Introducing Natural Language Explanations in Advanced Scan
Today, we take a major step forward in helping reviewers better identify and understand signs of AI or human writing. Our latest update to the Advanced Scan feature shows what AI-generated text looks like as well as the reasoning behind each detection.
At GPTZero, our mission is to protect what makes writing uniquely human. With AI writing tools rapidly evolving, the ability to express genuine human thought matters more than ever. Today, we take a major step forward in helping reviewers better identify and understand signs of AI or human writing. Our latest update to the Advanced Scan feature shows what AI-generated text looks like as well as the reasoning behind each detection.
From Sentence highlights to Human Understanding
Our Advanced Scan feature already helps our users identify which sentences disproportionately contribute to AI-ness score - highlighting the excerpts that have the greatest impact when it comes to flagging your text as AI or human written. But can we go further? With our latest release, we introduce one more layer of transparency by not only showing you which sentences stand out, but also explaining why.
GPTZero leads the way with the first detection tool that explains, in plain language, why your text might resemble AI or human writing. For each important sentence under the advanced scan view, we now provide explanations based on carefully determined Human and AI text features detailing potential contributing reasons.
How we built this: Mining Meaningful Differences
The first step to achieving an explainable system is to unpack the characteristics or text features that are commonly associated with AI-written text and separate them from the human-written documents. To do this, we deep-dived into our large corpus of AI-generated and corresponding Human written sources, and used powerful state-of-the-art machine learning techniques to extract core features that distinguish AI from Human writing.
We then grouped these features into meaningful clusters - capturing patterns such as but not limited to-vocabulary, literary devices, grammatical structure and voice. Spurious and ambiguous features were identified and discarded. After identifying these meaningful patterns we trained a classifier model to differentiate Human and AI text solely based on these features. We were able to achieve an accuracy of around 90% validating the representative power of the identified feature sets (both AI and Human).
What we found: Distinctive Traits Found in AI and Human Text
Here is a snapshot of some of the top AI and Human features based on our text analysis:
Table 1: Prominent AI and Human text features and their associated feature tags
Example in Action: Document with Natural Language Insights
Figure 1 shows our current advanced scan view for an example mixed text where the first paragraph is a human-written excerpt from an article discussing climate change and the second paragraph is AI-generated using gpt-4o.
Figure 3: (a) Example of our Advanced Scan View with document-level features and descriptions. (b) Example of our Advanced Scan View with sentence-level explanations.
As seen in Figure 3, you can now observe features and their descriptions/explanations at both document and sentence-levels. Specifically, for this example our system identified AI features such as “Mechanical Precision”, “Sophisticated Clarity” and “Robotic Formality” and human features such as “Contemporary Relevance”, “Informative Analysis” and “Journalistic Style”.
Figure 4: Sentence-level human features and corresponding explanations for high-impact Human sentences
Users can now click on the collapsible drop-down at the sentence level to see sentence-specific explanations tailored to answer the question: How are the identified AI feature (e.g., Figure 5) and human feature (e.g., Figure 4) reflected in the selected sentence?
Figure 5: Sentence-level human features and corresponding explanations for high-impact AI sentences
The human writing is noticeably distinct in terms of overall style compared to the AI-generated text. For example, the sentences use a narrative structure and diverse sentence structures blending academic concepts with a human touch as indicated in Figure 4. It also clearly indicates contemporary relevance by comparing conclusions in previous publications (New Yorker, 1989) with current happenings (flooding and wild fires in North Carolina). Features such as these and other prominent indicators are captured by our system.
On the other hand the AI sentences are run-on, with multiple comma separated clauses and unnecessary jargon inducing a robotic tone in the writing (Figure 5). For example, phrases such as “Hydrological Extremes” and technical terms like “pluvial flooding” artificially induce technical jargon into the writing which is captured by features such as “Sophisticated Clarity” and “Mechanical Precision”. Additionally, AI also has the tendency to produce intricate long comma separated clauses which are captured by features such as “Run-On Sentences” and “Complex Sentences”.
Together these features enable reliable differentiation between authentic human writing and AI generations.
What’s Ahead?
Our feature is currently in beta. As the LLM landscape and human-writing styles continue to evolve, we are continuously refining our system to update our features to most closely match the adaptations in various writing styles. Our goal is to give you the most important and actionable insights to distinguish authentic human writing from AI-generated text and improve the overall transparency of our AI detection system.
