Ethics and bias

EduEval is a comprehensive hierarchical benchmark for evaluating LLMs in Chinese K-12 education, comprising 24 task types with over 11,000 questions organized across six cognitive dimensions (Memorization, Understanding, Application, Reasoning, Creativity, and Ethics) based on Bloom's Taxonomy and Webb's Depth of Knowledge. The benchmark uses authentic educational materials including real exam questions, classroom dialogues, student essays, and expert-designed prompts spanning primary through high school levels.

This paper presents LearnLM-Tutor, a fine-tuned Gemini model for educational use, and introduces a comprehensive evaluation framework spanning seven diverse benchmarks (quantitative, qualitative, automatic, and human evaluations) grounded in learning science principles to assess pedagogical quality in K-12 AI tutoring systems. The work includes real-world deployment at Arizona State University and systematic evaluation of pedagogical dimensions including Socratic dialogue, adaptive scaffolding, and learning-centered interactions.

EduGuardBench is a dual-component benchmark designed to evaluate LLMs acting as simulated teachers, measuring both pedagogical fidelity (role-playing accuracy, teaching competence) and adversarial safety (resistance to jailbreaking, handling of academic misconduct requests). The benchmark identifies harmful teaching behaviors (incompetence, indolence, offensiveness) and uses persona-based adversarial prompts to test ethical boundaries specific to educational contexts.

This paper presents an LLM-based instructional module to teach prompting literacy to K-12 students through scenario-based practice with AI chatbots, deployed across 11 secondary education classrooms. The study evaluates an AI auto-grader's capability to assess student prompts, measures changes in students' prompting performance and confidence in using AI for learning, and analyzes the quality of assessment materials.

This paper evaluates LLM-based and statistical NLP methods for automated scoring of nationwide high school graduation essay exams in Estonia, comparing machine-generated scores against human raters across multiple rubric dimensions including content, argumentation, and language quality. The study demonstrates that automated scoring achieves reliability comparable to human raters while also examining bias, prompt injection risks, and providing personalized feedback capabilities.

This mixed-methods study directly evaluates ChatGPT-5 use with 109 secondary (age 16) students in Greek high schools, measuring attitudes (SATAI), anxiety (AIAS), and student responses to deliberately triggered hallucinations across multiple task modalities. The research identifies pedagogical affordances, constraints, and documents how students develop 'epistemic safeguarding' strategies after encountering incorrect AI outputs.

This paper provides a comprehensive review of evaluation practices for LLM-powered Intelligent Tutoring Systems (ITSs), critically analyzing existing benchmarks and proposing three pedagogy-driven research directions for establishing unified, scalable evaluation methodologies grounded in learning science principles. It emphasizes cognitive offloading concerns, citing empirical studies showing students' over-reliance on AI tutors leading to reduced independent problem-solving skills.

This paper directly evaluates AI integration in secondary education through teacher surveys and a sixth-grade classroom experiment, examining impacts on student engagement, critical thinking, retention, and over-reliance. The study specifically investigates cognitive offloading risks when students use AI tools like ChatGPT for homework, measuring whether AI promotes genuine learning or creates dependency.

MinorBench is a hand-built benchmark designed to evaluate how well LLMs refuse unsafe or age-inappropriate queries from children, based on a real-world case study of middle-school students using an LLM chatbot and a novel taxonomy of content-based risks specific to minors.

Safe-Child-LLM introduces a comprehensive benchmark with 200 adversarial prompts designed to systematically evaluate LLM safety across two developmental stages (children 7-12 and adolescents 13-17), testing leading models including ChatGPT, Claude, Gemini, and others for age-appropriate responses and ethical refusal behaviors. The benchmark addresses critical safety gaps in child-facing AI interactions by measuring how well LLMs handle developmental-stage-specific vulnerabilities and harmful content targeting minors.

