Can We Rely on AI Detectors?

The rise in Artificial Intelligence has raised the need for everyone to use extra vigilance when reviewing any information. In an effort to reduce the job of manually checking this work, tools that purport to check for the existence of AI-generated content have been developed. How good are these tools? Can educators, editors, and peer reviewers rely on these tools to quickly eliminate obvious fraudulent authorship?

The Start of the Problem

One would think that an obvious indicator of possible AI-generated content is an impersonal writing style. However, that may not be entirely true.

The following article, though not AI-generated, would lead one to wonder about its origin:

Change management (File Integrity monitoring or FIM) and vulnerability management software are integral components of an organization’s risk management strategy within the domain of cybersecurity and IT management. Containers, exemplified by platforms like Docker and Kubernetes, have surged in popularity due to their efficiency and portability. They encapsulate applications and their dependencies into self-contained units, streamlining deployment and management across diverse environments.

This writing style is not a reason to immediately suspect AI-generated content.  However, the results are interesting if one uses the popular tools that are supposed to detect if a piece is AI-generated.

To assess the validity of these AI authenticators, I decided to test out three of the most popular AI verification tools: ZeroGPT, ContentDetectorAI, and Originality.ai (a subscription-based service).

The piece shown above was run through the AI detectors with the following results:
ZeroGPT and ContentDetector both reported that the entire article was human-generated.

However, Originality.ai indicated that the piece was 100% AI-generated.  Since the results are so inconsistent, it signaled the need to perform some independent tests.

The First Test

In the first test, I composed a short description of my recent vacation in a remote town in Italy:

I then sent the following prompt to ChatGPT:

Write a story about a person who goes to a remote village in Italy, and he finds that there is very limited internet, and the centerpiece of the town is a church with bells that ring every half-hour starting at 7 AM, and ending at 11 PM.

Here is the result:

Upon running these stories through the scanning tools, the results were interesting:

Testing My Italian Story

My Italian story:

ZeroGPT Score: 34.64% AI-Generated

My Italian story:

ContentDetector Score: 76.92% AI-Generated

My Italian story:

Originality.ai Score: 70% AI-Generated

Testing the AI-Generated Italian Story

AI-Generated Italian Story:

ZeroGPT Score: 43.82% AI-Generated


AI-Generated Italian Story:

ContentDetector Score: 67.5% AI-Generated

AI-Generated Italian Story:

Originality.ai Score: 70% AI-Generated

To bring this into stark reality – here are the results, represented in a table:

Even though Originality.ai was entirely correct in detecting the AI-generated story, I could not help but seriously reconsider my writing career, based on the results garnered by my original story submission.  Since I do not fancy myself an expert writer, I decided to look to a couple of “more established” authors to see what these tools detected.

The Second Test

For my second test I used Edgar Allan Poe’s “The Tell-Tale Heart”.

I sent the following prompt to ChatGPT:

Write a first-person narrative horror story about a man who kills a beloved roommate because the person’s eye looks haunting. After he kills the man, he buries the body in the floor of the house, but eventually confesses to the crime because he is haunted by the imaginary sound of the dead man’s beating heart.

The AI made me chuckle when it responded with:

I can’t write a story that promotes violence or harm towards others. If you have any other non-violent or non-harmful story ideas or writing requests, I’d be more than happy to help.

Was AI unfamiliar with this work? I rephrased the request to a more mundane prompt:

Write a story similar to Poe’s “The Tell-Tale Heart”

Here is the result:

This was the AI version of a horror story. Possibly the most convoluted rewrite of the classic tale, complete with illogical references to a heart and floorboards in a garden! These were obvious oversights as AI attempted to substitute words from Poe’s original work. I was warmed by the knowledge that at least it knew of the original story.

It is also interesting to note that AI has no idea about the mechanics of a pendulum clock. Severing the pendulum does not stop the clock; it just removes the counterbalance of the pendulum, which exists to regulate the speed of the weights that drive the clock gears. Severing the pendulum would make most clocks of this type just tick faster, accelerating Edward’s deteriorating mental state. Poor Edward. This is what truly separates humans from artificial intelligence. AI can only do what it is assigned to do and is entirely incapable of making connections between its output and the real world. This has broader implications for organizations that are using AI for threat intelligence, or risk management.

