What Is AI Writing? – ChatGPT and the AI Revolution
In autumn 2022, OpenAI released ChatGPT version 3.5 to the public. A Large Language Model (LLM) trained on vast quantities of human text, ChatGPT 3.5 grasped the statistical patterns of words and sentences, predicting the most likely next word and generating replies within a chat window. While not the world’s first chatbot, it swiftly became the first to reshape the world.
Within five days of its release, ChatGPT amassed a million users; eight weeks later, 100 million. The fastest-growing software product in history. Its rapid adoption signaled the arrival of a new paradigm and ignited concerns about the authenticity of the written word itself.
The AI Black Box
At its core, ChatGPT is a software black box. Its inner workings are inscrutable. It appears almost magical. From the simple ‘spell’ of the prompt, coherent sentences emerge. It’s extraordinary. But, while it and the AI writing tools that followed have streamlined writing by making the process faster and more accessible, the concerns they create about authenticity and human expression have created an urgent need to verify whether a human or machine created a particular piece of text.
The Scrutiny of Authenticity: The Potential for False Positives
The need to verify AI-generated content led to the creation of AI content detectors, a development which carries its own consequences. False positives produced by AI content detectors can have severe repercussions, as illustrated by the experience of Robert Topinka, a senior lecturer in media studies at Birkbeck, University of London.
In the spring of 2023, while marking his undergraduate student essays, Topinka encountered a problem. Turnitin, a US-based education company that had evolved its plagiarism detection software to serve the growing need for accurate AI content detection, had flagged a promising student paper as 100% AI-generated.
The university’s policies were clear: They demanded such events be reported immediately. The consequences were potentially severe. The student could be expelled. Topkinka faced a dilemma.
The student denied using AI software to write the essay, and Topinka deliberated. He knew that the flagged essay was in a style similar to the student’s previous submissions, written before the advent of AI tools.
Topinka decided that the flagged essay was the product of an AI mistake. It was a “false positive.” Relying on his judgment, he exonerated the student. He noted it was a “high stakes call without reliable evidence” and a “distressing experience” for the student. It was also, he said, an experience “being repeated across the sector.” ((Link)
The Hunt For False Positives in AI Content Detectors
Topkina’s story highlights the central problem: the need to verify the origin and authenticity of a piece of writing. For writing, AI is transformative. The ability to input a prompt and have a language model produce a torrent of coherent, college-level sentences is a game changer.
Generative AI’s don’t just output coherent essays; they displace the reverence traditionally assigned to the act of writing itself. In so doing, they force us to re-measure the worth and value of literacy.
If anyone can produce a plausible essay, how do we know they did? If a student submits an essay to what extent can we be sure that the work is a product of their own mind?
This question that lies at the heart of the challenge content detectors attempt to decipher. And, as AI writing tools become more capable, the hunt for false positives becomes increasingly critical.
What is AI Content Detection? Assessing the Contents Coherence
AI Content detection offers a solution: disentangling the human from the machine. However, distinguishing between human-created and machine-generated text is no simple task. To do so, content detectors must analyze two key metrics: perplexity and burstiness.
Perplexity
Perplexity can be defined as a certain analog unpredictability —a subjective, human trait, based on our tendency to make mistakes and under or overcomplicate sentences. Writing that is perhaps slightly less than optimal. Of course, AI models display no such unpredictability. They construct sensible sentences with low perplexity scores. This is a measurable metric. The higher the perplexity the higher the probability it has been written by a human.
Burstiness
Our unpredictability produces another very human characteristic: variability. Our thoughts vary in length and complexity, and so do our sentences. This variability is also measurable. AI models construct more uniform sentences, following conventional grammar rules, which results in sensible (some might say boring) writing.
So, if we’ve established that there are metrics we can measure, what is the further problem? The answer lies in the occurrence of false positives.
