This study will look at the 5 best AI content generators for long-form SEO blog posts. But most importantly it will focus on using brand new AI that is the first to be able to detect GPT-3 AI-generated content all to try and answer… can Google detect AI content?
Since OpenAI released its GPT-3 API a large number of amazing AI content generation tools have been built. These tools can do a lot of things from writing emails to creating a Facebook ad. However, for this study we are focusing our efforts on long-form SEO blog posts that have the objective of ranking in Google.
We will be using Originality.AI which is the world’s first AI trained to detect if a piece of content has been created by AI tools using OpenAI’s GPT-3 API which tools like Jasper.ai use or other popular NLP APIs.
What Originality.AI Was Built to Do… AI Exposing AI
The unchecked proliferation of AI content is not without its concerns. How do we responsibly/ethically use this awesome power of AI content creation? How do we ensure that Google does not penalize us for leveraging AI content?… these are all questions I think I am in a unique position to try and help answer.
Who Am I to Talk about AI Content?
Although I am an engineer, I am not an AI developer however I…
- Founded (and now sold) a done-for-you AI Content Agency: Rocket Content… where we were the heaviest user of Jasper.AI, and our team partnered with Jasper to develop a course and webinar for Jasper.AI’s agency users.
- Manage a portfolio of content sites some of which we are ethically leveraging AI content
- Founded Originality.AI where we built our own AI that is able to predict if content was produced by AI (the world’s first GPT-3 AI detection tool!)
However, clearly I don’t know it all since I have my name on an AI patent that was not approved!
Before we get to the results of the study that determines if Google can detect content generated using GPT-3 (or any of the popular NLP models) we need a little background…
What is AI Generated Content?
AI Generated Content is when a machine produces content such as a blog post, article, headline, blog intro, ad copy, social media content, product description, content idea, chatbot response, email or even images. It has exploded since May 2020 when OpenAI provided access to the API of their pre-trained NLP called GPT-3.
The Secret about ALL AI Content Generators
There are multiple pre-trained natural language processing models but it was really when GPT-3 came on the scene that the quality of the content produced became useable for marketers.
So why was GPT-3 such a big deal in allowing an AI content generator to become useful? It was not the first or the only NLP model that has an API which developers can build tools with.
Pre-trained NLP model capabilities, up to this point, are best measured by their number of parameters, the table below shows the most popular NLP models that AI tools could choose to use and the number of parameters:
- Open AI GPT-2: 1.5 Billion
- EleutherAI GPT-J: 2.7 Billion
- EleutherAI GPT-Neo: 6 Billion
- Open AI GPT-3: 175 Billion (DaVinci)
The “secret” is that ALL tools rely on the same API’s to create their content. These pre-trained NLP models are so incredibly costly to train that it would be impossible for even the best-funded AI tool to compete with its own NLP model. At a 30:1 difference in the number of parameters for GPT-3 to its nearest NLP competitor, the reality is that all effective AI content generation tools are using GPT-3 as their foundation.
Even though they are all using the same base there are still differences that can result in a big change in the quality of the content produced.
Fun Fact – EleutherAI (GPT-J, GPT-NEO) (scary sounding name) is actually truly open source while OpenAI(GPT-2, GPT-3) is very much for profit. The community, such as huggingface, of developers around AI is incredible!
Okay, enough nerding out…
Can Google Detect AI Content?
Yes, the study below shows that it is possible for Google to build (or has already built) its own AI that can successfully predict if content was created with an AI tool using GPT-3 (or other NLP model). We know this because at Originality.AI we built and trained our own AI that can predict if a piece of written content was created by an AI content generator using GPT-3 with 94% accuracy and higher for all other available NLP models.
Originality.AI has shown that the rumour GPT-3 created content was undetectable is wrong. Originality.AI was able to build a model that can detect content produced by all of the popular NLP frameworks with a high degree of confidence:
- GPT-3 Detection Accuracy is 94.06%
- GPT-J’s Detection Accuracy is 94.14%
- GPT-Neo’s Detection Accuracy is 95.64%.
As we head into 2023 there are no other tools that can do this and we are gearing up to be ready when GPT-4 goes live.
So what? Why do we as content marketers care if Google (or others) can detect content that was created by a bot?
Google has come out and made it clear that it does not want AI-generated SPAM content in their Helpful Content Update they basically said as much. However, where it gets a little more mirky is what about content that was initially created by AI but then a human verified its accuracy and added some additional value? Will this get impacted by Google? No one knows!
The risk for all web publishers that don’t write every word themselves, right now is that writers they have hired are using AI writing aids to help create content faster and are now publishing content that could be identified as AI-generated.
As covered in this article or this article or this article it is widely understood that Google is scared of AI content damaging the usefulness of the text-based web which would result in Google making less money.
Now that the stage is properly set let’s get to the studies results!
STUDY: Examples From 5 Of The Most Popular AI Content Generators and Determining if Google can Detect Their Content?
