The proliferation of generative AI in academic contexts has revealed a fundamental truth that institutions have been reluctant to acknowledge:
AI detection tools achieve accuracy rates often below 80% in independent testing.11 Wakjira et al. (2025) False positive rates can be as high as 50% across widely-used platforms.22 Weber-Wulff et al. (2023) There is also documented systematic bias, with over 61% of non-native English writing flagged as AI-generated.33 Liang et al. (2023) The current approach of “detect and punish” thus creates more harm than it prevents.
Meanwhile, the integration of AI into scholarly work has already occurred at scale. Studies indicate that 13.5% to 22.5% of academic papers now show evidence of AI assistance.44 Liang et al. (2025a) Peer review processes are similarly affected, with at least 15.8% of reviews showing AI assistance in some conferences.55 Latona et al. (2024) This whitepaper argues that the question facing higher education is no longer how to prevent AI use, but rather how to build institutional frameworks that make AI integration transparent, bounded, and pedagogically sound.
The path forward requires abandoning unreliable surveillance in favor of transparency architectures—tools and policies designed from inception to make AI contributions visible, auditable, and appropriately constrained. This shift represents not a capitulation to technological inevitability, but an opportunity to re-imagine academic integrity for a new era.
Part I: The Epistemological Limits of AI Detection
The widespread deployment of AI detection tools in academic settings rests on a flawed premise: that AI-generated and human-generated text represent distinguishable categories that can be reliably separated through algorithmic analysis. Empirical evidence accumulated since late 2022 demonstrates that this assumption fails under rigorous testing.66 Weber-Wulff et al. (2023); Sadasivan et al. (2023)
The most comprehensive independent evaluation of AI detection tools to date, conducted by Weber-Wulff and colleagues and published in the International Journal for Educational Integrity, tested fourteen detection tools, including twelve publicly available platforms and two commercial systems.77 Weber-Wulff et al. (2023) The findings were unequivocal: all tools scored below 80% accuracy, with only five exceeding 70%. The study concluded that these tools are neither accurate nor reliable, exhibiting a main bias toward classifying output as human-written rather than detecting AI-generated text.
The performance degradation becomes more severe under realistic conditions. Studies indicate that even with relatively straightforward AI-generated text, detection tools exhibit notable issues, and their accuracy drops significantly when faced with manipulated content or obfuscation techniques such as paraphrasing.88 Perkins et al. (2024c); Weber-Wulff et al. (2023) Some AI content detectors have been found ineffective in identifying paraphrased texts.99 Liu et al. (2024)
Turnitin, a widely used commercial system in academia, has attempted to address limitations through deliberate calibration, reportedly accepting a certain percentage of false negatives to maintain a low false positive rate at the document level.1010 Jabarian and Imas (2025) This strategic trade-off, while a product design choice, underscores a fundamental challenge: detection systems often cannot simultaneously achieve high sensitivity (catching AI-generated content) and high specificity (avoiding misidentifying human-written content). The difficulty for traditional AI detectors to reliably differentiate between legitimate human paraphrasing and AI generation undermines their utility for academic integrity enforcement.1111 Bordalejo et al. (2025)
The theoretical foundations for this limitation were established by Sadasivan and colleagues at the University of Maryland.1212 Sadasivan et al. (2023) Their mathematical analysis demonstrated that for sufficiently capable language models, even optimal detectors can only marginally outperform random classification, especially after applying transformations like paraphrasing.1313 Cheng et al. (2025); Sadasivan et al. (2023) As Professor Soheil Feizi summarized, “There is an upper bound on our detectors that fundamentally limits them, so it’s very unlikely that we’ll be able to develop detectors that will reliably identify AI-generated content.”1414 University of Maryland (2025)
This is not merely a technical limitation awaiting an engineering solution. As AI language models improve and their outputs converge with human writing patterns, the statistical signatures that detection relies upon diminish. This convergence makes it increasingly difficult for both human users and automated systems to reliably distinguish between human-generated and AI-generated text.1515 Knott et al. (2024); Frank et al. (2024) Sadasivan and colleagues provided a theoretical impossibility result indicating that as language models become more sophisticated and better at emulating human text, the performance of even the best-possible detector decreases. This suggests a fundamental constraint on detection accuracy that cannot be overcome through improved algorithms.
