Content that is being created by artificial intelligence is continuing to fill the internet, but AI checkers and detection may not be the most accurate way to tell if someone is using an AI to generate their blog, website pages, or emails. Many of our clients have been seeing false positives from common AI detectors, including ones we use such as Grammarly, Copyscape, ZeroGPT, and more. Our team even asked an AI bot how it felt about AI detectors, and it provided an interesting response. This blog will dive deeper into how AI checkers work, what they look for, and when they may not be accurate.
How Do AI Detectors Work?
Most detectors are based on language models similar to the ones used in AI writing tools. The language model looks at the content and asks itself, “Is this something I would have written?” If the answer is yes, it concludes that the content must be AI-generated.
However, there is more that goes on behind the scenes when an AI detector reads your pasted content. Here are some models that AI looks for and what to watch for when writing your own content.
Low Perplexity
Perplexity refers to how unpredictable the text is, meaning how likely it is to confuse the average person. AI uses low perplexity, making it easier to read, while humans tend to write with higher perplexity or creative language. For example, a low perplexity sentence may be “I couldn’t get to sleep last night.” While a higher perplexity sentence may be “I couldn’t get to sleep last summer on nights when it was very hot.” AI detectors will view lower perplexity sentences as an indication that the content is written by an AI. For businesses wanting to target different audiences, lower perplexity sentences may be necessary in order for your content to be relevant. This is why AI detectors are not always accurate at detecting human over AI.
Variation Of Sentence Length
Another model used during AI detection is burstiness, which refers to sentence structure and length. Content with little variation in sentence structure tends to be labeled as AI-generated, especially because of the low perplexity writing that is common with AI. The short sentence structure is often why readers feel that AI-written text is monotonous, because it is easy to predict what word will come next. Well-trained writers know that keeping sentences and paragraphs short is easier for our brains to read. But this could backfire when putting content into an AI detection service.
What Are False Positives
So what does all of this mean for businesses, marketing agencies, or freelancers? Well, it is common for AI detectors to produce false positives and false negatives. They are not truly accurate and cannot predict what is written by AI or what is written by a human trying to help content rank in a changing digital world.
Here are some examples from research done by our team.
Our account manager had AI write a small amount of text about a city. She then put it in an AI detector and was told that it was written by a human. After this result, she and Gemini proceeded to have a very humorous conversation.
Interesting. So even the AI detectors we are using to prove that someone is sticking with their morals about being AI-free can’t be relied on?
Another example: I put this blog into a few different AI detectors to see if I would get varying results. Grammarly stated that what I had written was 100% human. ZeroGPT said there was 17% AI text detected. Copyleaks also determined my blog was 100% human.
I decided to test another piece of content, just to really show how widely the results can vary from person to person. I pulled a blog from our site, written by another content writer on our team; here were the results:
- Grammarly: 16% of text is AI
- ZeroGPT: 25% AI
- Copyleaks: 100% human
Write Content With Search Intent In Mind
While it is concerning to see that your well-researched blog or hard work has been flagged as AI, the best way to provide results to your readers and clients is to write with their audience in mind. Our marketing team understands that we are catering to different industries and reading levels, which makes our content more unique than AI-generated responses. When writing a blog about a complex topic or answering a common question your customers have, the best way to ensure the content does well is to truly answer the question in a digestible way.
PDM writers focus on creating blogs, webpages, location pages, and emails that are easy to read by your audience. While we want your website to build authority and pages to rank well on search engines, providing informative content to your audience is our ultimate goal. Consistently adding human-written content to your website is the best way to stay relevant.
Detecting AI Writing Manually
Many of us have seen many different AI content across the internet, including blogs, social media captions, and marketing emails. Instead of using online AI detectors, readers may want to try to determine if something they are reading is AI or human. Here are some identifying features you can spot while reading content.
- Overly polite language or sounding like a helpful assistant
- Hedging phrases like “X is widely regarded as…” or “some may say”
- Inconsistent voice
- Incorrect citations or misinformation
- Frequent bullet points, especially when they repeat the same points made earlier in the writing
- Predictable and few surprises in structure or word use
Content Developed By PDM
Our team has in-house writers who are creating human content every day. AI-generated content is not being produced by our team or being edited to look like a human-written or researched blog. While we don’t use these tools to create content, we do understand how they are being used for marketing and search engine results. Our team constantly adjusts to provide content that ranks well with Google while also being used in AI responses, helping your website to build authority. If you are ready to level up your marketing game, reach out to us to build a custom website that will continue to rank in this changing world.