On Seeding Watermarks to Detect Verbatim LLM Copy-Paste Responses

arXiv:2605.16336 · cs.CR, cs.AI, cs.CL, cs.CY, cs.ET · Submitted 2026-05-07 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.

Nadia: We're looking at the paper: "On Seeding Watermarks to Detect Verbatim LLM Copy-Paste Responses".

Elias: Here's the gist of the paper. Large language models (LLMs) have made fluent essay writing, code drafting, and quiz answering instantly available to students at every level from secondary school through graduate study <ref:2605.16336#pg2>.

Nadia: Many educators do not object to LLM use per se; what they need to detect is the case in which a student pastes the assignment prompt into a chatbot and submits the model’s reply verbatim without engaging with the work <ref:2605.16336#pg3>.

Elias: The paper proposes SteganoPrompt, an inputside watermark embedded inside the visible assignment prompt itself, which an LLM reads to write a tell-tale signature into its reply specifically exposing the copy-and-paste pathway <ref:2605.16336#pg2>.

Nadia: The Problem with Existing Detection Methods Existing posthoc AI-text detectors remain unreliable and have been shown to penalise non-native English writers, while output-side watermarks require cooperation from the model provider <ref:2605.16336#pg3>.

Paper summary: Nadia: So we're talking about this paper called "On Seeding Watermarks to Detect Verbatim LLM Copy-Paste Responses." Essentially, they’re looking at how students are using AI to write essays and the problem is that teachers can't easily tell when a student just pastes the whole assignment prompt into a chatbot and turns in the AI's reply without actually doing any of the work >

Elias: Exactly. The paper says existing methods for spotting this, like those looking at text after it’s been written, aren't very reliable because they can end up punishing students who write in a way that isn't standard English or non-native speakers have >

Priya: And output watermarks are another problem because you need the model provider to cooperate, which most teachers don't have control over for their classrooms >

Nadia: So what they propose is something different. They suggest an inputside watermark, meaning they embed an invisible instruction directly into the assignment prompt itself, and if an AI reads that prompt verbatim it writes a secret signature into its response to expose the copy-paste pathway >

Elias: That specific technique is called SteganoPrompt, and it uses a method where they encode something printable in ASCII into those Unicode tags block U+E0000 to U+E007F which looks exactly like the original text >

Nadia: It sounds like they're trying to create a channel that the student doesn't even see, so if an AI ingests the prompt verbatim, it quietly reads that hidden instruction and writes a tell-tale signature into its reply >

Elias: The method involves setting A equal to Ap union Aw and defining T as c plus zero E000 colon where C is in A >

Priya: I'm wondering about the practical side of this. If the goal is to detect verbatim copying, how does this inputside approach actually survive the process of a student using it? >

Nadia: Well, they focused on making sure it has visual identity and that it survives common distribution pipelines, which means testing if it gets stripped out by different text-editing tools >

Elias: They found that models like Claude and Gemini emitted this integrity token in every trial, but GPT series showed none because of something called input scrubbing >

Priya: That's interesting about the model compliance. But what about the channel survival part? Does it actually stay intact when people copy that text to different places? >

Nadia: They found that the payload is preserved through every text-editing pipeline they tested, with only some social platforms stripping extended Unicode characters >

Elias: They also mentioned linguistic steganography as another area, like using context-aware lexical substitution to trace where a passage came from, but they noted those methods change the wording which isn't what you want for an assignment brief that needs to stay verbatim >

Priya: So the Unicode tag approach is different because it modifies the byte stream while leaving every rendered glyph unchanged, which keeps the text looking normal >

Nadia: Right. And they built this tool as a single page web tool that does all the computation right in your browser, which makes it pretty accessible for someone trying to test this out >

Elias: The recommended way to use it involves four steps: first disclosing the policy in the syllabus, second encoding each brief with SteganoPrompt and pasting it into the learning management system, third verifying before releasing by decoding the brief to confirm the payload is bitidentical, and finally searching submissions for that literal token at grading time >

Priya: I wonder about that verification step. If you decode it to check for a bitidentical payload, does that really give you confidence in catching someone who might be slightly altering the prompt? >

Nadia: The paper argues that this workflow generalizes the manual word of the day check without needing visible student action, which is a big deal for teachers trying to keep up with this AI use >

Elias: Beyond just detecting verbatim copying, they framed it as an alternative that you control directly, sitting between output watermarking and indirect prompt injection attacks >

Priya: What are the ethical considerations here? The paper touches on the need to disclose the policy to have a deterrent effect and using a kind payload that doesn't embarrass students if they make a mistake >

Nadia: They suggest that if someone hits this watermark, it should be treated as a starting point for conversation instead of an immediate verdict on their integrity >

Elias: And they are careful about using the technique against systems you don't own or haven't been authorized to test because it can be used in adversarial ways, like indirect prompt injection >

Priya: So, what’s the actual future direction for this work? What do they suggest doing next after proving this inputside watermark works for detecting copy-paste? >

Nadia: They suggested embedding an HMAC over assignment metadata to get stronger verification of the assignment itself >

Elias: They also proposed combining that Unicode-Tag channel with a content level watermark to help it survive light paraphrasing, and they plan a multi section, multi instructor field study to estimate the actual detection rates across different scenarios >

Priya: That sounds like a solid plan for getting more robust data on how effective this is in the real world, especially when dealing with different types of student work >

Conclusion: Nadia: So we're wrapping up this talk on "On Seeding Watermarks to Detect Verbatim LLM Copy-Paste Responses." Basically, the paper shows an inputside way to watermark assignment prompts so if an AI copies it verbatim, it leaves a traceable signature in its answer.

