The Times Australia
The Times World News

.

We pitted ChatGPT against tools for detecting AI-written text, and the results are troubling

  • Written by Armin Alimardani, Lecturer, University of Wollongong
We pitted ChatGPT against tools for detecting AI-written text, and the results are troubling

As the “chatbot wars[1]” rage in Silicon Valley, the growing proliferation of artificial intelligence (AI) tools specifically designed to generate human-like text[2] has left many baffled[3].

Educators in particular are scrambling to adjust to the availability of software that can produce a moderately competent essay on any topic at a moment’s notice. Should we go back to pen-and-paper assessments? Increasing exam supervision? Ban the use of AI entirely?

All these and more have been proposed. However, none of these less-than-ideal measures would be needed if educators could reliably distinguish[4] AI-generated and human-written text using AI detector tools like ZeroGPTt.

We dug into several proposed methods and tools for recognising AI-generated text. None of them are foolproof, all of them are vulnerable to workarounds, and it’s unlikely they will ever be as reliable as we’d like.

Perhaps you’re wondering why the world’s leading AI companies can’t reliably distinguish the products of their own machines from the work of humans. The reason is ridiculously simple: the corporate mission in today’s high-stakes AI arms is to train ‘natural language processor’ (NLP) AIs to produce outputs that are as similar to human writing as possible. Indeed, public demands for an easy means to spot such AIs in the wild might seem paradoxical, like we’re missing the whole point of the program.

A mediocre effort

OpenAI – the creator of ChatGPT – launched a “classifier for indicating AI-written text[5]” in late January.

The classifier was trained on external AIs as well as the company’s own text-generating engines. In theory, this means it should be able to flag essays generated by BLOOM AI[6] or similar, not just those created by ChatGPT.

We give this classifier a C– grade at best. OpenAI admits it accurately identifies only 26% of AI-generated text (true positive) while incorrectly labelling human prose as AI-generated 9% of the time (false positive).

OpenAI has not shared its research on the rate at which AI-generated text is incorrectly labelled as human-generated text (false negative).

A promising contender

A more promising contender is a classifier created by a Princeton University student during his Christmas break.

Edward Tian[7], a computer science major minoring in journalism, released the first version of GPTZero[8] in January.

This app identifies AI authorship based on two factors: perplexity and burstiness. Perplexity measures how complex a text is, while burstiness compares the variation between sentences. The lower the values for these two factors, the more likely it is that a text was produced by an AI.

We pitted this modest David against the goliath of ChatGPT.

First, we prompted ChatGPT to generate a short essay about justice. Next, we copied the article – unchanged – into GPTZero. Tian’s tool correctly determined that the text was likely to have been written entirely by an AI because its average perplexity and burstiness scores were very low.

GPTZero measures the complexity and variety within a text to determine whether it is likely to have been produced by AI. GTPZero[9]

Fooling the classifiers

An easy way to mislead AI classifiers is simply to replace a few words with synonyms. Websites offering tools that paraphrase AI-generated text for this purpose are already cropping up all over the internet.

Many of these tools display their own set of AI giveaways, such as peppering human prose with “tortured phrases[10]” (for example, using “counterfeit consciousness” instead of “AI”).

To test GPTZero further, we copied ChatGPT’s justice essay into GPT-Minus1[11] — a website offering to “scramble” ChatGPT text with synonyms. The image on the left depicts the original essay. The image on the right shows GPT-Minus1’s changes. It altered about 14% of the text.

GPT-Minus1 makes small changes to text to make it look less AI-generated. GPT-Minus1[12]

We then copied the GPT-Minus1 version of the justice essay back into GPTZero. Its verdict?

Your text is most likely human written but there are some sentences with low perplexities.

It highlighted just one sentence it thought had a high chance of having been written by an AI (see image below on left) along with a report on the essay’s overall perplexity and burstiness scores which were much higher (see image below on the right).

Running an AI-generated text through an AI-fooling tool makes it seem ‘more human’. GPTZero[13]

Tools such as Tian’s show great promise, but they aren’t perfect and are also vulnerable to workarounds. For instance, a recently released YouTube tutorial explains how to prompt ChatGPT to produce text with high degrees of – you guessed it – perplexity and burstiness.

Watermarking

Another proposal is for AI-written text to contain a “watermark” that is invisible to human readers but can be picked up by software.

Natural language models work on a word-by-word basis. They select which word to generate based on statistical probability.

However, they do not always choose words with the highest probability of appearing together. Instead, from a list of probable words, they select one randomly (though words with higher probability scores are more likely to be selected).

This explains why users get a different output each time they generate text using the same prompt.

One of OpenAI’s natural language model interfaces (Playground) gives users the ability to see the probability of selected words. In the above screenshot (captured on Feb 1, 2023), we can see that the likelihood of the term ‘moral’ being selected is 2.45%, which is much less than ‘equality’ with 36.84%. OpenAI Playground[14]

Put simply, watermarking involves “blacklisting” some of the probable words and permitting the AI to only select words from a “whitelist”. Given that a human-written text will likely include words from the “blacklist”, this could make it possible to differentiate it from an AI-generated text.

However, watermarking also has limitations. The quality of AI-generated text might be reduced if its vocabulary was constrained. Further, each text generator would likely have a different watermarking system – so text would next to checked against all of them.