Part II: The Systemic Transformation of Scholarly Production
While institutions are hyper-aware of AI use in undergraduate contexts, they are not noticing the far more significant transformation that has already occurred within professional academic research. The evidence suggests that AI-assisted writing has become normative in scholarly publishing within two years of ChatGPT’s release.1616 Matsui (2025); Yip et al. (2025)
The most methodologically rigorous study of this phenomenon, published by Kobak and colleagues in Science Advances in July 2025,1717 Kobak et al. (2025) analyzed 15.1 million English-language biomedical abstracts indexed by PubMed from 2010–2024. Using an “excess word” methodology inspired by epidemiological studies of COVID-19 mortality, the researchers identified words appearing more frequently than predicted by pre-ChatGPT trends. Their findings indicate that at least 13.5% of 2024 biomedical abstracts—approximately one in seven—were processed with large language models. This lower bound varied significantly by subcategory, reaching 40% for certain disciplines, countries, and journals.
The researchers identified 379 “style words” revealing AI usage, including terms like “delve,” “crucial,” “meticulous,” and “intricate” that showed dramatic frequency increases post-November 2022. The study concluded that “LLMs have had an unprecedented impact on scientific writing in biomedical research, surpassing the effect of major world events such as the COVID pandemic.”1818 Kobak et al. (2025)
The peer review process has been similarly affected. Liang and colleagues’ analysis presented at the International Conference on Machine Learning in 20241919 Liang et al. (2024) found that between 6.5% and 16.9% of peer reviews at major AI conferences were substantially modified by LLMs—defined as modifications beyond mere spell-checking. The study examined reviews from ICLR 2024, NeurIPS 2023, CoRL 2023, and EMNLP 2023, finding that word frequencies for terms like “commendable” increased 9.8-fold and “meticulous” increased 34.7-fold in ICLR 2024 reviews compared to pre-ChatGPT baselines. Behavioral patterns associated with higher LLM usage included lower reported confidence, submissions closer to deadlines, and decreased likelihood of responding to author rebuttals.
Institutional responses are beginning to acknowledge this reality. In October 2025, arXiv reportedly indicated changes to its computer science category regarding review articles or position papers.2020 Sun et al. (2025); Jianghao et al. (2025) This decision was reportedly driven by a significant increase in LLMassisted survey papers, which were often described as little more than annotated bibliographies.2121 Liang et al. (2025b); Wang et al. (2024) The platform receives hundreds of such articles monthly and faces challenges in distinguishing quality submissions.
These data points describe a systemic transformation of scholarly production, not an isolated student cheating problem. Policies focused on undergraduate detection while professional practice evolves unchecked represent a category error in institutional response.
Part III: Detection Bias as an Equity Crisis
Beyond unreliability, AI detection tools exhibit systematic bias that transforms their deployment into a civil rights concern.2222 Gorichanaz (2023); Giray (2024) This bias is most striking for non-native English speakers, but also extends to neurodivergent students and other populations whose writing patterns diverge from algorithmic expectations of “humanness.”2323 Gegg-Harrison and Quarterman (2024); Chambers and Kelley (2025)
The landmark study on this issue, published by Liang and colleagues in Patterns, evaluated seven widely-used GPT detectors against two corpora: 91 TOEFL essays written by non-native English speakers and 88 eighth-grade essays written by native English speakers.2424 Liang et al. (2023) While detectors achieved near-perfect accuracy for native speaker essays, they misclassified over half of the non-native English writing samples as AI-generated. Specifically, GPT-based detection tools misclassified 61.3% of non-native English speaker samples.2525 Perkins et al. (2024c) The bias was pervasive: 97.80% of TOEFL essays were flagged as AI-generated by at least one detector, and 19.78% were unanimously misclassified by all seven.2626 Arango et al. (2024)
The mechanism underlying this bias is algorithmic rather than coincidental.2727 Liang et al. (2023) Detection systems operate by measuring “perplexity”— how predictable word choices are.2828 Park (2024) Non-native English speakers naturally produce text with less linguistic diversity and more predictable word choices due to limited lexical richness, reduced syntactic diversity, and lower grammatical complexity. The TOEFL essays that were unanimously misclassified showed significantly lower perplexity. The system thus penalizes exactly the linguistic patterns that characterize competent second-language writing.