SproutBench is a comprehensive safety evaluation benchmark comprising 1,283 developmentally-grounded adversarial prompts designed to assess whether LLMs are safe for children and adolescents across three age groups (0-6, 7-12, 13-18 years), measuring 20 distinct child safety dimensions including emotional dependency, privacy violations, and age-appropriate responses. The benchmark evaluates 47 LLMs across cognitive, emotional, and social developmental domains, uncovering significant safety vulnerabilities and trade-offs between interactivity and age appropriateness.

This paper evaluates bias in large language models when they act as personalized tutors, analyzing how 9 LLMs generate and select educational content across different demographic groups (race, ethnicity, sex, gender, disability, income) using over 17,000 educational explanations. The study introduces two bias metrics (MAB and MDB) to measure whether models provide difficulty-appropriate content fairly across student demographics.

This paper presents a survey of 395 students aged 13-25 in France and Italy investigating how they use Large Language Models like ChatGPT in educational settings, examining usage patterns, gender disparities, and potential risks to cognitive skill development including critical thinking and foundational knowledge.

This paper presents an automated multilingual pipeline for generating, solving, and evaluating K-10 math problems using three commercial LLMs (GPT-4o-mini, Gemini 2.5 Flash, Qwen-plus) across English, German, and Arabic. The study reveals consistent linguistic bias, with English solutions rated highest and Arabic lowest, highlighting equity concerns in AI-based educational tools.

This survey paper systematically reviews LLM applications in education from a technology-centric perspective, covering student/teacher assistance, adaptive learning, and commercial tools, while organizing related datasets, benchmarks, and identifying deployment challenges. It provides a comprehensive taxonomy of how LLMs are being applied across multiple educational contexts including tutoring, assessment, content generation, and personalized learning.

This survey paper comprehensively reviews LLM agents for education, proposing a task-centric taxonomy that categorizes agents based on teaching assistance (classroom simulation, feedback generation, curriculum design) and student support (adaptive learning, knowledge tracing, error correction). The paper analyzes architectures, datasets, benchmarks, and deployment challenges including ethics, hallucination, and overreliance.

This paper examines how large language models can be integrated into educational grading practices, analyzing their potential to automate assessment, provide personalized feedback, and improve consistency while addressing challenges like bias, data privacy, and ethical concerns in automated grading systems.

This white paper examines ethical implications of AI-powered tools in educational measurement, focusing on automated scoring, personalized feedback, and data analytics while addressing concerns about validity, reliability, transparency, fairness, algorithmic bias, and equity in educational assessment contexts.

This paper evaluates the robustness of AI-generated content detectors specifically in the context of student essays, constructing the AIG-ASAP dataset with various adversarial perturbation methods to test whether current detectors can identify AI-written student essays that have been modified to evade detection.

This paper develops methods to detect boundaries between human-written and AI-generated content in hybrid essays where students may have used LLMs like ChatGPT to complete portions of their writing assignments. The approach identifies transition points in student essays to help educators detect AI assistance in academic writing.

This paper conducts a systematic literature review of safety, privacy, developmental, and security risks associated with children's interactions with LLMs, and proposes a protection framework with measurable evaluation targets including content safety, age-appropriate readability, bias checks, and prompt-injection robustness. The framework is designed to guide developers, educators, and policymakers in assessing child-safe LLM deployments in educational and companion contexts.

This paper describes the development and deployment of a production-grade guardrail model for a K-12 educational platform, focusing on content safety, appropriateness, and compliance with educational regulations like FERPA and COPPA. The authors detail training methodology, benchmarking against both proprietary education-specific metrics and general safety benchmarks, and production optimization choices.

This paper presents a user study with 16 DHH (d/Deaf and Hard-of-Hearing) learners interacting with LLM-powered AI tutors that embody different personas based on varying levels of experience in DHH education. The study evaluates accessibility preferences, cultural knowledge, transparency needs, and multimodal support requirements (particularly sign language) for DHH learners using conversational AI tutors in online learning contexts.

This paper presents a five-step scaffolding framework for teaching high school students (ages 14-15) to conduct algorithm audits of AI/ML systems, demonstrated through a workshop where youth audited peer-designed TikTok filters. The framework supports systematic critical evaluation of AI systems by having students formulate questions, design tests, collect data, analyze patterns, and communicate findings.