Elias: It’s about embedding something into the prompt itself that the AI reads and then writes back. The authors used something called SteganoPrompt which uses those unused Unicode tags block U plus E zero zero zero to U plus E zero seven f which looks exactly like normal text.

Priya: So, what does this mean for us? We're talking about making sure students aren't just pasting the whole question into a chatbot and submitting the AI's answer without doing the actual thinking.

Nadia: Right, it’s a direct way to check that specific copying behavior without having to look at the final essay text. The authors tested this across different language models like Claude and Gemini, and they saw that most of them put out this integrity token in every test run.

Elias: But the results weren't perfect. They found that some models, specifically a GPT series model, didn't show any token because it had some input scrubbing mechanism. That’s a big caveat for us to keep in mind when we talk about model compliance across the board.

Priya: So, even with this watermark on the prompt side, you still have to be careful about which AI you are testing against and how that specific model handles instructions before you can trust the detection. The data shows that channel survival is good; they found it stays intact through most text editing pipelines.

Nadia: It’s a tool sitting right between trying to detect what the AI writes and trying to control what the AI produces at output time. It gives educators a way to catch that direct copy-paste action during grading, which is currently pretty hard because of all the paraphrasing students do.

Elias: Exactly. The authors suggest this inputside watermark is a working classroom integrity tool, but they also point out it’s not perfect on its own. They plan more work to combine this Unicode channel with a content level watermark to handle light paraphrasing better.

Priya: And they’re looking at embedding something called an HMAC over assignment metadata for even stronger verification of the assignment itself before you even get to the student submissions. That suggests a layered approach might be needed here.

Nadia: So, this paper gives us a concrete method—SteganoPrompt—that allows instructors to search for that literal token at grading time, which is a big step toward making academic integrity checks more practical and less reliant on just looking at the final paper.

George Washington University

cs.CR, cs.AI, cs.CL, cs.CY, cs.ET

Submitted: 2026-05-07

Updated: 2026-08-07

Comments: We are pleased to announce that this paper has been accepted by the 9th AAAI/ACM Conference on AI, Ethics, and Society (AIES 2026). We appreciate the valuable feedback from the reviewers and look forward to sharing our findings with the community

Journal ref: Proceedings of the Ninth AAAI/ACM Conference on AI, Ethics, and Society (AIES 2026)

DOI: 10.1609/aies.v9i1.43126

Code: https://github.com/Ezharjan/SteganoPrompt

Project page: https://ezharjan.github.io/SteganoPrompt

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 82/100

The gist: Large language models (LLMs) have made fluent essay writing, code drafting, and quiz answering instantly available to students at every level from secondary school through graduate study…

Key concepts

SteganoPrompt
This is an inputside watermarking technique where a hidden, printable-ASCII message is encoded directly into the visible text of an assignment prompt using specific Unicode blocks. The goal is to leave the text visually unchanged while embedding a secret signal for detection.
Inputside Watermarking
Instead of marking the output, this method modifies the input (the prompt) before it reaches the LLM. SteganoPrompt achieves this by encoding data into unused Unicode tags, making it an 'invisible' modification that triggers a specific response from the model.
Channel Survival
This refers to how well the embedded watermark survives common text processing steps, like copying and pasting or passing through different text-editing pipelines. The technique is designed to ensure the hidden payload remains intact even when the prompt is manipulated by students.
Cross-Model Compliance
This measures whether different large language models (like Claude or Gemini) correctly recognize and respond to the hidden instruction embedded in the SteganoPrompt watermark. The paper found that some models respected these hidden instructions while others did not.

Terminology

Summary

Large language models (LLMs) have made fluent essay writing, code drafting, and quiz answering instantly available to students at every level from secondary school through graduate study <ref:2605. p. 0

Where the paper says it

On Seeding Watermarks to Detect Verbatim LLM Copy-Paste Responses Aizierjiang Aiersilan Q Artin Yousefi Robert Pless The George Washington University {alexandera, artinyousefi, pless}@gwu. edu Abstract Large language models (LLMs) have made fluent essay writing, code drafting, and quiz answering instantly available to students at every level, from secondary school through graduate study.

16336#pg2>. Many educators do not object to LLM use per se; what they need to detect is the case in which a student pastes the assignment prompt into a chatbot and submits the model’s reply verbatim without engaging with the work <ref:2605. p. 0
Where the paper says it

edu Abstract Large language models (LLMs) have made fluent essay writing, code drafting, and quiz answering instantly available to students at every level, from secondary school through graduate study. Many educators do not object to LLM use per se; what they need to detect is the case in which a student pastes the assignment prompt into a chatbot and submits the model’s reply verbatim, without engaging with the work.

16336#pg3>. The paper proposes SteganoPrompt, an inputside watermark embedded inside the visible assignment prompt itself, which an LLM reads to write a tell-tale signature into its reply specifically exposing the copy-and-paste pathway <ref:2605. p. 0
Where the paper says it

We propose an alternative that the educator controls directly: an inputside watermark in which an invisible instruction is embedded inside the visible assignment prompt itself. An LLM that ingests the prompt verbatim quietly reads the hidden instruction and writes a tell-tale signature into its reply, exposing the copyand-paste pathway specifically.

16336#pg2>.

The Problem with Existing Detection Methods

Existing posthoc AI-text detectors remain unreliable and have been shown to penalise non-native English writers, while output-side watermarks require cooperation from the model provider <ref:2605.16336#pg3>. The paper identifies three options for countermeasures: post-hoc detection, output-side watermarking, and inputside watermarking <ref:2605. p. 1

Where the paper says it

Taken together, post-hoc detection is informative as a signal but not as a verdict. Output-Side Watermarking of LLM Text A second, complementary line embeds a watermark at the model end.

