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 incorporates 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 education, and introduces a comprehensive evaluation framework consisting of seven diverse benchmarks (quantitative, qualitative, automatic, and human evaluations) grounded in learning science principles to assess pedagogical capabilities of AI tutoring systems. The work includes real-world deployment at Arizona State University's Study Hall and demonstrates that LearnLM-Tutor is consistently preferred by educators and learners over prompt-tuned Gemini across multiple pedagogical dimensions.

MinorBench is a hand-built, open-source benchmark that evaluates LLMs' ability to refuse unsafe or age-inappropriate content requests from children, using a taxonomy of content-based risks specific to minors derived from real middle-school chatbot deployment. The benchmark tests six prominent LLMs under different system prompts to assess child-safety compliance.

This paper describes the design and classroom evaluation of an LLM-based instructional module that teaches K-12 students prompting literacy through scenario-based practice with AI chatbots, including an AI auto-grader for evaluating student-written prompts. The study deployed the module across 11 secondary education classrooms in two iterations, measuring students' prompting performance improvements, confidence changes, and the effectiveness of different assessment question types.

This mixed-methods study examines 109 Greek secondary students' (age 16) attitudes, anxiety, and responses to hallucinated outputs when using ChatGPT-5 in classroom settings across multiple tasks. The research measures cognitive and behavioral attitudes toward AI, AI-related anxiety, and documents students' 'epistemic safeguarding' strategy of restricting AI use to domains where they can verify outputs after encountering hallucinations.

EduGuardBench is a dual-component benchmark that evaluates LLMs as simulated teachers by measuring both professional fidelity (using Role-playing Fidelity Score to detect pedagogical harms like incompetence, indolence, and offensiveness) and adversarial safety (using persona-based jailbreak prompts to test vulnerability to harmful requests including academic misconduct). The benchmark tests 14 leading models and is publicly available with automated evaluation code.

Safe-Child-LLM introduces a comprehensive benchmark with 200 adversarial prompts and standardized ethical refusal scales to systematically evaluate LLM safety across two developmental stages: children (7-12) and adolescents (13-17). The paper evaluates leading LLMs including ChatGPT, Claude, Gemini, and others, revealing critical safety deficiencies in child-facing scenarios, with both datasets and evaluation code publicly released.

This paper evaluates bias in LLMs acting as personalized tutors by measuring how models generate and select educational content differently for students with varying demographic characteristics (race, gender, disability, income, etc.). The study introduces two bias metrics (MAB and MDB) and applies them to 9 LLMs using over 17,000 educational explanations across multiple difficulty levels and subjects.

This paper presents survey results from 395 students aged 13-25 in France and Italy investigating how they use LLMs like ChatGPT in educational settings, finding widespread adoption across age groups and disciplines but revealing concerning patterns including gender disparities in usage and lack of critical evaluation among younger users. The study examines usage frequency, purposes, proofreading habits, and potential risks to cognitive skill development.

This paper develops a comprehensive taxonomy of content harms specific to children interacting with LLMs and creates Child User Models based on child psychology literature to evaluate the safety of six state-of-the-art LLMs through red-teaming. The evaluation reveals significant safety gaps in LLMs for child-specific harm categories that are not captured by standard adult-focused safety evaluations.

This paper presents an automated multilingual pipeline that generates, solves, and evaluates 628 math problems aligned with the German K-10 curriculum across English, German, and Arabic using three commercial LLMs (GPT-4o-mini, Gemini 2.5 Flash, Qwen-plus), finding consistent linguistic bias with English solutions rated highest and Arabic lowest. The pipeline includes automated generation, translation, solving, and LLM-judge evaluation to measure quality disparities in educational AI outputs across languages.

This paper reports on a participatory design workshop with 17 high school students to understand their perspectives on GenAI in education, identifying concerns about bias, misinformation, plagiarism, over-reliance, and false accusations of academic dishonesty, and proposing design guidelines for EdTech developers. Students co-designed GenAI tools and school policies addressing these concerns through structured activities.

