A mentee arrived last week with a question that has become common in recent months. She works with undergraduates and wanted to know how to respond when students ask whether running an outline through a large language model counts as their own work. Early on I might have recited the latest policy language. Now the first step is different: what do the last year of actual cases at comparable institutions show about where lines have been drawn and enforced?
Generative AI has shifted the ground under academic integrity faster than most policies anticipated. Usage numbers tell part of the story. A large-scale survey of more than 95,000 undergraduates across twenty research universities found that roughly two-thirds had used generative AI tools, with nearly 40 percent doing so at least monthly. Among those users, at least 9 percent reported employing the tools to cheat, with daily users showing markedly higher rates than occasional ones.
National figures from the United Kingdom point in the same direction. Proven cases of AI-related misconduct reached nearly 7,000 in the 2023-24 academic year, or 5.1 incidents per 1,000 students, up sharply from the prior year. Projections for the following year suggested further increases. These recorded cases sit against broader surveys showing 88 percent of UK undergraduates had turned to generative AI for assessments by early 2025.
The numbers alone do not capture the daily decisions students face. Many describe AI as a study aid for brainstorming, clarifying concepts, or checking grammar. Others admit crossing into full essay generation under deadline pressure. Surveys of several hundred students indicate that personal ethical beliefs predict behavior more reliably than awareness of institutional rules. Students who view AI output as fundamentally different from traditional plagiarism often treat the boundary as movable rather than fixed.
One recurring theme is the distinction between acceleration and outsourcing. A student who uses the tool to generate practice questions, then works through explanations and checks understanding against primary sources, is building skill. Another who pastes a prompt and submits the result unchanged is not. The same technology supports both paths, which is why blanket prohibitions have proven difficult to sustain.
Photo by Brett Jordan on Unsplash
Faculty concerns track these developments closely. Research released in early 2026 showed 92 percent of instructors worried about AI-enabled plagiarism or dishonesty, with especially sharp unease in writing-intensive disciplines. At the same time, many recognize that detection software remains unreliable, producing both false positives that unfairly flag original work and false negatives that miss generated text. Institutions have therefore turned toward redesign rather than reliance on detection alone.
Examples include a return to in-person examinations in some departments, requirements that students document their process, and prompts that ask for personal reflection or data collected by the student themselves. The UK’s Quality Assurance Agency has advised institutions to state clearly what forms of AI use are acceptable, to require explicit acknowledgment of tool use, and to place responsibility for the final submission on the student.
Policy language helps, yet enforcement still depends on shared understanding. When guidelines remain vague or shift term to term, students report uncertainty about whether a particular workflow will later be treated as misconduct. Training that focuses on ethical reasoning rather than lists of prohibitions appears more effective at shaping decisions. Students who can articulate why a given use preserves or undermines their own learning tend to make more consistent choices under pressure.
Cultural and disciplinary differences matter as well. Fields that already emphasize iterative drafting and collaboration have found it easier to integrate transparent AI use than disciplines built around single-author, high-stakes submissions. International students sometimes face additional layers of confusion when home-country norms around assistance differ from host-institution expectations.
Practical responses are emerging from these patterns. Clear process documentation, scaffolded assignments that build from outline to draft to revision, and explicit discussion of acceptable versus unacceptable prompts all reduce ambiguity. Some programs now include short modules on AI ethics early in the term, treating the tools as part of the contemporary research environment rather than an external threat. The goal in each case is legibility: making visible the intellectual labor that remains the student’s responsibility.
Questions remain about equity. Students with reliable access to premium models or with stronger prior preparation may gain advantages that others lack. Institutions are beginning to address this by providing approved tools through campus licenses and by designing assignments that do not assume equal starting points with the technology.
Photo by Brett Jordan on Unsplash
What the evidence shows so far is that neither outright bans nor unchecked adoption serve students well. The workable middle path treats AI as a powerful but imperfect collaborator whose output must still be evaluated, verified, and integrated by a human author. That standard places the burden on institutions to articulate expectations plainly and on students to document their own contributions honestly.
A small step this month is to choose one assignment or advising conversation and ask explicitly what the student or colleague intends the AI to do and how they will verify or extend the result. The answer often reveals more about the underlying understanding than any policy paragraph can.