The civil rights implications are substantial. English Language Learner students—comprising over 10% of the U.S. student population2929 Alejandro (2024)—are protected under Title VI of the Civil Rights Act on the basis of race and national origin. False accusations based on AI detector bias could constitute actionable discrimination,3030 Jacob et al. (2024); Ovalle (2023); Jiang et al. (2024) a concern highlighted in educational contexts. For international students, such accusations can extend beyond academic penalty, potentially impacting their immigration status and causing significant harm to their academic careers.3131 Bannister et al. (2024); Giray (2024) Students often find themselves in a position where they must defend their work against algorithmic accusations.3232 Gorichanaz (2023)
Neurodivergent students face parallel vulnerabilities. Research demonstrates that neurodivergent writers, similar to second-language writers, are more likely to receive false positive results from AI text-detection tools.3333 Gegg-Harrison and Quarterman (2024); Chambers and Kelley (2025) The mechanism mirrors the non-native speaker problem: autistic students often exhibit writing patterns characterized by rigid style, limited vocabulary, and difficulties expressing thoughts, which may inadvertently resemble algorithmic expectations of AI output.3434 França et al. (2024); Chambers and Kelley (2025); Carik et al. (2024)
Detection-based enforcement thus creates a two-tier system based on linguistic privilege. Students whose writing naturally matches algorithmic expectations of “humanness”—typically native English speakers with neurotypical communication patterns—face minimal risk. Those whose legitimate writing diverges from these expectations bear a disproportionate risk of accusation.3535 Giray (2024)
Part IV: Transparency Architecture as an Alternative Paradigm
If detection fails both technically and ethically, what alternative exists? The answer lies in building transparency into AI tools themselves—designing systems that make AI contributions visible and bounded rather than attempting to identify them retrospectively through surveillance.
This approach draws on the Transparency by Design framework articulated by Felzmann and colleagues in Science and Engineering Ethics, which adapts principles from GDPR’s Privacy by Design to AI contexts.3636 Felzmann et al. (2020) The framework proposes integrating transparency “in a step-by-step manner and as an ex-ante value, not as an afterthought.” For academic AI tools, this means designing systems where AI contributions are inherently traceable rather than requiring post-hoc forensic detection.
Jenni AI was developed with this transparency-first philosophy as a foundational design principle. Rather than positioning AI as a black box that generates content for subsequent human review, the platform’s architecture ensures that human judgment remains central to the writing process while AI assistance remains visible and bounded. Several specific design choices distinguish this approach from bulk generation systems.
Sentence-level autocomplete rather than bulk generation maintains human engagement throughout the writing process. Jenni AI’s autocomplete function produces one sentence at a time, requiring users to review and actively accept each suggestion before proceeding. This contrasts with systems that generate entire paragraphs or documents, which create conditions for wholesale substitution of AI for human thought. The engagement pattern matters: sentence-by-sentence interaction keeps the human writer continuously making decisions about content, structure, and argumentation rather than passively reviewing AI output.
