Elamur Research⌕
☰
home / papers / 2026 / Integrity Shield A System for Ethical AI Use & Authorship Transparency in Assessments

Integrity Shield A System for Ethical AI Use & Authorship Transparency in Assessments

Paper · 2026 · Ashish Raj Shekhar, Shiven Agarwal, Priyanuj Bordoloi et al.

Summary

Multimodal LLMs can now take an uploaded exam PDF and solve it end to end, and current countermeasures fall short. Output-based AI-text detectors are unreliable on short or mixed-authorship answers, proctoring is invasive, and decoder-level watermarking requires access to the model, which is impossible when students use proprietary chatbots. Integrity Shield modifies the assessment document instead. It exploits the render-parse gap, the difference between what a PDF shows people and what parsers extract. Hidden text, CMap glyph remapping and off-canvas overlays steer models toward planned distractors or signature phrases, while the visible page stays identical.

The pipeline has three parts:

  • An extraction step builds an item schema.
  • An LLM planner chooses tactics for each question type (MCQ, true/false, long-form).
  • A calibration and authorship service compares model behaviour on original and shielded files, then scores later responses against the stored signatures.

On 30 exams covering STEM, humanities and medicine, the system blocked 91–94% of exams across four commercial MLLMs. It recovered watermark signatures in 89–93% of cases. The authors present it as a privacy-friendly triage aid for instructors, not as proof of misconduct.

Key points

  • The system has two watermark presets: IS-v1 uses light hidden text with minimal glyph remapping, and IS-v2 uses denser multi-layer perturbations. Visual equivalence was checked in Adobe Reader, Chrome and macOS Preview.
  • The evaluated models were GPT-5, Claude Sonnet-4.5, Grok-4.1 and Gemini-2.5 Flash. All were queried as black boxes via their APIs at temperature 0, using fixed prompts for MCQ, T/F and long-form items.
  • The baselines were ICW (in-context watermarking), code-glyph and TrapDoc.
  • The demo has three stages: upload and planning, embedding and calibration, and an authorship dashboard. For long-form answers, a judge LLM scores how closely a response follows the watermark's key phrases.
  • The stated limitations include that robustness was never evaluated for rasterized copies, files re-encoded with Ghostscript, documents round-tripped through Word, or PDF sanitizer tools. Screenshots and retyping are out of scope, and the watermarks will need recalibration as model PDF parsers change.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓Arizona State UniversityTempe · United States50 works · 2,032 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksSocial Sciences · 1,772 worksInformation Systems · 1,437 worksSafety Research · 543 works

🧭 Topics 1Research subjects and languages the work is about.Bias, fairness & safety · 517 works

Cite

Cite as Ashish Raj Shekhar, Shiven Agarwal, Priyanuj Bordoloi, Yash Shah, Tejas Anvekar and Vivek Gupta. 2026. Integrity Shield A System for Ethical AI Use & Authorship Transparency in Assessments. Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 3: System Demonstrations).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.eacl-demo.29

@inproceedings{shekhar2026integrity,
  title = {{Integrity Shield A System for Ethical AI Use \& Authorship Transparency in Assessments}},
  author = {Ashish Raj Shekhar, Shiven Agarwal, Priyanuj Bordoloi, Yash Shah, Tejas Anvekar and Vivek Gupta},
  booktitle = {Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 3: System Demonstrations)},
  year = {2026},
  doi = {10.18653/v1/2026.eacl-demo.29},
  url = {https://doi.org/10.18653/v1/2026.eacl-demo.29}
}

References

14 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Yousef Atoum, Liping Chen, Alex X. Liu, Stephen D. H. Hsu, and Xiaoming Liu. 2017. Automated Online Exam Proctoring. IEEE Transactions on Multimedia, 19(7):1609–1624. find
  2. Bradley Emi and Max Spero. 2024. Technical Report on the Pangram AI-Generated Text Classifier. Technical report, Pangram Labs. find
  3. Hyundong Jin, Sicheol Sung, Shinwoo Park, SeungYeop Baik, and Yo-Sub Han. 2025. TRAPDOC: Deceiving LLM Users by Injecting Imperceptible Phantom Tokens into Documents. EMNLP Findings. ArXiv:2506.00089. find
  4. John Kirchenbauer, Jonas Geiping, Yuxin Wen, and 1 others. 2023. A Watermark for Large Language Models. arXiv preprint arXiv:2301.10226.
  5. Debnath Kundu, Atharva Mehta, Rajesh Kumar, Naman Lal, Avinash Anand, Apoorv Singh, and Rajiv Ratn Shah. 2024. Keystroke Dynamics Against Academic Dishonesty in the Age of LLMs. In Proceedings of the IEEE International Joint Conference on Biometrics (IJCB). find
  6. Yepeng Liu and Yuheng Bu. 2024. Adaptive Text Watermark for Large Language Models. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pages 30718–30737. PMLR. find
  7. Yepeng Liu, Xuandong Zhao, Christopher Kruegel, Dawn Song, and Yuheng Bu. 2025. In-Context Watermarks for Large Language Models. arXiv preprint arXiv:2505.16934. find
  8. Eric Mitchell, Yoonho Lee, Alexander Khazatsky, Christopher D. Manning, and Chelsea Finn. 2023. DetectGPT: Zero-shot Machine-Generated Text Detection Using Probability Curvature. Proceedings of the 40th International Conference on Machine Learning.
  9. Chenhao Niu, Kevin P. Yancey, Ruidong Liu, Mirza Basim Baig, André Kenji Horie, and James Sharpnack. 2024. Detecting LLM-Assisted Cheating on Open-Ended Writing Tasks on Language Proficiency Tests. EMNLP Industry Track. find
  10. OpenAI. 2023. ChatGPT. OpenAI blog. find
  11. Teo Susnjak. 2022. ChatGPT: The End of Online Exam Integrity? Preprint, arXiv:2212.09292. find
  12. Gemini Team. 2024. Gemini 1.5: Unlocking Multimodal Understanding Across Millions of Tokens of Context. arXiv preprint arXiv:2403.05530.
  13. Katherine Thai, Bradley Emi, Elyas Masrour, and Mohit Iyyer. 2025. EditLens: Quantifying the Extent of AI Editing in Text. Preprint, arXiv:2510.03154. — [arXiv:2510.03154](https://arxiv.org/abs/2510.03154) find
  14. Junjie Xiong, Changjia Zhu, Shuhang Lin, Chong Zhang, Yongfeng Zhang, Yao Liu, and Lingyao Li. 2025. Invisible Prompts, Visible Threats: Malicious Font Injection in External Resources for Large Language Models. Preprint, arXiv:2505.16957. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/shekhar-2026-integrity-shield-system-ethical-ai/