Watermarking could also be circumvented by paraphrasing tools, which might insert blacklisted words or rephrase essay questions.

An ongoing arms race

AI-generated text detectors will become increasingly sophisticated. Anti-plagiarism service TurnItIn recently announced[15] a forthcoming AI writing detector with a claimed 97% accuracy.

However, text generators too will grow more sophisticated. Google’s ChatGPT competitor, Bard, is in early public testing[16]. OpenAI itself is expected[17] to launch a major update, GPT-4, later this year.

It will never be possible to make AI text identifiers perfect, as even OpenAI acknowledges[18], and there will always be new ways to mislead them.

As this arms race continues, we may see the rise of “contract paraphrasing”: rather than paying someone to write your assignment, you pay someone to rework your AI-generated assignment to get it past the detectors.

There are no easy answers here for educators. Technical fixes may be part of the solution, but so will new ways of teaching and assessment (which may including harnessing the power of AI).

We don’t know exactly what this will look like. However, we have spent the past year building prototypes of open-source AI tools for education and research in an effort to help navigate a path between the old and the new – and you can access beta versions at Safe-To-Fail AI[19].

References

  1. ^ chatbot wars (www.independent.co.uk)
  2. ^ human-like text (plato.stanford.edu)
  3. ^ left many baffled (www.nytimes.com)
  4. ^ reliably distinguish (www.theguardian.com)
  5. ^ classifier for indicating AI-written text (openai.com)
  6. ^ BLOOM AI (www.notion.so)
  7. ^ Edward Tian (www.npr.org)
  8. ^ GPTZero (gptzero.me)
  9. ^ GTPZero (gptzero.me)
  10. ^ tortured phrases (www.nature.com)
  11. ^ GPT-Minus1 (www.gptminus1.com)
  12. ^ GPT-Minus1 (www.gptminus1.com)
  13. ^ GPTZero (gptzero.me)
  14. ^ OpenAI Playground (beta.openai.com)
  15. ^ TurnItIn recently announced (www.turnitin.com)
  16. ^ early public testing (blog.google)
  17. ^ expected (www.theverge.com)
  18. ^ OpenAI acknowledges (youtu.be)
  19. ^ Safe-To-Fail AI (safetofailai.streamlit.app)

Read more https://theconversation.com/we-pitted-chatgpt-against-tools-for-detecting-ai-written-text-and-the-results-are-troubling-199774

Times Magazine

Headless CMS in Digital Twins and 3D Product Experiences

Image by freepik As the metaverse becomes more advanced and accessible, it's clear that multiple sectors will use digital twins and 3D product experiences to visualize, connect, and streamline efforts better. A digital twin is a virtual replica of ...

The Decline of Hyper-Casual: How Mid-Core Mobile Games Took Over in 2025

In recent years, the mobile gaming landscape has undergone a significant transformation, with mid-core mobile games emerging as the dominant force in app stores by 2025. This shift is underpinned by changing user habits and evolving monetization tr...

Understanding ITIL 4 and PRINCE2 Project Management Synergy

Key Highlights ITIL 4 focuses on IT service management, emphasising continual improvement and value creation through modern digital transformation approaches. PRINCE2 project management supports systematic planning and execution of projects wit...

What AI Adoption Means for the Future of Workplace Risk Management

Image by freepik As industrial operations become more complex and fast-paced, the risks faced by workers and employers alike continue to grow. Traditional safety models—reliant on manual oversight, reactive investigations, and standardised checklist...

From Beach Bops to Alpine Anthems: Your Sonos Survival Guide for a Long Weekend Escape

Alright, fellow adventurers and relaxation enthusiasts! So, you've packed your bags, charged your devices, and mentally prepared for that glorious King's Birthday long weekend. But hold on, are you really ready? Because a true long weekend warrior kn...

Effective Commercial Pest Control Solutions for a Safer Workplace

Keeping a workplace clean, safe, and free from pests is essential for maintaining productivity, protecting employee health, and upholding a company's reputation. Pests pose health risks, can cause structural damage, and can lead to serious legal an...

The Times Features

The Role of Your GP in Creating a Chronic Disease Management Plan That Works

Living with a long-term condition, whether that is diabetes, asthma, arthritis or heart disease, means making hundreds of small decisions every day. You plan your diet against m...

Troubleshooting Flickering Lights: A Comprehensive Guide for Homeowners

Image by rawpixel.com on Freepik Effectively addressing flickering lights in your home is more than just a matter of convenience; it's a pivotal aspect of both home safety and en...

My shins hurt after running. Could it be shin splints?

If you’ve started running for the first time, started again after a break, or your workout is more intense, you might have felt it. A dull, nagging ache down your shins after...

Metal Roof Replacement Cost Per Square Metre in 2025: A Comprehensive Guide for Australian Homeowners

In recent years, the trend of installing metal roofs has surged across Australia. With their reputation for being both robust and visually appealing, it's easy to understand thei...

Why You’re Always Adjusting Your Bra — and What to Do Instead

Image by freepik It starts with a gentle tug, then a subtle shift, and before you know it, you're adjusting your bra again — in the middle of work, at dinner, even on the couch. I...

How to Tell If Your Eyes Are Working Harder Than They Should Be

Image by freepik Most of us take our vision for granted—until it starts to let us down. Whether it's squinting at your phone, rubbing your eyes at the end of the day, or feeling ...