This paper comprehensively evaluates nine Automated Essay Scoring (AES) methods across 25,000+ student essays, examining the interplay between predictive accuracy, fairness (bias across demographic groups), and generalizability in both prompt-specific and cross-prompt settings. The study reveals that prompt-specific models achieve higher accuracy but exhibit greater bias toward students of different economic statuses compared to cross-prompt models.

This perspective article examines how AI integration affects classroom-based formative assessment practices, discussing both opportunities (e.g., automated feedback, data analytics) and challenges (e.g., potential for cognitive offloading, reduced critical thinking, over-reliance on AI-generated content) for teachers implementing formative assessment strategies.

This paper introduces KidLM, a language model specifically designed for children through domain-specific pre-training data collection and a novel Stratified Masking training objective. The work evaluates the model's ability to understand age-appropriate text, maintain safety by avoiding stereotypes, and capture children's unique preferences.

This paper develops a comprehensive taxonomy of content harms specific to children interacting with LLMs, creates diverse child user models based on child psychology literature, and evaluates six state-of-the-art LLMs for safety gaps through red-teaming. The work focuses on identifying safety issues unique to children (ages below 18) that standard adult-focused safety evaluations miss.

This paper surveys students and teachers on their perceptions and experiences using ChatGPT in education, examining both opportunities (personalized feedback, accessibility, interactive learning, lesson preparation) and challenges (cheating risks, reduced critical thinking, academic integrity concerns). The study includes subjective experiments demonstrating ChatGPT's effectiveness in programming education and analyzes survey results to understand its impact on learning and teaching.

This qualitative study investigates programming instructors' perceptions of ChatGPT's potential benefits and drawbacks for computer programming education at a university, exploring how it can support personalized learning, code debugging, and instructional efficiency while raising concerns about over-reliance, accuracy, and ethical issues.

This paper reports on a participatory design workshop with 17 high school students to understand their concerns about generative AI in education and co-design solutions, focusing on issues like bias, misinformation, over-reliance, and academic dishonesty. The work yields design guidelines for educational technology developers based on student perspectives.

This Indonesian case study examines English language teachers' and learners' perceptions of using AI for literary writing instruction, finding that AI helps with lesson planning, content generation, and writing skill development, but also creates dependency that limits creativity and critical thinking. The research uses questionnaire data to assess both pedagogical benefits and concerns about cognitive offloading.

This paper addresses answer bias in knowledge tracing (KT) systems, which monitor students' evolving knowledge states through their interactions with concept-related questions. The authors propose a counterfactual reasoning framework (CORE) that helps KT models better understand actual student knowledge rather than memorizing question-specific answer patterns.

This paper investigates how student gaming behaviors (hint abuse, random guessing, rapid clicking) contaminate data quality in knowledge tracing models used by adaptive learning systems, conceptualizing these behaviors as data poisoning attacks and evaluating their impact on learner modeling performance. The authors simulate diverse gaming patterns to systematically assess vulnerabilities in KT models and explore unsupervised detection approaches.

This paper investigates whether LLM judges can be manipulated through persuasive language when evaluating mathematical solutions, finding that strategically embedded rhetorical techniques cause LLMs to assign inflated scores to incorrect answers by up to 8%. The study tests 14 LLM models across six math benchmarks using seven persuasion strategies grounded in Aristotelian rhetoric.

This paper performs qualitative content analysis of social media discussions to understand early adopter perceptions and experiences with ChatGPT across educational sectors, including K-12 and higher education. The study identifies common use cases, attitudes, and concerns including potential overdependence, superficial learning habits, and erosion of critical thinking skills.

This paper introduces CulturAIEd, an LLM-powered tool that helps K-12 teachers adapt AI literacy curricula to students' cultural contexts through Culturally Relevant Pedagogy (CRP) principles. A pilot study with four K-12 teachers explored how the tool influences their confidence and ability to design culturally responsive AI literacy activities.