Integrated citation verification through Jenni’s Research Library addresses a core vulnerability of AI-assisted academic text: fabricated or misattributed sources. Jenni AI enables users to import existing research materials via BibTeX files, DOI/PMID/arXiv identifiers, or direct PDF upload. AI suggestions are then grounded in these verified sources, with citations linked to specific documents in the user’s library.3737 Walters and Wilder (2023); Spennemann (2025); Mustaffa et al. (2025)
Configurable contribution limits enable institutions and users to enforce appropriate boundaries on AI assistance. Research indicates that AI tools are frequently used and viewed as acceptable for supporting routine tasks such as improving readability, grammar checking, literature reviews, information seeking, editing, and structuring academic papers.3838 Shimray and Subaveerapandiyan (2025); Xu (2025); Liao et al. (2024); Arroyo-Machado et al. (2025) This aligns with a broader understanding that while AI can assist in these areas, core analytical and argumentative work is expected to remain human-authored.3939 Kwon (2025); Rughiniș et al. (2025)
Multilingual support addresses the equity concerns documented in Part III. Jenni AI operates across more than thirty languages, enabling researchers to work in their native language while producing English-language output. This design choice acknowledges that the goal of academic writing is clear communication of ideas, not demonstration of native English fluency. Non-native speakers using AI for language assistance are not circumventing academic integrity requirements; they are using tools to communicate ideas they have independently developed.4040 Filimonovic and Rutzer (2025); Prakash et al. (2025); Amano et al. (2025); Rueda et al. (2025); Sebastian and Baron (2024); Lepp and Sarin (2024)
Process documentation and export flexibility support institutional disclosure requirements. Jenni AI maintains records of AI-assisted versus human-authored content throughout the writing process, aligning with general recommendations for transparency and documentation in academic integrity.4141 MLA-CCCC Task Force (2023)
This transparency-first approach repositions AI writing tools as ethical accelerants rather than integrity threats. The tool itself becomes the disclosure mechanism, making AI contributions visible by default rather than requiring after-the-fact forensic investigation.
Part V: Institutional Implementation Pathways
Transitioning from detection to transparency requires coordinated changes in policy, assessment design, and faculty development. Several leading institutions have begun this shift, providing models for broader adoption.
Moving from detection to disclosure requirements represents the foundational policy change. Cornell University’s guidance explicitly states: “We currently do not recommend using current automatic detection algorithms for academic integrity violations using generative AI, given their unreliability.”4242 Cornell University (2025) Similarly, Columbia University acknowledges that “AI detection should be treated as a guideline and not a grading metric,” recognizing “risks of misidentification, which can have consequences in the classroom.”4343 Columbia University (2025) The University of Pittsburgh has declined to endorse or support any AI detection tools, citing “substantial risk of false positives.”4444 University of Pittsburgh (2025a)
The alternative is clear disclosure requirements with procedural protections. COPE’s position statement establishes the framework: authors using AI must be “transparent in disclosing in the Materials and Methods (or similar section) of the paper how the AI tool was used and which tool was used,” while remaining “fully responsible” for all content.4545 da Veiga (2025)
Assessment redesign offers more robust integrity evidence than detection. The MLA-CCCC Joint Task Force recommends requiring students to share and comment on version history for written work, using tools that replay writing history to make visible “when text was copied and pasted from another source.”4646 MLA-CCCC Task Force (2023)
Interactive oral examinations, where students articulate reasoning in real time, verify understanding regardless of how written products were produced.4747 Theobold (2021); Gaines and Burrows (2024); Smith and Tanner (2010) Portfolio approaches with scaffolded reflections document learning processes across time, making wholesale AI substitution impractical.4848 Chan and Colloton (2024); Generative AI in Higher Education (2025b); Kong et al. (2024); Pedersen et al. (2018)
The AI Assessment Scale developed by Perkins, Furze, Roe, and MacVaugh provides a five-level framework for calibrating appropriate AI integration across assignment types.4949 Perkins et al. (2024b); Perkins et al. (2024a) Level 1 applies to supervised settings testing knowledge retention. Levels 2 and 3 permit AI for ideation and editing respectively. Level 4 allows AI-assisted writing with clear disclosure. Level 5 enables full AI exploration with focus on human curation and oversight. Pilot implementation at one institution showed 5.9% increase in student attainment and 33.3% increase in module passing rates while reducing academic misconduct cases.5050 Furze et al. (2024)
Faculty development is prerequisite to effective implementation. AI literacy frameworks, which often include competency areas such as technical understanding, evaluative skills, practical application, and ethical considerations, guide this development.5151 UNESCO (2024); Tadimalla and Maher (2024); Chiu et al. (2024) Faculty-specific outcomes include developing AI-informed assignments, creating course-specific policies, and fostering responsible AI usage culture.5252 Gupta et al. (2024)
Institutional investment in approved AI tools with transparency features enables consistent implementation, with some universities providing university-managed access and privacy protections for AI tools.5353 Ng et al. (2025); Wu et al. (2024) Many universities, including Carnegie Mellon, are engaged in discussions and policy development to manage AI tools, focusing on institutional authentication and data protection.5454 Pastva et al. (2024)
The shift from detection to transparency requires cultural change alongside policy revision.5555 Zhai et al. (2024); Miah (2024); Laflamme and Bruneault (2025); Lund et al. (2025) The broader academic discourse emphasizes that AI transparency is crucial for intellectual honesty in academic settings.5656 Zhang et al. (2025) The goal is neither prohibition nor permissiveness, but the development of reflective practice regarding when, how, and why AI assistance serves legitimate purposes.5757 Hysaj et al. (2025); Figueroa et al. (2024)
Conclusion
The evidence presented in this whitepaper leads to an uncomfortable but unavoidable conclusion: institutions cannot detect their way to academic integrity in the age of generative AI.
