinely useful.
1. Content at Scale
Test page: https://contentatscale.ai/ai-content-detector/
Free use covers up to 2500 characters, and content over 25 words gives better results. Here, the tool flags content created by Chat GPT-3 as fake:
And when I check an original post, you can see below that it identifies it as human-written:
2. Writer.com
Test page: https://writer.com/ai-content-detector/
Another free tool worth testing. When I ran a human-written text through it, it clearly identified it as human-written, as shown below:
A piece of content written by GPT-3 scored 17%. Worth noting, a few more tests on AI-generated content I didn't touch scored much higher:
3. GPT-2 Output Detector Demo
Test page: https://openai-openai-detector.hf.space/
As the page itself states, this is a calculation tool that gives more reliable results after 50 tokens. Here, it clearly flags content produced by GPT-3 as 99% fake:
And when I test human-written content, it shows 73% of it as real:
Test page: https://www.gpt-detector.com/
Like its counterparts, this tool can check at a fairly basic level whether content was written by AI. Here's the result for AI-generated content:
And for human-generated content, it shows a low probability, meaning a human likely wrote it:
5. Gltr.io
Test page :http://demo.gltr.io/client/index.html
Gltr.io uses the GPT-2 117M language model. Purple and red in the output specifically point to text written by a human. And if the histogram results skew right, you can already tell fairly clearly whether an article was human-written or not. Below is the result for a piece called What is SEO written by AI: mostly green and yellow.
And when we look at human writing, the model shows purple and red much more often:
These tools aren't just useful on your own content, they work for analyzing your competitors' too. Some of your competitors may be running autonomous or semi-autonomous content on parts of their site. Take the Washington Post, for example: they use AI to generate content:
These tools can also help you see how your competitors build their FAQs, and even let you track how their relevant pages perform over time, so you can shape your own action plans around what you find:
Will Google Devalue AI-Generated Content or Take Manual Action?
Google's John Mueller said that content automatically generated by AI counts as spam under webmaster guidelines, which can hurt a site's quality and even lead to a penalty.
The Web Spam team can go as far as taking manual action against a site over this. Automatically generated content was already a guideline violation before AI made it this easy. Worth noting: Google has LLMs even larger than GPT-3, like PalM, and here's an example of what that looks like:
Let's close this article with Chat GPT-3 itself. I asked it to write a closing sentence:
At the end of the day, remember these tools aren't perfect and can get it wrong sometimes. The links to these pages may also go stale or change over time. If you spot a broken one, let me know and I'll update it as soon as I can. :)