We ran samples from 5 of the most popular long-form content AI generators through Originality.AI to test the confidence with which the content could be identified as not human-generated. In addition, we ran a control set of human-generated content to verify the probability score was identifying human as human and bot as bot.
Here are the results…
|Source of Content||AI Probability Score|
Jasper produced the content that had the lowest probability of being produced by the GPT-3 API. This likely reflects a lot of the extra work they have done on top of the OpenAI API.
To help show the tool is providing accurate results we ran the 5 most recent articles at VentureBeat covering AI through the tool and you can see the results below.
Although the AI at Originality.AI can not say with absolute certainty if content was created by GPT-3. You can see from this study it does an impressive job at determining what is human-created content and what is bot-created content.
Originality.AI Exposing GPT-3 Content…
Here are the AI probability scores for the editorial articles from Venture Beat we safely assume are human-generated…
Below is a detailed breakdown of each tool and a video showing an example of the output of each tool being tested by Originality.AI
Jasper.AI is an incredible tool, they have established themselves as the leading AI content creation tool with a $125M raise.
This tools popularity and recognition is for good reason, they were our go-to tool when we built RocketContent.ai
The downside is that with their popularity, they have the challenge of needing to be a tool for all marketers looking to leverage AI. They are not focused on long form SEO content in particular like some of these other tools are.
They are definitely the best tool if you are signing up for an AI Content Creation tool for the first time or need an AI tool that can do it al!
Can Google Detect Jasper.AI Content:
Yes, using Originality.AI we are able to identify when content has been created by Jasper.AI.
In the video below I show live as I use Jasper to create content and then test it with Originality.AI to see if it can be detected as generated by a GPT-3 tool.
- Rich Features
- Quality output
- Community & Support
- Trying to serve many masters
Jasper.ai Is Best For:
Anyone buying their first AI content creation tool. With its diverse group of tools it can support any marketer whether they are writing emails, sales pages or creating blog posts. It has established itself as the lead tool for good reason.
If you only care about producing long-form SEO blog posts with a team I currently prefer WriteSonic.
See Jasper.AI Examples – Here
Rytr is an interesting option for writers… it has the most affordable “unlimited” monthly plan.
When we used it we struggled to have it efficiently produce full blog posts that didn’t wander wildly off-topic. It seems that other AI tools have produced the ability to extract more focused content as the text gets longer but Rytr definitely struggles with this.
The product is not as polished, easy to use or feature-rich as Jasper.AI so it is not best for people’s first tool.
The content produced does not seem to be at the same level as other tools. Within GPT-3 you can select which model you want to use and my guess is that Rytr has not selected the premium.
However, it is fast and the most economical unlimited tool.
Can Google Detect Rytr-Generated Content?
Yes, using Originality.AI we are able to identify when content had been created by Rytr. Originality.AI showed with an average of 95% confidence that the content created by Rytr was produced by a bot.
In the video below I show live as I use Rytr to create articles and test them for originality.
- Affordable unlimited word count monthly plan
- Plagiarism checker integrated
- Easy to use
- Not great for long-form blog posts
- Not as feature-rich as competitors
Rytr is Best for:
I think of Rytr as the cheaper version of Jasper.ai. It is best for freelance writers who are producing a lot of lower-cost content and looking for a budget-friendly AI solution to help them.
Rytr is your best low cost option, even though all these tools rely on the same NLP the results as you can see in the examples do vary. The Rytr examples are not as good as most of the other tools.
See Rytr Examples – Here
Write Sonic has become our go-to tool for creating long-form SEO blog posts. The content stays on point and flows well.
Where many tools such as Jasper.ai and Rytr try and be a swiss army tool for digital marketers Write Sonic has focused more heavily on the use case for writers producing blog posts.
It has some nice features to make it a good fit for teams of writers including the ability for multiple users and to select different quality content options.
It is definitely not the lowest-cost option with the cost per word being as high as $0.01/word.
Can Google Detect Write Sonic Generated Content?
Yes, using Originality.AI we are able to detect WriteSonic content as being bot-generated. For the example articles that were run through the tool it predicted with a confidence of 99.4% that the content was bot-generated.
In the video below I show live as I demo WriteSonic, show examples of content and run them through the tool at Originality.AI:
- Great long-form content
- Able to add multiple users
- Able to get started quickly
- Does have a free-trial
- No Unlimited plan
- Fewer features than others (which is also a positive)
This is not the lowest cost option but it can produce excellent long-form content if that is what you are looking for.
Write Sonic is Best For…
Web publishers with a team of writers looking to leverage AI to increase their team’s efficiency.
WriteSonic and Jasper.AI are our current tools.
Write Sonic Examples – Here
Frase is a phenomenal tool that started as a great way to make content briefs easily.
It has evolved a lot over the last 2 years and has now added AI content writing to its features.
Because it was originally built not to be just an AI content creation tool the complexity is higher than most.
Can Google Detect Frase.io AI Writer-Generated Content?