16336#pg3>. Post-hoc detectors are unreliable because they can be defeated by light paraphrasing and disproportionately misclassify essays written by non-native English speakers <ref:2605. p. 0
Where the paper says it

2023; Krishna et al. 2023) results have shown that such detectors are unreliable, can be defeated by light paraphrasing, and disproportionately misclassify essays written by non-native English speakers.

16336#pg5>. Output watermarking is only available to parties who control the model, which classroom teachers generally do not possess <ref:2605. p. 0
Where the paper says it

2023; Christ, Gunn, and Zamir 2024; Zhao et al. Output watermarking is technically attractive but is only available to parties who control the model, while classroom teacher or university instructor generally does not.

16336#pg3>.

How the Training Signal is Built

SteganoPrompt encodes an arbitrary printable-ASCII payload into the deprecated Unicode Tags block U+E0000–U+E007F, which is visually identical to the original text <ref:2605.16336#pg2>. The encoding procedure involves setting A = Ap ∪ Aw and defining T = c + 0xE000: c ∈ A (1) <ref:2605. p. 2

Where the paper says it

The instructor must be able to decode any text and inspect the payload before distribution, so that the hidden message contains exactly what was intended. Encoding Procedure Let Ap = { c ∈ N : 0x20 ≤ c ≤ 0x7E } denote the 95 printable ASCII code points, and let Aw = {0x09, 0x0A} denote the horizontal tab and the line feed, which we carry through encoding so that the layout of multi-line payloads is preserved.

16336#pg5>. The watermarked document is the concatenation wmπ(P, W) = P[1:k] ∥ E(W) ∥ P[k+1:P] (4) <ref:2605. p. 2
Where the paper says it

\n is encoded as E(W) = ⟨ U+E0048, U+E0069, U+E0021, U+E000A ⟩, none of which renders as a glyph in any common font, but each of which is tokenized by current chatbots as the corresponding ASCII byte. Given a visible text P ∈ A∗ and a payload W ∈ A∗ , the watermarked document is the concatenation wmπ(P, W) = P[1:k] ∥ E(W) ∥ P[k+1:|P |] , (4) where ∥ denotes string concatenation, P[i:j] is the substring of P from position i to position j (and is empty if j < i),

16336#pg5>. The insertion index k depends on the placement parameter π, which can be start, end, or mid-sentence boundary index µ(P) <ref:2605. p. 3
Where the paper says it

, |P|} depends on the placement parameter π. Define the first-sentence-boundary index µ(P) = min j : Pj ∈ {.

16336#pg5>. The tool is a single-page, zero-dependency web tool that performs all computation in the user’s browser <ref:2605. p. 0
Where the paper says it

An LLM that ingests the prompt verbatim quietly reads the hidden instruction and writes a tell-tale signature into its reply, exposing the copyand-paste pathway specifically. We describe SteganoPrompt, a single-page, zero-dependency web tool that encodes an arbitrary printable-ASCII payload into the deprecated Unicode Tags block (U+E0000–U+E007F).

16336#pg3>.

Where the Gains Come From

The technique is designed around five goals, including visual identity, channel survival, reader transparency, and privacy and offline use <ref:2605.16336#pg5>. The evaluation focuses on two technical axes: cross-model compliance with the hidden instruction and cross-channel survival of the payload through common distribution pipelines <ref:2605. p. 4

Where the paper says it

The visible assignment text is never altered, and the Verify panel lets the instructor confirm the exact payload before distribution. Evaluation We evaluate SteganoPrompt on two technical axes that are controllable independently of the student population: (i) cross-model compliance with the hidden instruction, and (ii) cross-channel survival of the payload through common distribution pipelines.

16336#pg5>. Cross-Model Compliance shows that models like Claude, Gemini, and Grok emitted the integrity token in every trial, while GPT-5 series models showed none due to input scrubbing <ref:2605. p. 7
Where the paper says it

The latest GPT-5- series models appear to apply such a filter, and independent audits report the same input scrubbing in some other hosted chat products (Markopoulos and Fagan 2025; Toulas 2025). Other providers have not adopted it: Google has stated that it does not treat the channel as a security issue for Gemini (Toulas 2025), and the Claude, Grok, Qwen, DeepSeek, and Codestral pipelines all delivered the Tag block to the model unfiltered.

16336#pg8>. Channel Survival indicates the payload is preserved through every text-editing pipeline tested, with only some social platforms stripping extended Unicode <ref:2605. p. 5
Where the paper says it

Cross-Channel Survival Table 2 summarizes the channel-survival study. The payload is preserved through every text-editing pipeline we tested, including Microsoft Word (.

16336#pg8>.

Recommended Deployment Workflow

The recommended workflow for an instructor involves four steps: 1) Disclose the policy in the syllabus <ref:2605.16336#pg8>; 2) Encode each brief using SteganoPrompt and paste it into the LMS <ref:2605. p. 6

Where the paper says it

Encode each brief. Paste the visible assignment text into SteganoPrompt, choose (or write) a payload, and paste the encoded result into the learning-management system.

16336#pg8>; 3) Verify before release by decoding the brief to confirm the payload is bitidentical <ref:2605. p. 6
Where the paper says it

Verify before release. Use the Verify panel to decode the brief from the same channel a student would read it through, and confirm that the recovered payload is bitidentical to the intended one.

16336#pg8>; and 4) Search submissions for the literal token at grading time <ref:2605. p. 6
Where the paper says it

Search at grading time. Search submissions for the literal token (by default, SteganoPrompt-OK-2026).

