Study Finds AI-Assisted Writing Demands More Thought

A peer-reviewed study, "Threshold concepts for writing with AI: Experimentation, expertise, agency," published in Computers and Composition, examined an experimental AI and Writing course at Iowa State University that followed 38 undergraduate students across 22 majors through fall 2023 and fall 2024 (Computers and Composition; The Jerusalem Post). Researchers Abram Anders (Iowa State) and Emily Dux Speltz (Embry-Riddle Aeronautical University) analyzed student reflections after AI writing exercises and self-directed projects, identifying three concepts they found essential to productive use: that writing with AI is an experimental process, that it requires human expertise and dialogue, and that it should strengthen rather than replace the writer's own agency (The Jerusalem Post). One journalism student found ChatGPT could help draft leads but did not follow expected journalistic structure until the student supplied context about lead-writing conventions, illustrating how AI output improved once students brought their own domain expertise to the exchange.

Course exercises were built around the specific problem of linguistic fluency masking factual inaccuracy. In a "Create a Fluent Hallucination" assignment, students deliberately generated false but plausible AI outputs, including fabricated events and invented sources, to practice recognizing when confident-sounding AI text is wrong. Other assignments included a prompt competition, an "AI Ethics Tutor" exercise, and a task where students decided which parts of a writing process a human versus an AI assistant should handle.

The study suggests assessment and tool design should explicitly teach and reward verification behavior rather than assume it emerges naturally. For educators, that means rubrics that credit source-checking, iterative prompt refinement, and disclosed human-AI division of labor. For developers of AI writing tools, it points to demand for built-in provenance signals, confidence indicators, or citation-checking features that help users catch fluent-but-false output rather than relying solely on user skepticism.

Whether the threshold-concepts framework (experimentation, expertise, agency) is replicated at larger scale or across different institutions and model families, and whether the study's course materials and assignment templates (a "Fluent Hallucination" exercise, an "AI Ethics Tutor" task) get adopted or adapted by other writing programs. The authors themselves note further research is needed to test whether the reported shifts in student thinking persist over time or translate into measurably stronger writing.

This adds to a broader pattern in AI-in-education research: the more consequential question is rarely whether students use AI, but whether their institutions teach the specific skill of verifying AI output against source material and domain expertise. Courses that build that skill directly into the curriculum, rather than treating AI use as inherently suspect, appear to produce more capable AI-assisted writers - though this remains a single, small-sample study rather than a settled result.