Exams Soar, Writing Slumps in ChatGPT Era, Study Finds
After ChatGPT, online exam scores soared while research marks slipped. For writers: use AI for speed and structure; rely on your judgment, sourcing, and voice for high-stakes work.
AI Made Exams Easier, Research Harder - What That Means for Writers
A large study of more than 3,300 college students found something counterintuitive. After ChatGPT launched, online exam scores jumped by nearly 22 points, yet writing project marks fell by about 10.
Translation for writers: AI boosts fast, structured tasks. It struggles where original thinking, context, and source-aware judgment do the heavy lifting.
In a nutshell
- Exam pass rates rose from about 50% to 86%.
- Research project scores dropped over 10 points, with pass rates falling into the low 70s.
- Students who were already passing generally gained; struggling students improved on exams but saw lower overall marks.
- Creative research proposals showed no change, suggesting little advantage from AI on tasks requiring originality and course-specific methods.
What changed - and why it matters
The course, assessments, textbook, and teaching approach stayed consistent across six semesters. The timing of the shift lined up with the growth in AI usage, especially on online, unsupervised exams.
The pattern points to a simple truth writers see daily: AI is strong on quick answers and structure. It weakens as constraints increase and stakes rise for accuracy, originality, and context.
The signal for working writers
Think of exams as listicles, summaries, and basic briefs - AI performs well there. Think of research projects and proposals as investigative features, white papers, and client-specific strategies - this is where your judgment, voice, and sourcing win.
AI can draft fast. It can also produce polished but inaccurate content if you skip verification. The quality gap now lives in how you set constraints and how you check facts.
How to use AI without losing your edge
- Use AI for scaffolding: outlines, headline options, structure, checklists.
- Inject constraints: audience, brand voice, banned claims, required sources, formatting rules.
- Make sourcing non-negotiable: request citations, follow links, replace weak sources with primary data.
- Run two passes: one for clarity, one for truth. Highlight every claim and verify it.
- Add what AI can't see: interviews, first-hand examples, proprietary data, lived experience.
- Stress-test the draft: ask the model to argue the opposite, list counterexamples, and flag weak logic.
- Track ROI: time saved on scaffolding vs. time spent on fact-checks and edits. Keep what nets out.
- Build a prompt library aligned to your deliverables and style guide. Update it with post-mortems from each project.
For content leads and editors
- Reward depth: weigh originality, sourcing, and insight higher than volume.
- Redesign briefs: require mandatory sources, quotes, and point-of-view sections.
- Separate AI-friendly tasks from human-first work to set turnarounds and expectations.
- Add a verification step before copy hits layout. Treat it like QA, not a favor.
A practical checklist for your next piece
- Define a clear thesis and audience constraints.
- Generate an outline with AI; prune fluff and add unique angles.
- Collect sources first. Prioritize primary data and recent studies.
- Draft fast, then verify every claim and stat.
- Add original inputs: examples, quotes, or a small dataset.
- Rewrite for voice. Cut any sentence that could live in a generic blog.
Why this aligns with the study's details
Easy-to-grade assessments saw the biggest gains. Deeper tasks that require human review either fell or held steady. That mirrors content work: templated outputs get quicker; high-leverage writing still demands expert attention.
Source
Study published in the International Journal of Mathematical Education in Science and Technology. DOI link: https://doi.org/10.1080/0020739X.2025.2539711