16336#pg8>. This workflow generalizes the manual word-of-the-day check without requiring visible student action <ref:2605. p. 6
Where the paper says it

The deterrent effect of an announced policy is, by all available evidence on classroom integrity, larger than the catch effect on any individual assignment, and disclosure is also the ethically required posture. The workflow generalizes the manual word-of-the-day check that originally motivated this work, without requiring the student to do anything visible: the tripwire travels with the prompt rather than with the student.

16336#pg8>.

Ethics and Disclosure

The responsible deployment of input-side watermarks has four properties, including disclosing the policy to maximize deterrent effect <ref:2605.16336#pg9>. The paper argues for using a kind payload that helps students in good faith, avoiding content that would embarrass them <ref:2605. p. 7

Where the paper says it

Use a kind payload. Prefer payloads that, when followed, help a student who is using AI in good faith.

16336#pg9>. A watermark hit should be treated as a starting point for a conversation rather than a verdict <ref:2605. p. 7
Where the paper says it

Treat hits as a starting point. A watermark hit is a probable-cause signal for a conversation, not a verdict.

16336#pg9>. The technique is also noted to be an indirect prompt injection in adversarial hands, which means it should not be used against systems the instructor does not own or have not been authorized to test <ref:2605. p. 7
Where the paper says it

The same technique is, in adversarial hands, an indirect prompt injection. We do not endorse its use against any system the instructor does not own or has not been authorized to test.

16336#pg9>.

Conclusion and Future Work

The authors present SteganoPrompt as a working classroom integrity tool that sits between output-side watermarking and indirect prompt injection <ref:2605.16336#pg10>. Future work suggests embedding an HMAC over assignment metadata for stronger verification <ref:2605. p. 8

Where the paper says it

Several directions are worth pursuing. First, the current payload is a literal string; a stronger design would embed an Hash-based Message Authentication Code (HMAC) over the assignment metadata, so that the instructor could verify not only that a watermark is present but that it is the watermark associated with this assignment, this term, and this section.

16336#pg10>. They also propose combining the Unicode-Tag channel with a content-level watermark to survive light paraphrasing and conducting a multi-section, multi-instructor field study to estimate detection rates <ref:2605. p. 8
Where the paper says it

2022; Ueoka, Murawaki, and Kurohashi 2021), would survive a student who lightly paraphrases the brief before pasting it. Third, a pre-registered, multi-section, multi-instructor field study would allow detection and deterrence rates to be estimated with statistical confidence; we did not run such a study and we leave it to follow-on work.

16336#pg10>. The tool is released under the MIT licence at https://ezharjan. p. 8
Where the paper says it

Fourth, as model pipelines incorporate input sanitizers, the design space for integrity watermarks will narrow; we are interested in a principled negotiation between the educational community and model providers on which input transformations should and should not be applied silently to user-pasted text. The tool is released under the MIT licence at https: //ezharjan.

github.io/SteganoPrompt/ <ref:2605. p. 0
Where the paper says it

This work was motivated by two recent observations. First, the program chairs of ICML 2026 announced that reviewerfacing materials would be circulated with hidden watermarks, so that a reviewer who pasted a call into an LLM and copied the model’s reply back into the review form could be flagged arXiv:2605.

16336#pg10>.

References

Aaronson, S.; and Kirchner, H. p. 0

Where the paper says it

The second family is output-side watermarking, in which the model provider biases its sampling distribution at generation time so that a downstream verifier can later test for the watermark (Kirchenbauer et al. 2023; Aaronson and Kirchner 2023; Kuditipudi et al.

2023. Watermarking of large language models. p. 8
Where the paper says it

; and Kirchner, H. Watermarking of large language models.

Bommasani, R. p. 0

Where the paper says it

com/Ezharjan/SteganoPrompt Introduction The release of broadly capable instruction-following language models (Ouyang et al. 2022; Bommasani et al.

; Hudson, D. p. 8
Where the paper says it

Bommasani, R. ; Hudson, D.

A.; Adeli, E. p. 8
Where the paper says it

; Hudson, D. ; Adeli, E.

; Altman, R. p. 8
Where the paper says it

; Adeli, E. ; Altman, R.

; Arora, S. p. 8
Where the paper says it

; Altman, R. ; Arora, S.

; von Arx, S. p. 8
Where the paper says it

; Arora, S. ; von Arx, S.

; Bernstein, M. p. 8
Where the paper says it

; von Arx, S. ; Bernstein, M.

S.; Bohg, J.; Bosselut, A. p. 8
Where the paper says it

; Bernstein, M. ; Bosselut, A.

; Brunskill, E. p. 8
Where the paper says it

; Bosselut, A. ; Brunskill, E.

, et al. 2021. On the opportunities and risks of foundation models. p. 8
Where the paper says it

; Brunskill, E. On the opportunities and risks of foundation models.

Boucher, N. p. 1

Where the paper says it

SteganoPrompt sits within this conceptual frame, but with an inverted threat model: the party embedding the hidden instruction is not an attacker, but the legitimate author of the document, and the victim of the injection is a copy-and-paste workflow that the author wishes to discourage. Unicode and Character-Level Channels Boucher et al.

; Shumailov, I. p. 8
Where the paper says it

Boucher, N. ; Shumailov, I.

; Anderson, R. p. 8
Where the paper says it

; Shumailov, I. ; Anderson, R.

; and Papernot, N. p. 8
Where the paper says it

; Anderson, R. ; and Papernot, N.

2022. Bad characters: Imperceptible nlp attacks. p. 8
Where the paper says it

; and Papernot, N. Bad characters: Imperceptible nlp attacks.

Christ, M.; Gunn, S.; and Zamir, O. p. 0

Where the paper says it

2023; Aaronson and Kirchner 2023; Kuditipudi et al. 2023; Christ, Gunn, and Zamir 2024; Zhao et al.

