New educational research identifies distinct ways college students rely on generative AI during academic writing. The findings suggest schools should distinguish thoughtful AI use from dependence instead of treating every student who uses AI as academically dishonest.
Editorial Note
This article provides independent educational reporting and analysis based on newly released research concerning undergraduate use of generative artificial intelligence. It does not provide legal advice, academic-integrity rulings, or institutional policy recommendations for a particular school or university.
New To Education is an independent publication. It is not affiliated with, sponsored by, endorsed by, or acting on behalf of the study’s authors, participating institution, Northwestern University, artificial-intelligence companies, colleges, students, or organizations discussed in this article.
The underlying study was posted on July 15, 2026, rather than July 20. It was included in this July 20 report because it is among the newest available pieces of educational research and directly relates to a university research workshop held on July 20 concerning generative AI in computing education. The paper is a preprint and should not be treated as a final peer-reviewed consensus.
The Question Is No Longer Whether Students Use AI
Generative artificial intelligence has moved rapidly from an unfamiliar classroom experiment to an ordinary part of student life.
College students now use AI systems to generate ideas, clarify readings, revise sentences, organize arguments, locate possible sources, explain difficult concepts, create outlines, and sometimes produce entire assignments.
Much of the public debate has treated those activities as though they were essentially the same. A student either used AI or did not use it. The first student is viewed with suspicion, while the second is assumed to have completed the work independently.
New educational research suggests that this simple distinction misses what may matter most.
The study argues that students rely on AI in meaningfully different ways. Some use it deliberately while continuing to evaluate and control their own work. Others use it mainly as a convenient tool for completing isolated tasks. Some engage with it as a conversational partner. Another group may become dependent on it to the point that the technology begins replacing important parts of the learning process.
The difference between those patterns could determine whether AI strengthens learning or weakens it.
Researchers Developed a New AI-Reliance Scale
Researchers Shahin Hossain and Tukhbita Afroz Nawmi developed what they call the Generative AI Reliance Types Scale, or GenAI-RTS.
The 20-item measurement tool was designed to identify how undergraduate students use and depend on generative AI during academic writing.
The researchers drew on survey responses from 382 undergraduate students at a U.S. minority-serving institution. They also conducted interviews with 14 students selected to provide more detailed accounts of how AI affected their writing practices.
Rather than measuring only how frequently students used AI, the researchers attempted to measure the nature of that reliance.
This distinction is important. Two students may both use an AI assistant several times during an assignment while engaging in completely different learning processes.
One might ask the system to challenge an argument, compare possible structures, and identify weaknesses before independently writing the final paper. Another might ask it to generate the paper and then make only minor edits.
A frequency-based survey could classify both students as heavy AI users. A reliance-based framework attempts to show that their behavior is not educationally equivalent.
The Study Identified Five Measurable Patterns
The researchers initially organized AI reliance into four broad categories: strategic, instrumental, dependent, and dialogic use.
Statistical testing suggested that strategic reliance contained two distinguishable parts. The final model therefore identified five measurable dimensions: deliberate use, critical evaluation, instrumental reliance, dependent reliance, and dialogic reliance.
These categories do not necessarily describe permanent types of students. One student may use AI strategically during one assignment and dependently during another.
The framework instead describes patterns of interaction that teachers and researchers may be able to observe and measure.
That could help move school discussions away from the question of whether AI was present and toward the more useful question of what the student actually did with it.
Deliberate Use Keeps the Student in Control
Deliberate use describes situations in which students make conscious decisions about when AI is useful and when they should work independently.
A student using AI deliberately may first identify the purpose of the assignment, decide which tasks can appropriately involve AI, and reserve other tasks for personal reasoning.
The student might use a chatbot to brainstorm possible examples but write the analysis without automated assistance. Another student might ask for feedback on clarity after completing an original draft.
The defining feature is not that AI use is minimal. It is that the student remains responsible for directing the process.
This kind of use resembles the way capable writers use dictionaries, grammar tools, tutors, search engines, or peer feedback. The external support assists the work without fully determining its substance.
Schools trying to develop AI literacy may want to encourage this sense of control rather than simply rewarding students who avoid the technology altogether.
Critical Evaluation Requires Students to Question the Output
The second strategic dimension involves critical evaluation.
Students demonstrating this form of reliance do not assume that an AI-generated answer is accurate merely because it sounds confident or well organized.
They compare claims with reliable sources, check quotations and references, identify possible bias, and decide whether the response actually fits the assignment.
This is one of the most important abilities students can develop in an AI-rich information environment.
Generative systems can produce false information, invented citations, misleading summaries, cultural stereotypes, and arguments that overlook important context. Their fluency can make weak answers appear authoritative.
A student who evaluates AI critically is still performing substantial intellectual work. The system may provide material, but the student must determine what is trustworthy, relevant, ethical, and defensible.
