A field study conducted at the Technical University of Munich demonstrates that restructuring master’s level engineering coursework from traditional report generation to systematic error identification and correction significantly diminishes the operational effectiveness of generative artificial intelligence chatbots while safeguarding student mastery of higher-order academic competencies. Published in Human-Centric Intelligent Systems by researchers Daniel Renjewski and Alexandra Strasser, the investigation evaluated how graduate engineering students navigated written assessments when tasked with auditing deliberately corrupted technical documentation rather than generating original prose from scratch. Conducted within a specialized biorobotics curriculum, the experimental assignment format sharply lowered students’ reliance on automated language tools, prompting many participants to report that conversational software actively hindered their problem-solving workflows. Simultaneously, subjective self-assessments revealed that learners continued to achieve core educational benchmarks spanning Bloom’s revised taxonomy. By shifting the primary cognitive burden from composing descriptive passages to executing rigorous analytical critique, the instructional intervention established an empirical framework for designing asynchronous university examinations that remain resilient against generative automation.
The impetus for redesigning university assessment strategies stems from an escalating pedagogical dilemma confronting educators across engineering disciplines. In recent academic cycles, the widespread accessibility of advanced large language models has transformed homework evaluation, frequently leaving course instructors in the untenable position of assessing synthetic text generated by machine learning systems rather than authentic student analysis. Earlier departmental observations indicated that conventional report assignments—which traditionally evaluated data processing, visual presentation, and argumentative structure—became vulnerable to unreflective copy-and-paste practices that bypassed the deep reasoning necessary to internalize core principles. Compounding this challenge, automated computational tools designed to detect artificial intelligence text generation consistently demonstrated unreliability, failing to provide an equitable or dependable safeguard for academic integrity. Rather than abandoning asynchronous written evaluation entirely or reverting strictly to invigilated examinations, the researchers sought an instructional design capable of reducing the utility of generative tools. Specifically, the objective was to prevent conversational systems from autonomously completing assignments while ensuring that students engaged directly in disciplinary reasoning.
To evaluate this alternative paradigm, the research team implemented a field-experimental protocol during the fourteen-week summer term of 2025 within the master’s course Applied Biorobotics at the Technical University of Munich’s School of Engineering and Design. Established originally in 2018 under principles of constructive alignment, the course limits enrollment to twenty-four participants, drawing students from graduate programs in engineering, physics, and computer science. The curriculum operates under an inverted classroom model where lectures are replaced with instructional videos and technical readings, allowing classroom hours to center exclusively on collaborative problem-solving, analytical discussions, and interactive exploration. Over the semester, students complete a learning portfolio of ten sequential technical assignments, receiving individualized formative feedback and numerical marks ranging from zero to four points on each deliverable. To establish commitment, twenty-three regularly attending students co-signed a formal learning agreement with the instructor before coursework commenced. Additionally, during the opening week, an educational AI specialist provided structured instruction outlining best practices and cautionary examples for deploying large language models in scientific writing.
The core pedagogical intervention occurred in Assignment VI, which required students to examine a bipedal locomotion problem utilizing a simplified spring-mass walking model. In previous course iterations, students completed this module by authoring an original report on the mechanics and simulation results. For the redesigned assignment, instructors substituted the generative task with an error-correction exercise built around a hypothetical peer submission. Each student received a report attributed to a virtual classmate named Kelly Doe and was instructed to audit the work for technical inaccuracies and incompleteness. Students had to strike out erroneous content, highlight their additions using distinct frames or colors, and submit a corrected report alongside necessary simulation code in a digital archive. To prevent misconceptions, the faulty reports were explicitly marked as flawed artifacts rather than authoritative instructional guides. Marking followed the standard four-point rubric, awarding credit for accurate error detection and valid correction. Detailed individual feedback accompanied numerical grades to resolve any remaining misunderstandings.
To curate authentic and pedagogically demanding materials, the teaching staff analyzed prior cohorts’ submissions, identifying recurring modeling errors, miscalculations, and interpretative flaws. Using these real-world pitfalls, the instructors generated a baseline report containing domain-specific conceptual errors, such as flawed modeling assumptions, sensor data misinterpretations, reasoning errors in the discussion, and discrepancies between figures and text. Purely orthographic, grammatical, or structural errors were deliberately minimized to ensure that evaluation focused strictly on disciplinary competence. Because student collaboration was permitted throughout the course, the researchers distributed four distinct permutations of the faulty report, each containing a different subset of errors of calibrated difficulty. This permutation prevented students from discovering errors through simple document comparison, requiring each participant to develop an independent verification strategy. The standardized structure also kept grading manageable for instructors, who utilized pre-highlighted master sheets to quickly assess student markups against the specific error permutation assigned to each student.
Alongside the classroom trial, the researchers conducted a heuristic benchmark to evaluate how state-of-the-art artificial intelligence models handled the error-correction task. Using ChatGPT-5, developed by OpenAI, the team tested faulty reports 1, 3, and 4 due to pragmatic resource constraints, supplying the model with complete PDF documents, LaTeX source files, bibliographies, and graphical figures. The testing prompt instructed the model to audit the reports holistically, disregard language or formatting issues, and deliver structured breakdowns identifying each error, explaining the flaw, and proposing a concrete fix. The benchmark demonstrated significant operational gaps in the automated model’s capabilities: ChatGPT-5 detected between 50% and 79% of intended errors and successfully corrected between 50% and 71%. For unintended errors noticed after release, the model detected 50% to 75% of cases and corrected 38% to 62%. Furthermore, the chatbot identified irrelevant phantom errors in Reports 1 and 4, while in Report 3 it proposed a faulty correction that directly contradicted the underlying technical data.
