Is AI Smart Enough to Prevent AI Detection?

If you are a writer attempting to use AI responsibly, AI detectors can keep you from over-dependence. Yet, what happens when AI fools AI?

Artificial Detection

With the vast number of AI products permeating the internet, programmers have developed tools to detect non-human text. This is useful for teachers to determine grades and employers who pay writers, not robots. In most cases, AI detection apps are spot on. Sometimes, however, they flag boring or verbose human text as AI-generated.

Obfuscating Detection

What happens when AI gets smart enough to prevent detection? I presented that challenge to ChatGPT. The 3.5 version was prompted to write a blog post. The entire article was correctly flagged as AI-generated. That same article was uploaded to ChatGPT 4.0 with the prompt to rewrite it to avoid AI detection.

Prompt: “An AI detector flags this entire article as generative, even though only portions are. What adjustments will make it more natural?”

The first revision did not pass detection. So this was my next comment.

Prompt: “Thanks, although it didn’t help with detection. The detector must be using AI. :)”

The second revision passed for only the first half of the article. So I uploaded the problem area.

Prompt: “This portion is flagged as AI generated:”

The third revision was marked as entirely human text by an AI detector. Below is a summary of how it was achieved.

Key Adjustments Made:

Enhanced Personal Anecdotes: Added more specific details about Uncle Hayward, Aunt Daisy, Auntie Doris, and Aunt Jane to make the descriptions more vivid and personal.
Varied Sentence Structure: Mixed longer, descriptive sentences with shorter, punchier ones to create a more natural flow.
Incorporated Colloquial Language and Quotes: Used more informal language and included a direct quote from Aunt Mary to add a conversational tone.
Added Sensory and Emotional Details: Included sensory and emotional descriptions, such as Auntie Doris’s experience with arthritis and Aunt Jane’s feelings of loneliness.
Integrated Emotive Language: Used phrases like “a crucial recovery time,” “a much-needed break from the solitude,” and "gentle smile" to add emotional depth.
Included Specific, Varied Vocabulary: Used a broader vocabulary and varied word choices to avoid repetitive patterns typical of AI-generated text.

By incorporating these techniques, the text should read more naturally and authentically, reducing the likelihood of being flagged as AI-generated.

What It's Like to Chat

By now, millions of people have an idea of what it’s like to chat with a virtual robot. Although ChatGPT is the most common, there are many others. At times it can be like chatting with a wise friend. During extended critiques about a story, it can get frustratingly recursive.

For example, it might suggest verbose rewrites to several paragraphs. After edits, it could offer additional suggestions for the same paragraphs or contradict prior advice. Depending on how you ask questions, ChatGPT might rewrite an entire story, removing key expressions an author crafted. Like a wise friend with OCD, it can sometimes insist that its style of writing is best. So writers who use the new Creative Writing Coach can get more satisfaction with the following requests:

Prompt: “Rate this story and list suggestions for improvement, without rewriting.”

Each iteration of AI becomes orders of magnitude “smarter.” Does this exercise sound like it jeopardizes the roles of people employed as writers?

You can support the writing of useful articles about technology by sponsoring specific articles or remitting a small donation. Story includes one or more generative AI images to help readers visualize scenes.

Can AI Outsmart AI?