When Avoid AI Detection is turned “on” Article Forge content can pass AI detectors as human written at the same rate as content written by actual humans.
To test this, we ran content from the most popular AI writers and human content currently ranking #1 on Google through AI detectors and compared the results. Read the case study here.
However, it is important to understand how AI detectors work and the limitations of both AI content detectors and Article Forge.
AI Detectors
Without going into too much technical detail, AI detectors give their results as the probability that portions of the text could be generated by AI (called perplexity). So a 50% AI generated score means that there is a 50-50 chance that portions of the text could be generated by AI. Any result that is no better than chance should not be used to reject or penalize a piece of content.
While every AI content detector is different, they will all incorrectly classify human text as AI generated in some cases.
This is because AI detectors look for patterns that generative AI models tend to write with. However, some of these patterns can also be found in human-written content.
For example, a major limitation of AI detectors is judging factual or general text. Let’s say a piece of content included the sentence “The first ten digits of pi are: 3.1415926535.”, both a human and an AI writer would (hopefully) write the exact same thing. However, AI detectors would see that a machine could have written this and would lean to the side of calling it “AI-generated”.
Article Forge
Similar to the example above, if you request an article that should contain many factual statements, Article Forge will prioritize accuracy over passing AI detection. In addition, if a topic is very general, Article Forge will write generally about it. In these cases, Article Forge content can come back with higher AI generated probabilities.
How to judge AI content detection results
Therefore, when looking at AI content detection results, it’s important to:
- Prioritize the quality of the content. When you read through your text, does it flow well and is it useful? At the end of the day, humans want human-readable, valuable content. So, it is more important to judge the overall quality of a piece of content than use any one scoring factor.
- Consider the probability of the score. If a score is in the ~40-60% (or rated “unsure”, “possibly”, “parts could be” etc) range, that means that the detector does not know how the text was created. In general, this means the content is safe.
- Look at trends in scores. Even human written articles will be detected as AI at times, but as long as your overall detection results air on the side of human-written, your content will be safe.
What should you do if you are getting high “AI generated” scores
If you have gone through the above steps and determined that your content is getting high probabilities of AI generated content too often, there are a few steps you can take:
- Make sure the content topic is not too generic. You can incorporate more unique angles into your content by adding unique points to your AI writer’s prompts or adding in unique perspectives yourself. This will also improve the overall quality and usefulness of your content.
- Run the rewrite through WordAi. This is not always recommended as Article Forge already wrote the text to avoid AI patterns as much as possible the first time around without decreasing the value of the text. However, in some cases, this can help content get more human scores.
AI Detector Messaging Around Limitations
To give you a more detailed sense of some of the different AI detectors' limitations, we included their messaging below:
OpenAi Detector Limitations:
Image from OpenAi blog
Originality.ai Description and Limitations:
Image from Originality.ai
GPTZero Warning:
Image from GPTZero main page