Never 100%! The Unpredictable Accuracy of AI Content Detectors
A false positive occurs every time an AI content detector incorrectly claims an AI document was authored by a human. Turnitin, the software used by Topinka’s university, boasts a false positive rate of 1% across its detection algorithms. However, as Turnitin’s Chief Product Officer Annie Chechitelli demonstrates in this post, other circumstances and writing styles can increase the rate. For example, a document containing less than 20% AI-generated content can produce a higher rate of false positives. This is caused by the difficulty of detecting AI text housed within a mostly human-authored document.
As Chechitelli’s blog post outlines, Turnitin’s tool measures text at document and sentence levels. But, the sentence-level false positive rate is 4%. This means that four in a hundred sentences highlighted as AI-generated may have been human-written. These challenges underscore the need for high-quality content detection.
Chechitelli’s blog post reveals that of the 38.5 million college paper submissions measured by Turnitin, 3.5% contained more than 80% AI-generated content, and 9.6% reported over 20% AI writing.
Even if we allow for the fallibility of Turnitin’s measurement metrics and leave room for wider margins of error, AI-generated content is finding its way into millions of college and university essays.
This problem will only intensify as language models improve their outputs. The current limitation of content detection tools, including the potential for false positives, presents significant challenges that need to be addressed. Various AI content detectors, such as Turnitin [1], Originality.ai [2], Winston [3], GPTZero [4], and Smodin [5], are working to improve this.
How to Interpret the Results from AI Content Detectors
AI content detection is not without controversy, and has received a good deal of criticism, so for clarity, it might be useful to review how the makers of these tools view the matter.
Winston AI claim a 99.98% accuracy in predicting the influence of an AI writing tool on a piece of content. While this is an impressive result, it’s helpful to understand the methodology behind the result.
Winston AI state that their prediction method uses a probabilistic approach. This is significant, because it implies several things about how their content detection works:
- Modelling Uncertainty: Probability implies uncertainty. A thing may be ‘highly likely’ (probable) and still untrue. So, even with such a high accuracy rate there still exist margins of error that can produce false positives or negatives.
- Random Variables: Random variables exist within probabilistic reasoning models. These represent uncertain quantities. So, while a content detector can claim with a 99.8% certainty that its assumption about authorship is correct, there remains a small margin of error.
Smodin’s blog states that Tunitin’s content detection algorithms rely on breaking text up into small segments and analyzing it in portions. This is to improve accuracy and enable segmented scoring on a scale from 0-1. 1 is an indication that the content is AI-generated. At the end of this analysis, an aggregated score is provided, which is a final indication of the text being human or AI-generated.
GPTZero is using a by sentence approach, adding metrics for perplexity and burstiness. Originality AI suggests this may leave gaps, especially when the model is analyzing shorter texts.
It is crucial to understand that as impressive as these margins are none of these methods are infallible. Either at the algorithm level, or at the level of human trickery. Content detection tools are constantly improving and attempting to shrink these margins further.
As the technology improves, it is likely accuracy and reliability will continue to rise, helping those who sorely need such assurances to more effectively identify AI content.
The Proximity Problem
The false positives issue is further complicated when we add the role of proximity. More than half (54%) of false positive (human-authored) sentences are directly adjacent to AI-generated ones. 26% are only two sentences away.
The Need for Broader Discussion
While Turnitin is a market leader, it is not the only one. Its competitors claim similar or lower false-positive rates. However, we must acknowledge that AI detection software is not perfect and is unlikely to ever to be so. This does not negate its usefulness but it does underline the need for broader discussion.
As content detectors become more widespread, relying on a single tool running proprietary algorithms in isolation will not suffice. We must look for better solutions. Also, there is one other important point to consider. Just as we are concerned about the occurrence of false positives, we should also consider the issue of false negatives — instances where AI-generated content is incorrectly flagged as human.
False Negatives: The Other Side of the Coin
A false negative is the opposite of a false positive. Machine-generated content identified as human-authored. This can occur for many reasons, but the main reason is the evolving capabilities of AI models and the increasingly human quality of their outputs.
Each successive language model upgrades on its predecessor. ChatGPT 3.0 was trained on around 170 billion parameters, while the parameter count for version 4.0 is rumored to be trillions.