The research found that strategic reliance was positively associated with AI literacy. That connection supports the idea that responsible AI use requires knowledge and judgment rather than mere access to the technology.
Instrumental Reliance Treats AI as a Task-Completion Tool
Instrumental reliance occurs when students use AI primarily to complete specific academic tasks.
A student may ask for a grammar correction, a summary, a citation format, a translation, or a list of possible essay topics.
This type of use is not necessarily harmful. Many ordinary technologies are instrumental. A calculator performs calculations, a spell-checker identifies errors, and a database helps locate research.
The educational effect depends on what the task was supposed to teach.
Using AI to correct a few spelling mistakes may not interfere with an assignment designed to evaluate historical reasoning. Using it to write the central historical argument would be more significant.
The same AI action can therefore be appropriate in one context and inappropriate in another.
Schools need assignment-level rules explaining which forms of assistance are consistent with the learning objective. A universal statement that AI is either permitted or prohibited may be too vague to guide students effectively.
Dependent Reliance May Replace the Learning Process
Dependent reliance is the pattern most likely to concern educators.
It describes situations in which students struggle to begin, continue, or complete academic work without substantial AI assistance.
The technology may start replacing the student’s own planning, reasoning, writing, or problem-solving rather than supporting those processes.
A dependent student might repeatedly ask AI what to think, how to organize every paragraph, what examples to use, and how to respond to feedback. The final paper may appear polished while providing little evidence that the student developed the underlying skills.
Dependence can also create a feedback loop. The more often students avoid difficult intellectual work, the fewer opportunities they have to build confidence and competence. That can make independent work feel even harder the next time.
The concern is not simply academic dishonesty. A student could openly disclose AI use and still fail to learn because too much of the cognitive work was transferred to the system.
Dialogic Reliance Treats AI as a Conversational Partner
Dialogic reliance involves interacting with AI through an extended exchange.
Students may ask follow-up questions, challenge an answer, request alternative explanations, test a developing idea, or use the system to rehearse an argument.
This approach can resemble a conversation with a tutor or study partner.
The educational potential is significant. A student who is reluctant to ask questions in class may feel more comfortable requesting repeated explanations from an AI system. Learners can explore an idea at their own pace without worrying about embarrassment.
However, an AI system is not a human tutor.
It does not reliably recognize when a student misunderstands a concept, and it may reinforce a false assumption rather than correct it. It also lacks the professional responsibility, subject expertise, and personal knowledge that qualified educators bring to instruction.
Dialogic use may promote learning when students remain skeptical and connect the conversation with authoritative course materials. It may become risky when the student treats the system as an unquestionable expert.
Why a Single Category of “AI User” Is No Longer Enough
Many school surveys ask students whether they have used generative AI and how frequently they use it.
Those questions provide useful information about adoption, but they reveal little about learning.
A student who uses AI once to check an outline may develop more independent understanding than a student who uses it every evening as an interactive tutor. A frequent user may be highly critical and strategic, while an occasional user may submit an entirely generated assignment.
Frequency alone cannot explain whether the technology increased or reduced the student’s intellectual engagement.
The new scale attempts to provide a more detailed picture by measuring reliance profiles rather than counting interactions.
That could help researchers examine which forms of use are associated with stronger writing, deeper understanding, increased confidence, weaker critical thinking, or greater academic dependence.
The Scale Was Tested Across Different Student Groups
The researchers examined whether the measurement tool functioned consistently across gender, first-generation college status, and STEM versus non-STEM majors.
They reported evidence that students from these groups interpreted and responded to the scale in sufficiently similar ways for meaningful comparisons.
This is known as measurement invariance.
It matters because a survey can produce misleading comparisons when different groups understand its questions differently. A score should represent the same underlying concept regardless of who completes the instrument.
The study was conducted at a minority-serving institution, giving the research relevance to a student population sometimes underrepresented in early educational-technology research.
Still, one institution cannot represent the full diversity of higher education. The tool will need testing across community colleges, research universities, international institutions, online programs, and schools serving different age groups.
The Research Does Not Prove Which Pattern Produces Better Grades
The study developed and validated a measurement instrument. It was not a randomized experiment assigning students to different forms of AI use.
The researchers identified relationships among reliance patterns, AI literacy, writing processes, and educational outcomes. Those relationships can help generate hypotheses, but they do not by themselves establish cause and effect.
Students with stronger academic skills may be more likely to use AI strategically because they already know how to plan and evaluate writing.
Students who feel underprepared may become more dependent because they are struggling before AI enters the picture.
AI reliance could therefore be both a cause and a symptom of educational difficulty.
Long-term studies will be needed to determine whether dependent use reduces skill development, whether strategic use improves learning, and whether students can move from one pattern to another through instruction.