This paper explores K-12 teachers' AI literacy levels and how they integrate Project-Based Learning AI toolkits (AI Art Lab, AI Music Studio, AI Chatbot) into diverse subject areas through interviews and co-design sessions, examining pedagogical adaptations, challenges, and ethical concerns.

This is a comprehensive survey paper that systematically reviews the technological applications of large language models (LLMs) in K-12 education across multiple dimensions including student assistance, teacher support, adaptive learning, and commercial tools, while also identifying datasets, benchmarks, risks, and future research opportunities.

This paper introduces GEDE (Generative Essay Detection in Education), a benchmark dataset with over 900 student essays and 12,500 LLM-generated essays across various contribution levels (from human-written to fully AI-generated), and evaluates five state-of-the-art detection methods. The benchmark assesses detectors' ability to identify different levels of student vs. LLM contribution in educational writing assignments.

This paper proposes a federated learning framework for privacy-preserving detection of mind wandering, behavioral disengagement, and boredom in online learning environments using facial expressions and gaze features from webcam video. The approach is validated across five datasets and benchmarks multiple federated learning algorithms for automated learner state detection.

This paper proposes a boundary detection method to identify transition points between human-written and AI-generated content in hybrid essays, constructing a dataset by having ChatGPT fill in incomplete student essays and evaluating detection approaches. The work addresses academic integrity concerns by detecting AI-assisted writing in educational assignments rather than assuming essays are entirely human or AI-generated.

This paper conducts a systematic literature review of safety risks when children interact with LLMs, identifying concerns around harmful content, bias, developmental inappropriateness, and adversarial attacks, and proposes a protection framework with measurable evaluation targets for child-safe LLM deployment.

Edu-Values is a Chinese education values benchmark with 1,418 questions evaluating LLMs across seven core educational dimensions including professional philosophy, teachers' ethics, education laws, cultural literacy, educational knowledge/skills, basic competencies, and subject knowledge. The benchmark evaluates 21 LLMs using human feedback-based automatic evaluation and finds Chinese LLMs outperform English ones, with particular weaknesses in professional ethics and philosophy.

This paper describes the development and deployment of a domain-specific guardrail model for a K-12 educational platform that ensures content appropriateness, safety, and policy compliance. The authors benchmark their guardrail model against proprietary education-related benchmarks and public safety benchmarks, demonstrating superior performance in filtering inappropriate content for K-12 contexts.

This qualitative study examines primary school teachers' views on using artificial intelligence in education through interviews with 16 teachers, exploring perceived advantages (personalized learning, rapid feedback, time-saving) and disadvantages (student laziness, reduced social interaction, ethical concerns) of AI integration in K-12 classrooms.

This paper presents AI for Wellbeing, a middle school curriculum where students learn about conversational AI and ethics by building chatbot prototypes for wellbeing support. The study evaluates curriculum effectiveness through a 5-day virtual workshop with 23 middle school students and 5 teachers, measuring knowledge gains and student perspectives on AI.

This theoretical paper examines the integration of generative AI (especially LLMs like ChatGPT) in K-12 education, discussing pedagogical principles, potential benefits and risks, and instructional implications for teachers across different educational stages and subjects. It emphasizes the importance of using AI to supplement rather than replace cognitive effort, promoting critical thinking and deep learning while addressing concerns about cognitive offloading.

This paper presents a five-step framework for teaching high school students to systematically audit machine learning systems through hands-on activities, demonstrated via a case study where teens audited peer-designed TikTok filters to evaluate their limitations and biases. The study focuses on scaffolding critical evaluation skills and algorithmic accountability rather than evaluating AI tutoring systems or education-specific AI tools.

This paper presents a multi-agent LLM system with five role-based agents (Evaluator, Equity Monitor, Metacognitive Coach, Aggregator, and Reflexion Reviewer) that automatically scores learner reflections using rubrics and generates formative feedback comments while checking for bias and promoting metacognition. The system was evaluated in a 12-session AI literacy program with adult learners, achieving expert-level agreement in scoring and producing feedback rated as helpful and empathetic.