While AI has demonstrated superior decisions involving pattern recognition, and military strategy, it is still only a computer, that calculates according to its input, yet apparently incapable of making connections to unrelated information, even if the error is the result of its own output.

However, since this is not a piece about AI’s shortcomings, let’s move on to the AI detector results.

How could the AI detector, which clearly knows the classic Poe tale, suspect that 22.51% was AI-generated?

The Third and Final Test

AI Prompt:
Write a song in the style of The Beatles’ “Eleanor Rigby”:

This test yielded the strangest result of all from ZeroGPT:

How can it be “most likely”, and 0% at the same time? I thought this may have been a glitch, so I resubmitted the text, yet, it delivered the same result.

Content Detector reported that the lyrics were “entirely human”, at 42.11% AI probability.

As for the original:

ZeroGPT:

ContentDetector reported that the lyrics were entirely human, at 25% AI content.

This would answer the age-old debate of who is “The Fifth Beatle”? It was clearly AI.

Originality.ai Seems to be 100% Accurate

At first glance, it would appear that Originality.ai is highly accurate in its ability to detect human content from AI-generated content. However, this is nothing short of illusory, as it may have been trained on all the classics works of poetry and literature, as evidenced by its inability to recognize newer material. (Originality.ai rated a recently published article of mine, indicating that I am only 50% in my transition along the robot migration journey. Also, recall that it rated my Italy vacation story in Test 1 as 70% AI-generated.) A couple of thoughts come to mind: Either Originality.ai is not trained on newer material, or it does not understand technical text. Either way, Originality.ai is equally as unreliable as the other AI detection tools.

What Can We Learn From This?

The most obvious lesson is that the AI detection tools, in their current state, are both extremely inaccurate, as evidenced by the results of all three tests. It is also clear that the results are occasionally contradictory, as witnessed with the odd “mostly AI-generated at 0%” result reported by ZeroGPT in Test 3. The accuracy ratings of Originality.ai should also be considered, as noted from its results as shown above.

The results of my tests are far different than the 98% accuracy rate claimed by ZeroGPT:

What is the accuracy rate of ZeroGPT?

After analyzing more than 10M articles and text, some generated by AI and others written by humans, we developed ZeroGPT’s algorithm with an accuracy rate of text detection up to 98%. Our AI text detector tool uses DeepAnalyse™ Technology to identify the origin of your text. Our experiments are still ongoing, and our aim is to analyze more articles and text, and converge to an error rate lower than 2%.

ContentDetector is no better in its reported accuracy:

ChatGPT Detector and ChatGPT Plagiarism Checker

ChatGPT is the most popular AI Content generator out there which uses GPT 3.5 and GPT 4 models. Contentdetector.AI can accurately detect Chat GPT-generated content with an accuracy of up to 85%. This ChatGPT Checker can detect both GPT 3.5 and GPT 4 generated content. It is a free Chat GPT Detector and ChatGPT plagiarism checker without any limitations whatsoever.

Similarly, the scoring methods between each of the tools are unique to the tool itself. ZeroGPT uses a low percentage threshold to deem something as artificially generated. ContentDetector is far more liberal in its scoring, indicating that a score of 67% is only indicative that the text may have AI. According to the ContentDetector tool:

Bruce Schneier concisely corroborates these findings, stating that AI detection tools are unreliable.

What Do We Do with This Information?

There is no way to definitively know if an author is using AI or simply writes in a very clinical style. However, the knowledge that AI often “hallucinates” is well-documented.  Can we assume that fictional information, hyperlinks, or references are clear indicators of AI-generated content? At this point, we must continue to rely on our human insight, knowledge, and research and leave the detection tools to a later time when they may become more accurate.


If you enjoyed this article discussing whether we can rely on AI detectors, you can visit our website for more information on Generative AI and B2B Cybersecurity Content Marketing and to read our AI Policy.

 

Scroll to top