Detection fails on its own terms. Independent testing shows accuracy rates below 80%,5858 Weber-Wulff et al. (2023) false positive rates as high as 50% for widely-used platforms,5959 Walters (2023) and systematic bias that flags over 61% of non-native English writing as AI-generated.6060 Liang et al. (2023) These are not bugs awaiting patches. The theoretical impossibility result from Sadasivan and colleagues demonstrates that as language models improve, even optimal detectors converge toward random classification.6161 Sadasivan et al. (2023) Detection is not temporarily inadequate; it is fundamentally constrained.
Meanwhile, the transformation institutions hoped to prevent has already occurred. At least 13.5% of biomedical abstracts show evidence of AI processing,6262 Kobak et al. (2025) 22.5% of computer science papers evidence AI modification,6363 Liang et al. (2025a) and up to 16.9% of peer reviews at major conferences are substantially AI-assisted.6464 Liang et al. (2024) Policies that focus on undergraduate essays while professional scholarship evolves unchecked address a symptom while ignoring the condition.
The alternative is not permissiveness but transparency. Tools designed with sentence-level engagement, integrated citation verification, and contribution tracking make AI use visible by default rather than detectable after the fact. Combined with disclosure requirements aligned with COPE guidelines6565 da Veiga (2025) and process documentation as recommended by the MLA-CCCC Joint Task Force,6666 MLA-CCCC Task Force (2023) transparency-first approaches maintain academic integrity without the errors, inequities, and adversarial dynamics of detection-based enforcement.
The question facing institutions is no longer whether to permit AI use. That question has been answered by faculty, researchers, and students who have already integrated these tools into their work. The question now is whether AI use will be pedagogically integrated or covertly substituted.
Institutions that continue investing in detection alone will find themselves administering systems that harm the students they are meant to serve—disproportionately flagging non-native speakers and neurodivergent writers while sophisticated users evade detection entirely. Those that shift toward transparency can shape emerging norms rather than react to them, building frameworks that distinguish legitimate assistance from academic fraud through visibility rather than surveillance.
The path forward requires concrete action. Institutions should audit their current detection tool contracts against the accuracy and bias data presented here. They should develop disclosure policies that treat AI assistance as declarable rather than prohibited. They should invest in faculty development that builds AI literacy across functional, ethical, and rhetorical dimensions. And they should evaluate transparency-oriented tools—platforms designed to make AI contributions traceable—as infrastructure for the new academic environment rather than threats to the old one.
This is not capitulation to technological inevitability. It is recognition that academic integrity has always depended on norms, expectations, and honest disclosure rather than perfect surveillance. The goal remains unchanged: ensuring that credentials reflect genuine learning and that scholarly work represents authentic intellectual contribution. The method must change because the environment has changed. Detection answered a question that no longer admits of reliable answers. Transparency addresses the reality institutions now face.
During the preparation of this whitepaper the authors used Jenni AI in order to refine prose and verify citation formatting, as well as improve the clarity and quality of written communication. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication.





