A corpus-based analysis of frame markers in AI-generated and human-written argumentative essays

This research study compares the use of frame markers in 120 argumentative essays generated with ChatGPT (GPT-5.3 Instant; 63,908 words) and 120 human-written essays from the LOCNESS corpus (100,975 words). Using Hyland’s functional model—sequencing, labelling stages, announcing goals, and shifting topics—candidate forms were contextually validated and analysed with the essay as the observational unit. Frequencies were normalized per 10,000 words; essay-level distributions were compared using Mann–Whitney U tests, Holm correction, Cliff’s δ, and 95% confidence intervals. AI-generated essays demonstrated higher corpus rates for labelling stages (14.55 vs 1.19 per 10,000 words; δ = 0.617, adjusted p < 0.001) and announcing goals (7.82 vs 0.89; δ = 0.377, adjusted p < 0.001). Sequencing was higher in the full sample (12.67 vs 7.03; δ = 0.197, adjusted p = 0.007), but this direction reversed in the shared-topic sensitivity subset; topic shifting did not differ significantly (0.63 vs 1.29; δ = −0.057, adjusted p = 0.067). The AI texts were more concentrated (15 validated forms; the top three accounted for 72.8% of markers) than the human texts (29 forms; the top three accounted for 40.0%). These findings describe textual patterns under the reported conditions and do not equate marker frequency with writing quality, coherence, communicative intention, or reader response.