2024. Undetectable watermarks for language models. p. 8
Where the paper says it

; and Zamir, O. Undetectable watermarks for language models.

Cotton, D. R.; Cotton, P. p. 0

Where the paper says it

2021) has made it trivial for a student, whether in secondary school or at university, to obtain a fluent answer to almost any shortanswer or short-essay assignment by pasting the prompt into a chatbot. Surveys in education and rapid reviews of the resulting literature (Cotton, Cotton, and Shipway 2024; Susnjak and McIntosh 2024; Lo 2023; Perkins 2023; Stokel-Walker Copyright © 2026, Association for the Advancement of Artificial Intelligence (www.

A.; and Shipway, J. p. 0
Where the paper says it

2021) has made it trivial for a student, whether in secondary school or at university, to obtain a fluent answer to almost any shortanswer or short-essay assignment by pasting the prompt into a chatbot. Surveys in education and rapid reviews of the resulting literature (Cotton, Cotton, and Shipway 2024; Susnjak and McIntosh 2024; Lo 2023; Perkins 2023; Stokel-Walker Copyright © 2026, Association for the Advancement of Artificial Intelligence (www.

R. 2024. Chatting and cheating: Ensuring academic integrity in the era of ChatGPT. p. 8
Where the paper says it

; and Shipway, J. Chatting and cheating: Ensuring academic integrity in the era of ChatGPT.

Eger, S.; S¸ahin, G. p. 8

Where the paper says it

Innovations in education and teaching international, 61(2): 228–239. ; S¸ahin, G.

G.; Ruckl ¨ e, A. p. 8
Where the paper says it

; S¸ahin, G. ; Ruckl ¨ e, A.

; Lee, J.-U.; Schulz, C. p. 8
Where the paper says it

; Ruckl ¨ e, A. ; Schulz, C.

; ´Mesgar, M. p. 8
Where the paper says it

; Schulz, C. ; ´ Mesgar, M.

; Swarnkar, K. p. 8
Where the paper says it

; ´ Mesgar, M. ; Swarnkar, K.

; Simpson, E. p. 8
Where the paper says it

; Swarnkar, K. ; Simpson, E.

, and Gurevych, I. p. 8
Where the paper says it

; Simpson, E. ; and Gurevych, I.

2019. Text processing like humans do: Visually attacking and shielding NLP systems. p. 8
Where the paper says it

; and Gurevych, I. Text processing like humans do: Visually attacking and shielding NLP systems.

Goodside, R. p. 5

Where the paper says it

Compliance is the token-emission rate: high (every trial), medium (50–99%), low (1–49%), none (never). Ratings combine our experiments with public vendor and security disclosures (Goodside 2024; Markopoulos and Fagan 2025; Toulas 2025; Rehberger 2024b,c; Liu 2025; Yeung 2026).

2024. Invisible Prompt Injection via Unicode Tag Characters. p. 8
Where the paper says it

Goodside, R. Invisible Prompt Injection via Unicode Tag Characters.

Greshake, K. p. 1

Where the paper says it

Indirect Prompt Injection A separate strand in the security literature studies what happens when an LLM consumes content that the user did not author. Greshake et al.

; Abdelnabi, S. p. 8
Where the paper says it

Greshake, K. ; Abdelnabi, S.

; Mishra, S. p. 8
Where the paper says it

; Abdelnabi, S. ; Mishra, S.

; Endres, C. p. 8
Where the paper says it

; Mishra, S. ; Endres, C.

; Holz, T.; and Fritz, M. p. 8
Where the paper says it

; Endres, C. ; and Fritz, M.

2023. Not what you’ve signed up for: Compromising real-world llm-integrated applications with indirect prompt injection. p. 8
Where the paper says it

; and Fritz, M. Not what you’ve signed up for: Compromising real-world llm-integrated applications with indirect prompt injection.

Hines, K.; Lopez, G. p. 8

Where the paper says it

In Proceedings of the 16th ACM workshop on artificial intelligence and security, 79–90. ; Lopez, G.

; Hall, M.; Zarfati, F. p. 8
Where the paper says it

; Lopez, G. ; Zarfati, F.

; Zunger, Y. p. 8
Where the paper says it

; Zarfati, F. ; Zunger, Y.

; and Kiciman, E. p. 8
Where the paper says it

; Zunger, Y. ; and Kiciman, E.

2024. Defending against indirect prompt injection attacks with spotlighting. p. 8
Where the paper says it

; and Kiciman, E. Defending against indirect prompt injection attacks with spotlighting.

ICML 2026. 2026. On Violations of LLM Review Policies. p. 8

Where the paper says it

arXiv preprint arXiv:2403. On Violations of LLM Review Policies.

Kirchenbauer, J. p. 0

Where the paper says it

This problem is likely to deepen over time, as human learners of English are themselves exposed to and may internalize the stylistic conventions of AI-generated text. The second family is output-side watermarking, in which the model provider biases its sampling distribution at generation time so that a downstream verifier can later test for the watermark (Kirchenbauer et al.

; Geiping, J. p. 8
Where the paper says it

Kirchenbauer, J. ; Geiping, J.

; Wen, Y.; Katz, J.; Miers, I. p. 8
Where the paper says it

; Geiping, J. ; Miers, I.

; and Goldstein, T. p. 8
Where the paper says it

; Miers, I. ; and Goldstein, T.

2023. A watermark for large language models. p. 8
Where the paper says it

; and Goldstein, T. A watermark for large language models.

Krishna, K. p. 0

Where the paper says it

2023) and theoretical (Sadasivan et al. 2023; Krishna et al.

