Artificial intelligence (AI)–based feedback systems are increasingly used to support learning at scale, particularly in complex performance domains such as writing. While prior research has demonstrated that AI can generate feedback comparable in quality to human feedback, less is known about in which situations learners prefer AI feedback versus human feedback. Learners' preferences in specific feedback situations can predict whether learners will actively engage with the feedback. In this study, we investigated situation-specific psychological predictors of learners' preferences for AI versus human feedback using a vignette-based within-subject design in a sample of N = 282 US-college students on Prolific. Participants evaluated six feedback scenarios, each varying one of three factors - perceived effort, motivational need salience (autonomy vs. support), or anxiety - at high versus low levels. Each participant evaluated all six vignettes. After each vignette, participants indicated their preference for receiving feedback from an AI system or a human instructor. Mixed-effects logistic regression models were used to analyze feedback source preferences to account for repeated measures within participants for each factor. Results showed that preferences for AI versus human feedback were attributable to the characteristics of the scenarios: when the situation suggested high prior task effort and a salient need for support, scenarios were linked to less AI instead of human feedback preference, whereas high-anxiety scenarios regarding performance evaluation were associated with a substantially higher likelihood of preferring AI feedback. These findings suggest that AI and human feedback may serve complementary roles depending on learners' psychological needs. We discuss implications for the strategic deployment of AI feedback in hybrid and learning-at-scale educational environments.

Writing proficiency is an essential skill for secondary school students and can be supported through high-quality feedback. However, providing individualized feedback on student writing is timeintensive, which often limits its availability in classroom practice. Recent advances in large language models (LLMs) have raised interest in automated feedback as a potential way to scale and supplement teacher practice, yet evidence of its effectiveness relative to teacher feedback in authentic classroom settings remains limited. In this study, we examined the effects of LLM-generated feedback on students' essay revisions, subsequent writing performance, and feedback perceptions in an authentic English-as-a-foreignlanguage context in secondary schools (N = 391). We compared teacher-written feedback against both delayed and immediate LLMgenerated feedback. We further conducted Bayesian analyses to characterize the magnitude and uncertainty of differences between feedback conditions. Together, these findings point to a trade-off between small performance differences associated with teacher feedback that were neither statistically or practically significant. At the same time, results showed more positive perceptions of usefulness and motivation associated with immediate feedback, which only LLMs can provide at scale. These findings suggest that the primary educational value of LLM-based feedback lies less in outperforming teachers and more in redistributing instructional effort. When used to consistently support revision and short-term learning, AI feedback may free teachers to focus on higher-order instructional planning, targeted scaffolding, and pedagogical interaction, highlighting the potential of hybrid feedback systems in writing instruction.

In this paper, we present the DARIUS (Digital Argumentation Instruction for Science) corpus for argumentation quality on 4589 essays written by 1839 German secondary school students. The corpus is annotated according to a fine-grained annotation scheme, ranging from a broader perspective like content zones, to more granular features like argumentation coverage/reach and argumentative discourse units like claims and warrants. The features have inter-annotator agreements up to 0.83 Krippendorff’s α. The corpus and dataset are publicly available for further research in argument mining.