This paper presents a framework for building child-centered machine learning pipelines and describes two case studies: one predicting classroom engagement levels using video and biosensor data to support teachers, and another developing a math learning game with handwriting recognition for K-12 students.

This paper benchmarks LLM-generated text detection methods in educational contexts using a novel dataset (GEDE) of student essays and LLM-generated texts across different 'contribution levels' (from fully human-written to fully AI-generated). The work evaluates state-of-the-art detectors' performance in identifying AI-generated student submissions, particularly examining false positive rates and detection across different levels of human-AI collaboration.

This paper investigates counterfactual fairness—a causal approach to algorithmic fairness—in machine learning models applied to educational datasets, examining whether decisions for students would remain the same if their demographic attributes (race, gender) were different. The work evaluates fairness and bias in predictive models used for academic success prediction, at-risk detection, and other educational decision-making contexts.

This paper proposes a federated learning framework to detect mind wandering, behavioral disengagement, and boredom in online learning environments using video-based facial expressions and gaze features while preserving student privacy. The system aims to enable real-time learner support by detecting cognitive and behavioral disengagement without sharing sensitive student data.

This paper evaluates eight publicly-available LLM-generated text detectors (including GPTZero, CopyLeaks, GPTKit, and GLTR) using computing education assignments from university students, comparing their accuracy, false positive rates, and resilience to paraphrasing tools. The study focuses on detecting academic integrity violations when students use ChatGPT to complete assignments, finding significant variation in detector performance and high false positive rates.

This paper examines applications of large language models (LLMs like GPT-4, PaLM) for language teaching and assessment technology, covering content generation, automated grading, grammatical error correction, and personalized feedback. The work focuses on language learning contexts (not K-12 specifically) and evaluates LLM performance on tasks like question difficulty estimation, text generation quality, automated assessment, and error correction using established benchmarks.

This position paper discusses the potential of Artificial General Intelligence (AGI) and large language models like GPT-4 to transform K-12 education through personalized learning, intelligent tutoring, adaptive assessment, and pedagogical support. It explores AGI's capabilities in understanding emotions, providing feedback, and adapting to individual student needs while addressing ethical concerns around bias, fairness, and privacy.

This theoretical paper examines how heavy reliance on AI tools and discovery-based pedagogies may impair memory consolidation and learning, drawing on neuroscience evidence about declarative/procedural memory, cognitive offloading, and the relationship between knowledge and skills.

This paper investigates machine learning and large language model techniques for detecting AI-generated text in high school writing assessments, comparing various generative AI models (ChatGPT, Claude, Gemini) and their susceptibility to paraphrasing tools. The study uses essays from the ASAP Kaggle competition to train classifiers distinguishing student-written from AI-generated content.

Edu-Values is a Chinese education values evaluation benchmark with 1,418 questions testing LLMs on seven core educational values including professional philosophy, teachers' professional ethics, education laws and regulations, cultural literacy, educational knowledge and skills, basic competencies, and subject knowledge. The benchmark evaluates 21 state-of-the-art LLMs using multiple question types including multiple-choice, multi-modal, subjective analysis, adversarial prompts, and Chinese traditional culture questions.

This paper examines the relationship between generative AI and educational equity in K-12 contexts, discussing both opportunities (personalized learning, accessibility, AI tutoring with Socratic dialogue) and challenges (digital divides, bias perpetuation) of implementing AI technologies in education. It provides a conceptual analysis rather than empirical evaluation of AI systems.

This is a literature review discussing the potential of AI chatbots in e-learning, proposing a 'new Socratic method' for chatbot interactions to enhance critical thinking, and examining ethical concerns around sentiment analysis and student data privacy. The paper synthesizes existing research rather than presenting original evaluation methodologies or benchmarks for measuring chatbot effectiveness in K-12 settings.

This paper presents an open-source robotics learning module designed for high school and college students to build AI-powered robot companions, emphasizing hands-on learning about AI, machine learning, robotics, human-AI interaction, and AI ethics through an affordable platform (under $300).