Teachers Need Evidence of the Process, Not Only the Product
The research supports a larger shift already occurring in assessment.
A finished essay is becoming less reliable as evidence of student learning because AI can produce a polished response within seconds.
Teachers may need to examine the process behind the final submission.
That could include outlines, drafts, notes, source evaluations, prompt histories, revision explanations, brief conferences, oral defenses, or supervised writing samples.
The purpose should not be to create a surveillance system around every assignment.
Process evidence can be part of learning itself. Asking students to explain why they rejected an AI suggestion or how they verified a generated claim can strengthen judgment and reflection.
Students may also be asked to distinguish which parts of an assignment were independently produced, which involved AI assistance, and what intellectual decisions remained their responsibility.
A transparent process gives teachers more meaningful evidence than an unreliable AI-detection score.
AI Detectors Cannot Resolve the Deeper Problem
Schools have increasingly used automated tools that claim to identify AI-generated writing.
Research has repeatedly raised concerns about their accuracy. Detectors can falsely label human writing as generated, overlook AI text that has been edited, and disproportionately flag writing produced by multilingual students.
Even a perfect detector would answer only whether AI likely contributed to the wording.
It would not reveal how the student used it.
A student may have generated an early outline and then independently researched and rewritten the entire paper. Another may have copied an AI response but altered enough language to avoid detection.
The educational question is not simply where the sentences originated. It is whether the student demonstrated the knowledge, reasoning, communication, and decision-making the assignment was intended to develop.
The reliance framework could help educators focus on that deeper issue.
Schools Should Avoid Treating Every AI Use as Cheating
Academic-integrity policies remain necessary. Students should not misrepresent machine-generated work as their own when an assignment requires independent performance.
However, defining every AI interaction as cheating may become increasingly impractical.
AI features are being incorporated into search engines, word processors, learning platforms, translation tools, and accessibility software. Students may use automated assistance without entering a separate chatbot.
A blanket prohibition can also discourage honest disclosure. Students may hide helpful uses rather than discuss them with teachers.
A more workable policy would connect permitted AI use to the purpose of each assignment.
When the objective is independent writing fluency, students may need to compose without generative assistance. When the objective is evaluating evidence, AI-generated claims could become material for critical analysis. When the objective is revising communication, limited feedback may be appropriate.
Clarity helps students make responsible choices before misconduct occurs.
Dependent Use May Reflect Unmet Student Needs
Students may rely heavily on AI for reasons beyond laziness or dishonesty.
Some may lack confidence in academic writing. Others may be multilingual learners attempting to express complex ideas in a second language. First-generation students may be unfamiliar with hidden academic expectations. Students with disabilities may use AI for organization, language support, or accessibility.
Heavy reliance may also reflect large classes, limited instructor availability, inadequate tutoring, or assignment instructions that students do not understand.
Schools should not ignore misconduct, but punishment alone may fail to address why the dependence developed.
A student who cannot begin an essay without AI may need writing instruction, feedback, counseling, language support, disability accommodations, or clearer teaching.
The reliance scale could become useful as a diagnostic educational tool rather than merely another method of identifying prohibited behavior.
AI Literacy Should Include Self-Awareness
Most definitions of AI literacy focus on understanding how the technology works, recognizing bias, checking accuracy, and protecting personal information.
The new research suggests that self-awareness should also be included.
Students need to recognize when assistance is becoming dependence.
They should be able to ask themselves whether they could explain the argument without the chatbot, reproduce the skill under supervised conditions, or make progress if the tool were unavailable.
A student can receive an accurate AI answer and still lose an opportunity to learn.
Responsible use therefore involves deciding not only whether the output is correct but whether requesting it supports the learner’s long-term development.
That is a more difficult skill than writing an effective prompt.
Universities Are Beginning to Study AI as a Learning Environment
On July 20, Northwestern University hosted a workshop focused on generative AI in university-level computing education.
The event brought together researchers and practitioners to examine how AI is changing learning, teaching, and participation in computing courses. Organizers emphasized the need for inclusive and caring educational environments rather than treating AI solely as a technical or disciplinary problem.
The workshop did not publish the GenAI reliance study and should not be represented as a release event for it.
Together, however, the workshop and the newly posted research reflect a broader change in the field.
Educational researchers are moving beyond the early question of whether AI should be allowed. They are beginning to investigate how different forms of interaction affect learning, identity, access, assessment, and student independence.
What Educators Can Do Now
Teachers can make the purpose of each assignment explicit and explain which parts must demonstrate independent performance.
They can provide examples of acceptable and unacceptable AI assistance rather than relying on one general policy for every task.
Students can be asked to document their process, verify generated information, and reflect on how the tool affected their reasoning.
Educators can also include low-stakes opportunities for students to work without AI so that they continue developing independent fluency and confidence.