As AI models acquire deeper training knowledge, they produce more convincing outputs. Moreover, the jump from GPT-3 to GPT-4 was a quantum leap in reducing factually incorrect statements, further blurring the line between human and machine writing.
This constant evolution of AI content generators means that any checker will become less reliable without continuous improvement. As new language models appear, the need to update the software’s quality grows.
Tips on how to avoid False positives.
So, how do we ensure the value and validity of writing in a world where it is difficult to tell the difference?
- Use A Human Editor: There is no substitute. As much as we like to believe we are replaceable, we are not. A human author brings not only technical skills but a degree of empathy, instinct, and intuition.
- Review the Contents Coherence and Tone: Human-written words hit differently. They have a natural flow and structure lacking in their synthetic counterparts, but be careful; there are also incoherent human writers!
- Tone: Attitude or emotion requires a human writer’s unique voice. Personality, understanding, and intuition matter. AI performs impressive acts of stochastic mimicry but it struggles to maintain a truly human tone.
- Use AI Detector tools such as Content Guardian. 🙂
How Should AI Detectors Be Used?
The question of how AI content detector should be used might be considered an ethical one. Because without the ability to predict with complete certainty it is also difficult to convict with confidence. The gray areas, such as the occurrence of fals positives and negatives will always cast doubt on a prediction models reliability.
The responsible use of such tools might suggest they be employed as guides. Tools that can indicate probability and potential but are not definitive. In that way they can be used to guide conversation.
Can AI Detectors be tricked?
Undoubtedly. Though they continue to improve. The methods employed to deceive them will advance in tandem. Currently using tools like Quillbot to paraphrase and add burstiness is popular. But like any technological arms race, content detection tools are already beginning to account for this.
The Need for an Aggregate AI Content Detector Tool
Relying on a single model as your sole line of defense creates a single point of vulnerability. If any one tool struggles to keep pace, your journalists, students, or creatives could also be compromised.
No one model can do it all. The wide array of styles and the diversity of human writing make it impossible for any single checker to keep pace. The best solution is a content detection aggregator—a tool that combines the strengths of multiple content detectors. This can provide a more reliable assessment of a text’s authenticity.
All of which begs the question. “Well if the content detectos is imperfect, and can be fooled, why bother?
To which the answer is, that if we value human authenticity over machine generated assembly we must preserve the ability to highlight the areas where that quality is being compromised.
Enter Content Guardian
Content Guardian has been designed to integrate the services of the leading AI content-checking tools on the market, moulding them into an ‘all-star team.’ Our approach is not about catching AI content but promoting and upholding the clarity and insight unique to genuine human writing.
We remove the black box of single provider reliance and aggregate their knowledge, weighing their proficiency, evaluation methodologies, performance, and credibility against the language models available. Doing this lets us leverage each of their strengths and benefits.
Content Guardian runs the results from each content checker and provides an overall score. We call this “The Content Guardian Confidence Score.” You receive a holistic weighted measurement. One that indicates if a piece of content is human-generated. By combining the strengths of multiple tools, Content Guardian offers unparalleled confidence to anyone working in content creation.
Looking Ahead
The rapid advancements of AI threaten to transform our cognitive labour in ways that mirror the impacts of the Industrial Revolution. This transformation will occur with incredible speed. It will disorient many and require brave approaches to solve unprecedented challenges.
Old models of content creation and verification that have stood untouched for a century or more may disappear. It is an enormous opportunity that needs to be managed with exceptional care to avoid widespread displacement and disruption.
In an age when we can mechanize almost every cultural commodity and create generative versions of ourselves, confidence in the authenticity and believability of our creations is paramount. Our capacity to use language to communicate our ideas with language is perhaps our most profound gift. The future will be considerably more secure if we can develop tools and strategies to verify authenticity and enable trust.
This requires solutions like Content Guardian and a broader discussion about AI ethics across our society. If we can have that conversation, the benefits will be enormous, and the future will look much brighter.
