@article{4826,
  author = {Pit Pichappan},
  title = {Structural and Linguistic Markers for Distinguishing Human Manuscripts from AI-Generated Summaries},
  journal = {Journal of Digital Information Management},
  year = {2026},
  volume = {24},
  number = {3},
  doi = {https://doi.org/10.6025/jdim/2026/24/3/127-148},
  url = {https://www.dline.info/fpaper/jdim/v24i3/jdimv24i3_1.pdf},
  abstract = {The rapid development of Large Language Models (LLMs) has made it increasingly hard to distinguish
human written from AI-generated academic papers, raising questions about authorship, originality, and
academic integrity. This research presents a systematic approach to identifying and measuring
multidimensional markers that differentiate human written academic manuscripts from AI-generated and
AI-paraphrased summaries. After analysing 52 human written academic papers and their AI-generated
versions (created with GPT-4o, DeepSeek V4, and Qwen 3.7plus), we developed an automated, multi-layered
text-annotation tool. It incorporates rule based matching, zero shot classification, and structured LLMaided
annotation, and we validated it using human inter annotator agreement. We found that no single
language marker is enough for classification. However, the accurate classification (85-90%) requires a set
of markers which include the completeness of the scholarly apparatus (the most reliable one), significant
information loss (40-50% of granular details such as exact statistics or software name), decrease of syntactic
burstiness (by 40-50%), formulaic transition substitutions, and the drop in epistemic hedging by 60%.
Although the absence of structure and metadata yields perfect classification, the stylometric approach
alone produced an AUC of 0.78, underscoring the complexity of linguistic signatures. In summary, successfully
implementing AI for identifying text in academic environments requires a multi marker approach that favours
completeness, informativeness, and text depth over fluency and vocabulary sophistication, which are highly
prone to false positive results. Previous research has addressed AI academic detection; however, we shift
our emphasis from surface fluency (which LLMs excel at mimicking) to epistemic rigour and granular
information loss.},
}