; Song, Y.; Karpinska, M. p. 8
Where the paper says it

Krishna, K. ; Karpinska, M.

; Wieting, J. p. 8
Where the paper says it

; Karpinska, M. ; Wieting, J.

; and Iyyer, M. p. 8
Where the paper says it

; Wieting, J. ; and Iyyer, M.

2023. Paraphrasing evades detectors of ai-generated text, but retrieval is an effective defense. p. 8
Where the paper says it

; and Iyyer, M. Paraphrasing evades detectors of ai-generated text, but retrieval is an effective defense.

Kuditipudi, R. p. 0

Where the paper says it

The second family is output-side watermarking, in which the model provider biases its sampling distribution at generation time so that a downstream verifier can later test for the watermark (Kirchenbauer et al. 2023; Aaronson and Kirchner 2023; Kuditipudi et al.

; Thickstun, J. p. 8
Where the paper says it

Kuditipudi, R. ; Thickstun, J.

; Hashimoto, T. p. 8
Where the paper says it

; Thickstun, J. ; Hashimoto, T.

; and Liang, P. p. 0
Where the paper says it

2023), the (since-withdrawn) OpenAI classifier (OpenAI 2023), and a number of commercial services. Both empirical (Liang et al.

2023. Robust distortion-free watermarks for language models. p. 8
Where the paper says it

; and Liang, P. Robust distortion-free watermarks for language models.

Liang, W.; Izzo, Z.; Zhang, Y. p. 8

Where the paper says it

arXiv preprint arXiv:2307. ; Zhang, Y.

; Lepp, H.; Cao, H.; Zhao, X.; Chen, L.; Ye, H.; Liu, S., Huang., Z., et al. 2024. Monitoring ai-modified content at scale: A case study on the impact. p. 8
Where the paper says it

; Huang, Z. Monitoring ai-modified content at scale: A case study on the impact

Liang, W.; Yuksekgonul, M. p. 9

Where the paper says it

arXiv preprint arXiv:2403. ; Yuksekgonul, M.

; Mao, Y.; Wu, E.; and Zou, J. p. 9
Where the paper says it

; Yuksekgonul, M. ; and Zou, J.

2023. GPT detectors are biased against non-native English writers. p. 9
Where the paper says it

; and Zou, J. GPT detectors are biased against non-native English writers.

Liu, I. C. 2025. Invisible Prompt Injection: A Threat to AI Security. p. 9

Where the paper says it

Patterns, 4(7). Invisible Prompt Injection: A Threat to AI Security.

Liu, Y.; Jia, Y.; Geng, R.; Jia, J.; and Gong, N. p. 9

Where the paper says it

com/en us/research/25/a/ invisible-prompt-injection-secure-ai. ; and Gong, N.

Z. 2024. Formalizing and benchmarking prompt injection attacks and defenses. p. 9
Where the paper says it

; and Gong, N. Formalizing and benchmarking prompt injection attacks and defenses.

Lo, C. K. 2023. What is the impact of ChatGPT on education? p. 9

Where the paper says it

In 33rd USENIX Security Symposium (USENIX Security 24), 1831–1847. What is the impact of ChatGPT on education?

A rapid review of the literature. p. 1
Where the paper says it

(2024) survey responses by universities and recommend a combination of policy, assessment design, and detection. Perkins (2023) catalogues post-pandemic integrity considerations specific to LLMs, while Lo (2023) provides a rapid review of the early educational literature.

Markopoulos, V. p. 5

Where the paper says it

Compliance is the token-emission rate: high (every trial), medium (50–99%), low (1–49%), none (never). Ratings combine our experiments with public vendor and security disclosures (Goodside 2024; Markopoulos and Fagan 2025; Toulas 2025; Rehberger 2024b,c; Liu 2025; Yeung 2026).

; and Fagan, A. p. 5
Where the paper says it

Compliance is the token-emission rate: high (every trial), medium (50–99%), low (1–49%), none (never). Ratings combine our experiments with public vendor and security disclosures (Goodside 2024; Markopoulos and Fagan 2025; Toulas 2025; Rehberger 2024b,c; Liu 2025; Yeung 2026).

2025. Ghosts in the Machine: ASCII Smuggling Across Various LLMs. p. 9
Where the paper says it

; and Fagan, A. Ghosts in the Machine: ASCII Smuggling Across Various LLMs.

Mitchell, E. p. 0

Where the paper says it

The first is post-hoc detection, in which a classifier inspects the submitted text and decides whether it was produced by a model. Tools in this family include DetectGPT (Mitchell et al.

; Lee, Y.; Khazatsky, A. p. 9
Where the paper says it

Mitchell, E. ; Khazatsky, A.

; Manning, C. p. 9
Where the paper says it

; Khazatsky, A. ; Manning, C.

D.; and Finn, C. p. 9
Where the paper says it

; Manning, C. ; and Finn, C.

2023. Detectgpt: Zero-shot machine-generated text detection using probability curvature. p. 9
Where the paper says it

; and Finn, C. Detectgpt: Zero-shot machine-generated text detection using probability curvature.

OpenAI. 2023. New AI Classifier for Indicating AIWritten Text. p. 9

Where the paper says it

In International conference on machine learning, 24950–24962. New AI Classifier for Indicating AIWritten Text.

Ouyang, L.; Wu, J.; Jiang, X. p. 9

Where the paper says it

com/blog/new-ai-classifier-forindicating-ai-written-text. ; Jiang, X.

; Almeida, D. p. 9
Where the paper says it

; Jiang, X. ; Almeida, D.

; Wainwright, C. p. 9
Where the paper says it

; Almeida, D. ; Wainwright, C.

; Mishkin, P. p. 9
Where the paper says it

; Wainwright, C. ; Mishkin, P.

