Research shows that even if words are changed, AI-specific writing 'structures' cannot be hidden. A recently developed technique called 'SlopShape' detects AI content with 98% accuracy based solely on information placement and logical flow.
Imagine this: You have a favorite newsletter or blog you read every morning. You’re reading it as usual, only to find out that more than half of the piece was written by an AI, not a human. How would that make you feel?
Until now, we have mainly focused on ‘words’ to identify AI-written text. However, as the developers behind AI have become smarter, existing detection technologies have quickly become obsolete when instructed to “write like a human” or when words are slightly twisted. We have now entered a new era where we can see the ‘skeleton’ within the text, not just its outer appearance.
Why is this important?
The internet is an ocean of information. However, with generative AI recently flooding the web with massive amounts of content, it has become extremely difficult to distinguish what is genuinely written with human thought and what is a mechanical result generated by AI. Source: Responsible Detection and Mitigation Framework
Previous detection methods were like “fishing rods that only catch specific words.” They exhibited ‘brittleness,’ where the detector would fail to work if the AI’s word patterns changed even slightly. Source: SlopShape Research Paper However, the situation changes if a method that analyzes the ‘structure’ of the text is introduced. This is an important turning point that could change the standards of trust we use to judge the authenticity of content when consuming online information.
Easy to understand: From the wrapping of ‘words’ to the essence of ‘structure’
The secret to distinguishing AI-written text from human-written text can be easily explained with an analogy.
Simply put, think of ‘LEGO.’ Even if a LEGO castle built by a human and one assembled automatically by a machine look similar at first glance, the order in which the LEGO blocks were stacked or the way they were fastened can be different. If the word-based detectors we used previously were checking “what shape of blocks were used,” we are now checking the “overall process order (structure) of building the castle walls and raising the towers.”
Alternatively, you could compare text to a photo filter. While a word-level detector tried to detect the filter that adjusted the colors of the photo, a structural detector identifies the ‘essential composition’ of the photo, such as the placement of the subject in the photo, the angle of the lighting, and the camera’s framing.
The recently studied ‘SlopShape’ technique analyzes this essential ‘skeleton of text.’ Source: SlopShape Research Paper It learns the order in which information is presented, the logical steps taken to reach a conclusion, and how evidence is arranged.
Current situation: The emergence of detectors that see ‘structure’
In fact, according to studies published in 2026, this structural analysis method shows very powerful performance.
- Changes in Story Analysis: In the ‘StoryScope’ study (Russell et al., 2026), they succeeded in distinguishing between AI-generated stories and human-written stories without looking at the words at all. Source: YCombinator Discussion
- High Accuracy: In the ‘SlopShape’ study, when tested on commercial blog posts, it identified AI content with a phenomenal accuracy of 98.0% based solely on structural features. Source: SlopShape Research Paper Notably, this model showed consistent performance even on data from new companies it had never seen during the training process.
Technically, this suggests that no matter how much an AI chooses fluent words to polish its sentences, it is difficult to completely abandon the ‘writing habits’ or ‘logical structure of information delivery’ that the model itself has learned. Source: SlopShape Research Paper
AI’s Opinion
Sequences of words can be easily modified, but the ‘structural patterns’—the writing disposition AI has learned—pierce deeper into the essence. This research signals a new phase in AI content identification.
What will happen in the future?
In the future, we will live in a world where it will be difficult to bypass detectors with just ‘wordplay.’ AI developers will strive to create more natural structures, and detection technology will evolve to read deeper dimensions of logical flow. As readers, we will enter an era where we must look more carefully at the ‘intent’ and ‘logical skeleton’ contained within text, beyond the format of the writing, in a digital environment where AI- and human-created content are mixed.
References
- Slow processing speed
- Inability to grasp the context of content
- Detection performance drops drastically even with slight changes to sentences
- Word choice in sentences
- The structure of the text and the way information is arranged
- The number of images used
- 85.0%
- 92.5%
- 98.0%