Schools should provide additional support when students show signs of dependence. The goal should be to restore student agency, not merely catch violations.
Professional development will also be important. Teachers need time to examine AI tools, redesign assessments, discuss difficult cases, and develop consistent expectations across courses.
Key Takeaways
New research has developed a 20-item scale designed to measure how undergraduate students rely on generative AI during academic writing.
The study drew on survey responses from 382 students at a U.S. minority-serving institution and interviews with 14 students.
Researchers identified five measurable dimensions: deliberate use, critical evaluation, instrumental reliance, dependent reliance, and dialogic reliance.
The findings suggest that educators should not treat all AI use as equivalent. Some forms may support planning, evaluation, and conversation, while dependent use may replace important intellectual work.
Strategic reliance was positively associated with AI literacy, although the observational study does not prove that one caused the other.
The paper was posted on July 15, 2026. It was not first published on July 20, although its subject directly overlaps with educational-research discussions taking place on July 20.
Frequently Asked Questions
Was This Study Published on July 20, 2026?
No. The research paper was posted on July 15. It is included as one of the newest available education studies and relates directly to a university research workshop held July 20.
What Is the GenAI-RTS?
It is a 20-item survey instrument designed to measure different ways students rely on generative AI during academic writing.
What Types of AI Reliance Did Researchers Identify?
The final model included deliberate use, critical evaluation, instrumental reliance, dependent reliance, and dialogic reliance.
Is Every Form of AI Reliance Harmful?
No. Strategic or carefully evaluated use may support learning. The educational effect depends on the task, the student’s decisions, and whether the technology assists or replaces the intended learning process.
What Is Dependent Reliance?
Dependent reliance occurs when a student has difficulty completing important academic work without AI and transfers too much planning, reasoning, or writing to the system.
Did the Study Prove That AI Dependence Lowers Grades?
No. The study validated a measurement tool and examined relationships among several variables. It did not randomly assign students to different reliance patterns or establish long-term causal effects.
Can Teachers Use the Scale to Accuse Students of Cheating?
The research is better suited to studying and understanding patterns of reliance. It should not be treated as automatic proof that an individual student committed academic misconduct.
Why Is Measuring Reliance Better Than Counting AI Use?
Frequency does not reveal whether a student used AI critically, instrumentally, conversationally, or dependently. Two frequent users may engage in very different learning processes.
Support New To Education
New To Education provides independent reporting and analysis about educational research, artificial intelligence, teaching, student learning, academic integrity, and the policies affecting schools and universities.
Readers can support our work by sharing this article, exploring our tutoring and professional services, joining the New To Education community, or contributing through the support options available on our website.
Every visit, share, booking, and contribution helps us continue translating emerging educational research into accessible information for students, educators, families, and communities.
Related Articles
China’s July 20 AI Education Debate Asks How Schools Should Change When Technology Can Think Alongside Students
World AI Conference Opens in Shanghai as Global Competition and Governance Take Center Stage
Final Thoughts
The most important educational question about generative AI may not be whether a student used it.
It may be whether the student remained intellectually present.
A learner can use AI frequently while still directing the work, questioning the output, verifying evidence, and making independent decisions. Another can use it only once but allow it to complete the central task the assignment was designed to measure.
Those students should not automatically be treated as having engaged in the same behavior.
The new reliance framework gives researchers and educators a more precise language for examining the difference.
That precision is badly needed. Schools that rely only on bans and detection tools risk punishing honest students, missing hidden dependence, and overlooking forms of AI use that may support meaningful learning.
Students still need independent knowledge and skill. Without those foundations, they cannot recognize when an AI answer is wrong, biased, incomplete, or irrelevant.
They also need opportunities to learn how to work with technology responsibly because AI will be present in many of the careers and institutions they enter.
The goal should not be complete technological dependence or artificial isolation from modern tools.
It should be student agency.
AI should help learners explore, question, revise, and understand. It should not quietly become the author of their ideas, the source of their judgment, or the substitute for skills they never had the chance to build.
The strongest educational policies will therefore ask more than whether AI was used.
They will ask what the student understood before using it, what intellectual work remained with the student, and whether the learner could still think when the system was turned off.
Sources
Hossain and Nawmi — Measuring How Students Rely on Generative AI in Academic Writing: Development and Multi-Source Validation of the Generative AI Reliance Types Scale
https://arxiv.org/abs/2607.14301
Northwestern University — Workshop on Advancing Theory, Research, and Practice for Generative AI in University-Level Computing Education
https://planitpurple.northwestern.edu/event/643002
Education University of Hong Kong — New Study Reveals How Students Move Through Learning With Peers and AI
https://www.eurekalert.org/news-releases/1133970
OECD — Digital Education Outlook 2026
https://www.oecd.org/en/publications/oecd-digital-education-outlook-2026_062a7394-en.html_