In the classroom, student performance demonstrated that the redesigned assignment supported robust learning without creating insurmountable evaluative obstacles. Across the cohort, twenty-four students submitted Assignment II, twenty-three submitted Assignment IV, and twenty-two completed Assignment VI, with twenty-three students ultimately receiving a final course grade. On Assignment VI, fifteen of the twenty-two submitting students produced satisfactory solutions characterized by only minor errors or better, while only two students submitted insufficient work failing to meet passing standards. The instructor observed no widespread difficulties in identifying conceptual errors and noted that valid theoretical concepts were reliably recalled, alleviating prior didactic concerns that confronting flawed manuscripts might cement lingering misconceptions. Each assignment contributed to a cumulative point total across the ten portfolio modules, which converted to the standard German grading scale from 1.0 for excellent mastery down to 4.0 for passing and 5.0 for failure. The overall scoring distribution affirmed that students successfully adapted to the diagnostic examination format while demonstrating authentic mastery of technical locomotion principles.
Longitudinal survey findings administered through the SoSci survey platform across four collection points captured marked behavioral changes regarding artificial intelligence adoption. While declared tool usage reached 95.45% in Assignment II and 100% in Assignment IV, it dropped noticeably to 77.27% in Assignment VI. Furthermore, student perceptions of tool utility shifted significantly: whereas all respondents reported that artificial intelligence was helpful to varying degrees on traditional assignments, perceived helpfulness dropped sharply for the error-correction task, with nearly a quarter of respondents reporting that artificial intelligence actually hindered their progress. When evaluating submission quality, all respondents had reported positive or neutral contributions from artificial intelligence during Assignments II and IV; for Assignment VI, however, over 17% stated that using conversational models negatively affected the quality of their work. These empirical responses indicate that transforming the assignment into an error-auditing exercise noticeably decreased the practical utility of generative chatbots, discouraging uncritical reliance on automated assistance.
Detailed inquiries regarding specific workflows revealed that artificial intelligence usage declined across almost all designated activities, including finding ideas, structuring text, and explaining conceptual principles. Notably, the steepest decline occurred in finding and fixing errors—the explicit objective of Assignment VI. The only task showing an increase in automated usage was generating isolated report sections. When reflecting on task demands, over 77% of students did not disagree that starting from an existing report felt easier than facing a blank page, yet half agreed that the new format required a broader problem-solving approach. Only 20% and 25% disagreed that the approach was less clear or felt less familiar. Regarding learning outcomes evaluated against Bloom’s revised taxonomy, consistently more than 60% of students agreed or strongly agreed that they accomplished all stages of intended learning goals, while a maximum of 16% reported not reaching them, confirming that diagnostic tasks preserved high-level cognitive engagement.
The authors acknowledged several methodological limitations, noting that the study was conducted within a single small master’s course and evaluated against a single computational benchmark. Because surveys were pseudonymized to protect student privacy, instructors could not directly link self-reported perceptions to individual assignment grades, underscoring the need for future research to connect self-reports with observable performance measures. Furthermore, student artificial intelligence literacy did not exhibit substantial directional shifts between the beginning and conclusion of the term, likely because participating engineering students already possessed high familiarity with generative tools. Additionally, developing corrupted reports proved labor-intensive, though the authors noted that future iterations could automate error generation using historical submissions. Despite these constraints, the study provides a viable template for higher education, demonstrating that shifting written assessments from generative prose to diagnostic error correction effectively reduces automated shortcutting while preserving authentic reasoning and critical engagement in artificial intelligence environments.
Subject of Research: AI-resilient exam design using error-correction tasks in engineering education
Article Title: Designing AI-Resilient Exams: Error-Correction Assignments in Engineering Education
Article References: Renjewski, D., & Strasser, A. (2026). Designing AI-Resilient Exams: Error-Correction Assignments in Engineering Education. Human-Centric Intelligent Systems, 6(3), 377-389. https://doi.org/10.1007/s44230-026-00157-2
Image Credits: AI Generated
DOI: 10.1007/s44230-026-00157-2
Keywords: engineering education, generative AI, AI-resilient assessment, error-correction assignments, bipedal robotics, Bloom's taxonomy, portfolio examinations, Applied Biorobotics, Technical University of Munich, academic integrity, ChatGPT testing, spring-mass walker, heuristic benchmarking, higher-order learning, inverted classroom, constructive alignment, formative feedback
Cite Scienmag News
Denise Maddox. (November 18, 2026). Error-Correction Engineering Tasks Reduce Generative AI Effectiveness While Preserving Student Learning. Scienmag. https://scienmag.com/error-correction-engineering-tasks-reduce-generative-ai-effectiveness-while-preserving-student-learning/
Denise Maddox. "Error-Correction Engineering Tasks Reduce Generative AI Effectiveness While Preserving Student Learning." Scienmag, 18 November 2026, https://scienmag.com/error-correction-engineering-tasks-reduce-generative-ai-effectiveness-while-preserving-student-learning/. Accessed 18 November 2026.
Denise Maddox. "Error-Correction Engineering Tasks Reduce Generative AI Effectiveness While Preserving Student Learning." Scienmag. November 18, 2026. https://scienmag.com/error-correction-engineering-tasks-reduce-generative-ai-effectiveness-while-preserving-student-learning/