Yes, using Originality.AI we are able to identify when content was created by Frase. Originality.AI’s AI predicted with a full 100% confidence that the content was bot-generated.
- It is the best tool to quickly create brief analyzing the top 20 results on Google
- Feature rich tool for SEO writers that goes beyond simply having a bot create content
- Can be complex… the tool has a lot of pieces to it and if you are looking for a simple AI writer this is not it.
- Limited to 30 articles per month unless you pay a sizeable $149.99/month for unlimited
Frase AI Writer Tool is Best For:
If you are a serious SEO writer who writes a lot of your own content and would like to have an AI writing assistant paired with a lot of other more advanced SEO content tools to optimize your content to rank. I would suggest signing up to Frase because of the other features first (brief creation etc) not because of its AI writer.
We use Frase for brief creation to provide writers with the ability to ensure they completely cover a given topic.
5. Article Forge – Strongly Not Recommended
Article Forge is the only tool on the list that explicitly says they do NOT use GPT-3. However, it raises the first of several questions… what do they use if it’s not GPT-3.
There are a series of concerning issues when testing Article Forge that has me warning people to NOT USE article forge!
Article Forge has some very interesting functionality for SEO content managers:
- Bulk content creation
- Automatic uploading & scheduling
Just upload a list of keywords, add a login to your site and poof perfect SEO-optimized, factual and helpful content published on your site… but this claim I fear belongs in the bucket of too good to be true!
Here are the issues I was able to identify with the help of Originality.AI…
Issue 1 – Article Forge Content is Easily Identified as AI
Article Forge states they do not use GPT-3 but don’t explicitly state what they do use. Since our AI is able to consistently identify their content as AI it means they are likely using GPT-2, GPT-J or GPT-NEO. The quality of the output also suggests it is not GPT-3. However, since our tool is able to detect it that means it has to be one of the NLP models our AI is trained on so either GPT-2, GPT-J or GPT-NEO.
Issue 2 – Articles are Mostly Rewrites of top Ranking Articles
Here is how I went about testing this theory on how Article Forge works…
Theory: Article Forge finds top-performing pages for a given keyword and then uses AI to re-write sections of it.
- Create an article using Article Forge
- See that it passes keyword plagiarism (ie Copyscape or Originality.AI)
- Enter each LONG paragraph into Google and see how similar it matches up to the same article repeatedly
Article Forge uses an NLP API to re-write sections of top results in Google for a given keyword resulting in an un-original article in the eyes of Google.
Potential ISSUE 3 – Article Forge Content Does Not Rank in Google
The example articles at Article Forge do not rank well even if you search a full paragraph of the article. This could be a unique Google manual penalty to a portion of a page or it could mean that Google has been successful in identifying Article Forge content as not helpful.
There is indexable relevant content on their homepage (no googleoff tag or no index) that was produced by the tool. It should rank very high if we search an entire paragraph of the example article. However, when I grab the first paragraph and enter the entire thing in Google the Article Forge article is the 7th result.
Article Forge Summary
I would strongly recommend not using Article Forge since it appears that Article Forge is using either GPT-2, GPT-NEO, or GPT-J to re-write sections of top articles in Google for a given keyword. The result is generally a worse copycat of the original article that barely passes Copyscape but is in no way Original.
If either a user or Article Forge team member wants to communicate if/how I am wrong please reach out!
AI Content Generation Tool FAQ:
Here are some additional tools that will be studied in the near future as we work to discover the best AI content generator and if any are truly “undetectable” as GPT-3 generated content…
Yes, most tools have a free trial. There is not AI writer that is 100% free.
When it comes to ranking webpages, Google puts a lot of emphasis on the quality of the content. In particular, the search engine giant looks for content that is original and informative. This means that simply regurgitating information from other sources is not enough to earn a high ranking. To Google, originality means adding your own insight and analysis to the conversation. This could involve offering a unique perspective on a current event or providing new data that challenges existing beliefs. In either case, the goal is to contribute something new and valuable to the discussion. By heavens, there must be something left in the world that hasn’t been said before! If you can find it and say it well, Google will take notice.
Originality.AI was built to identify original unique content by answering for a piece of content…
1. Was it written by a human?
2. Did they copy/plagiarize it?
Will AI writers keep getting better?
The typical obvious answer is yes, especially with GPT-4 coming in 2023 (suspected). However, there is an interesting challenge on the horizon for NLPs. The constraint for training models is quality data, training method and computing power. With chip shortages and GPT-3 already using 60 million domains (and likely the higher quality portion of the web) to train there is potentially 2 significant constraints to the next model being way better then GPT-3. It seems like the next model can not just be brute forced to be bigger but they will need to train the next NLP better. Smart people are working on it and I am excited to see the results!
This article specifically looked at written content, if you are looking to understand how AI can be used to generate art check out this article this article. Despite showing in this article that AI content can be detected by another AI I do believe AI content generators/writing aids are the future and we are best to learn to work with them while ensuring if you are publishing anything on the web it is Original