; Zhang, C. p. 8
Where the paper says it

arXiv preprint arXiv:2307. ; Zhang, Y.

, Agarwal, S. p. 9
Where the paper says it

; Zhang, C. ; Agarwal, S.

, Slama, K. p. 9
Where the paper says it

; Agarwal, S. ; Slama, K.

, Ray, A., et al. 2022. Training language models to follow instructions with human feedback. p. 9
Where the paper says it

; Slama, K. Training language models to follow instructions with human feedback.

Perez, F.; and Ribeiro, I. p. 1

Where the paper says it

(2023) introduce indirect prompt injection, in which an attacker plants instructions inside data later retrieved by an LLM-integrated application. Perez and Ribeiro (2022) demonstrate the underlying ignore previous instructions attack, and Liu et al.

2022. Ignore previous prompt: Attack techniques for language models. p. 9
Where the paper says it

; and Ribeiro, I. Ignore previous prompt: Attack techniques for language models.

Perkins, M. p. 0

Where the paper says it

2021) has made it trivial for a student, whether in secondary school or at university, to obtain a fluent answer to almost any shortanswer or short-essay assignment by pasting the prompt into a chatbot. Surveys in education and rapid reviews of the resulting literature (Cotton, Cotton, and Shipway 2024; Susnjak and McIntosh 2024; Lo 2023; Perkins 2023; Stokel-Walker Copyright © 2026, Association for the Advancement of Artificial Intelligence (www.

2023. Academic integrity considerations of AI large language models in the post-pandemic era: ChatGPT and beyond. p. 9
Where the paper says it

Perkins, M. Academic integrity considerations of AI large language models in the post-pandemic era: ChatGPT and beyond.

Rehberger, J. p. 0

Where the paper says it

Output watermarking is technically attractive but is only available to parties who control the model, while classroom teacher or university instructor generally does not. A third option, less explored in the academic literature but broadly discussed by the prompt-security community (Rehberger 2024a), is to watermark the input that the model is asked to consume.

p. 0
Where the paper says it

Output watermarking is technically attractive but is only available to parties who control the model, while classroom teacher or university instructor generally does not. A third option, less explored in the academic literature but broadly discussed by the prompt-security community (Rehberger 2024a), is to watermark the input that the model is asked to consume.

p. 0
Where the paper says it

Output watermarking is technically attractive but is only available to parties who control the model, while classroom teacher or university instructor generally does not. A third option, less explored in the academic literature but broadly discussed by the prompt-security community (Rehberger 2024a), is to watermark the input that the model is asked to consume.

2024a. ASCII Smuggler Tool: Crafting Invisible Text and Decoding Hidden Codes. p. 9
Where the paper says it

Rehberger, J. ASCII Smuggler Tool: Crafting Invisible Text and Decoding Hidden Codes.

Rehberger, J. 2024b. Hidden Prompt Injections with Anthropic Claude. p. 9

Where the paper says it

Rehberger, J. Hidden Prompt Injections with Anthropic Claude.

Rehberger, J. 2024c. Security ProbLLMs in xAI’s Grok: A Deep Dive. p. 9

Where the paper says it

Rehberger, J. Security ProbLLMs in xAI’s Grok: A Deep Dive.

Sadasivan, V. p. 0

Where the paper says it

Both empirical (Liang et al. 2023) and theoretical (Sadasivan et al.

S.; Kumar, A. p. 9
Where the paper says it

Sadasivan, V. ; Kumar, A.

; Balasubramanian, S. p. 9
Where the paper says it

; Kumar, A. ; Balasubramanian, S.

; Wang, W.; and Feizi, S. p. 9
Where the paper says it

; Balasubramanian, S. ; and Feizi, S.

2023. Can AI-generated text be reliably detected? p. 9
Where the paper says it

; and Feizi, S. Can AI-generated text be reliably detected?

Stokel-Walker, C. p. 0

Where the paper says it

2021) has made it trivial for a student, whether in secondary school or at university, to obtain a fluent answer to almost any shortanswer or short-essay assignment by pasting the prompt into a chatbot. Surveys in education and rapid reviews of the resulting literature (Cotton, Cotton, and Shipway 2024; Susnjak and McIntosh 2024; Lo 2023; Perkins 2023; Stokel-Walker Copyright © 2026, Association for the Advancement of Artificial Intelligence (www.

2022. AI bot ChatGPT writes smart essaysshould professors worry? p. 9
Where the paper says it

Stokel-Walker, C. AI bot ChatGPT writes smart essaysshould professors worry?

Nature.

Susnjak, T. p. 0

Where the paper says it

2021) has made it trivial for a student, whether in secondary school or at university, to obtain a fluent answer to almost any shortanswer or short-essay assignment by pasting the prompt into a chatbot. Surveys in education and rapid reviews of the resulting literature (Cotton, Cotton, and Shipway 2024; Susnjak and McIntosh 2024; Lo 2023; Perkins 2023; Stokel-Walker Copyright © 2026, Association for the Advancement of Artificial Intelligence (www.

; and McIntosh, T. p. 0
Where the paper says it

2021) has made it trivial for a student, whether in secondary school or at university, to obtain a fluent answer to almost any shortanswer or short-essay assignment by pasting the prompt into a chatbot. Surveys in education and rapid reviews of the resulting literature (Cotton, Cotton, and Shipway 2024; Susnjak and McIntosh 2024; Lo 2023; Perkins 2023; Stokel-Walker Copyright © 2026, Association for the Advancement of Artificial Intelligence (www.

R. 2024. ChatGPT: The end of online exam integrity? p. 9
Where the paper says it

; and McIntosh, T. ChatGPT: The end of online exam integrity?

Education Sciences. p. 9
Where the paper says it

A rapid review of the literature. Education sciences, 13(4): 410.

The Unicode Consortium. p. 2

Where the paper says it

of the Unicode Standard (The Unicode Consortium 2022), which mirrors printable ASCII into a range that virtually no font renders. The prompt-security community has documented since 2024 that several frontier LLMs tokenize these tag characters as ordinary text and follow instructions encoded in them, a technique commonly called the ASCII Smuggler (Rehberger 2024a).

2022. The Unicode Standard, Version 15. p. 9
Where the paper says it

The Unicode Consortium. The Unicode Standard, Version 15.

0.0.

Toulas, B. 2025. Google Won’t Fix New ASCII Smuggling Attack in Gemini. p. 9

Where the paper says it

Mountain View, CA: The Unicode Consortium. Google Won’t Fix New ASCII Smuggling Attack in Gemini.

Ueoka, H.; Murawaki, Y. p. 8

Where the paper says it

Second, combining the Unicode-Tag channel with a contentlevel watermark, in the spirit of (Yang et al. 2022; Ueoka, Murawaki, and Kurohashi 2021), would survive a student who lightly paraphrases the brief before pasting it.

; and Kurohashi, S. p. 8
Where the paper says it

Second, combining the Unicode-Tag channel with a contentlevel watermark, in the spirit of (Yang et al. 2022; Ueoka, Murawaki, and Kurohashi 2021), would survive a student who lightly paraphrases the brief before pasting it.

2021. Frustratingly easy edit-based linguistic steganography with a masked language model. p. 9
Where the paper says it

; and Kurohashi, S. Frustratingly easy edit-based linguistic steganography with a masked language model.

Yang, X.; Zhang, J. p. 8

Where the paper says it

arXiv preprint arXiv:2307. ; Zhang, Y.

; Chen, K.; Zhang, W. p. 8
Where the paper says it

arXiv preprint arXiv:2307. ; Zhang, Y.

; Ma, Z.; Wang, F.; and Yu, N. 2022. Tracing text provenance via context-aware lexical substitution. p. 9
Where the paper says it

; and Yu, N. Tracing text provenance via context-aware lexical substitution.

Yeung, K. 2026. Tokenization Attacks on LLMs. p. 9

Where the paper says it

In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, 11613–11621. Tokenization Attacks on LLMs.

Zhao, X.; Ananth, P. p. 9

Where the paper says it

com/research/tokenization-attacks-onllms-how-adversaries-exploit-ai-language-processing. ; Ananth, P.

; Li, L.; and Wang, Y. p. 9
Where the paper says it

; Ananth, P. ; and Wang, Y.

-X. 2023. Provable robust watermarking for ai-generated text. p. 9
Where the paper says it

; and Wang, Y. Provable robust watermarking for ai-generated text.

Zou, A.; Wang, Z.; Carlini, N. p. 9

Where the paper says it

arXiv preprint arXiv:2306. ; Carlini, N.

; Nasr, M.; Kolter, J. p. 9
Where the paper says it

; Carlini, N. ; Kolter, J.

Z.; and Fredrikson, M. p. 9
Where the paper says it

; Kolter, J. ; and Fredrikson, M.

2023. Universal and transferable adversarial attacks on aligned language models. p. 9
Where the paper says it

; and Fredrikson, M. Universal and transferable adversarial attacks on aligned language models.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. inputside watermarking for assignment prompts allows an educator to embed an invisible instruction inside a visible assignment prompt, which an LLM will read and write a tell-tale signature into its reply, exposing the copy-and-paste pathway specifically.

  2. SteganoPrompt is a single-page zero-dependency web tool that encodes an arbitrary printable ASCII payload into the deprecated Unicode Tags block U+E0000–U+E007F and concatenates it with the visible assignment text. This ensures the encoded string is visually identical to the original on every text renderer, survives common copy-paste channels, and is tokenized as ordinary text by most frontier models.

  3. The system allows an instructor to verify the embedded payload before distribution using a Verify panel that computes verify(X) = D rhoT (X), returning W exactly if X is a well-formed watermarked document, thus confirming the literal token presence in a suspect submission.

Abstract

Large language models (LLMs) have made fluent essay writing, code drafting, and quiz answering instantly available to students at every level, from secondary school through graduate study. Many educators do not object to LLM use per se; what they need to detect is the case in which a student pastes the assignment prompt into a chatbot and submits the model's reply verbatim, without engaging with the work. Existing post-hoc AI-text detectors remain unreliable and have been shown to penalise non-native English writers, while output-side watermarks require cooperation from the model provider. We propose an alternative that the educator controls directly: an input-side watermark in which an invisible instruction is embedded inside the visible assignment prompt itself. An LLM that ingests the prompt verbatim quietly reads the hidden instruction and writes a tell-tale signature into its reply, exposing the copy-and-paste pathway specifically. We describe SteganoPrompt, a single-page, zero-dependency web tool that encodes an arbitrary printable-ASCII payload into the deprecated Unicode Tags block (U+E0000--U+E007F). The encoded string is visually identical to the original, survives common copy-paste channels (Word, Google Docs, PDF, Markdown, Slack, e-mail, the major learning-management systems), and is tokenized as ordinary text by most frontier models. We evaluate compliance across eight LLM families and a representative set of educational content channels. To keep deployment transparent, disclosure is built into the tool itself: every encoded brief also carries a brief, always-on academic-integrity reminder for the student, and we set out concrete guidance recommending for its disclosed, ethical use in the classroom. The tool is released under the MIT licence at https://ezharjan.github.io/SteganoPrompt/.

Sources

Related papers