This systematic literature review synthesizes 14 empirical studies examining how ChatGPT has been utilized in educational settings by both students and teachers, analyzing its applications, benefits, and drawbacks in learning contexts. The review identifies uses including virtual tutoring, writing assistance, personalized learning support, and teacher productivity tools, while noting concerns about overreliance affecting innovation and collaborative learning.

This paper introduces KidLM, a language model specifically designed for children through a novel data collection pipeline and training objective (Stratified Masking). The model is evaluated on its ability to understand lower grade-level text, avoid stereotypes, and capture children's unique preferences using both automated metrics and human evaluation.

This exploratory self-study examines how ChatGPT answers science education questions, explores potential pedagogical applications for science educators, and reflects on its use as a research tool, identifying both capabilities (generating units, rubrics, quizzes) and risks (presenting information as authoritative without evidence, potential bias).

This paper conducts a SWOT analysis of ChatGPT's role in learning research, examining its strengths (natural language processing, accessibility), weaknesses (lack of deep understanding, objectivity issues), opportunities (intelligent research assistant, writing support), and threats (risks to critical thinking, information security, educational equity). The study uses an ethnographic approach treating ChatGPT as a research participant to assess its impact on learning research practices.

This qualitative study examines programming instructors' perceptions of ChatGPT's potential benefits and drawbacks for computer programming education through interviews with 12 university faculty members, identifying advantages like personalized learning and code debugging alongside concerns about accuracy, over-reliance, and ethical issues.

This chapter develops a conceptual framework for bidirectional human-AI alignment in K-12 education, emphasizing mutual adaptation between AI systems and educational stakeholders (teachers, students, institutions) to ensure trustworthy learning environments grounded in equity, transparency, and student agency. It proposes strategies for aligning AI values, goals, and interaction norms with educational principles rather than presenting empirical evaluations or benchmarks.

This paper presents an explainable AI (XAI) educational tool designed for K-12 students (grades 4-9) to teach them about data-driven mechanisms in social media platforms, including data collection, user profiling, engagement metrics, and recommendation algorithms. The tool uses an Instagram-like interface with real-time visualizations and was tested with 209 children in 12 two-hour sessions, using learning analytics to track how students navigated and understood these AI-driven processes.

This paper provides a critical conceptual review of ethical challenges in integrating AI into K-12 STEM education, proposing a three-phased implementation roadmap and professional development model that emphasizes equity-centered design, bias audits, and AI literacy for both teachers and students.

This paper presents a human-subject research study evaluating a computational action curriculum that teaches 101 K-12 students (ages 9-18) to evaluate and create socially responsible AI through structured reflection and an impact matrix tool. The study measured students' perceptions of AI ethics and their ability to analyze positive and negative impacts of AI technologies using pre-post questionnaires and open-ended responses.

This qualitative study analyzes social media discourse to explore early adopters' perceptions and experiences with ChatGPT across educational sectors (K-12, higher education, skills training), examining usage patterns, attitudes, and concerns about cognitive offloading, critical thinking erosion, and ethical implications.

This paper presents CulturAIEd, an LLM-powered tool designed to help K-12 teachers adapt AI literacy curricula to students' cultural contexts using Culturally Relevant Pedagogy (CRP) principles. Through a pilot study with four teachers, the research evaluates how the tool influences teachers' confidence and ability to design culturally responsive AI literacy activities.

This paper proposes a framework for adapting machine learning pipelines to be child-centered and presents two case studies: predicting classroom engagement levels from video/biometric data to support teachers, and developing a handwriting recognition system for young learners with special educational needs.

This paper evaluates counterfactual fairness (a causal individual-level fairness notion) of machine learning models on educational datasets, examining how sensitive attributes like race and gender causally influence predictions of student outcomes. The study demonstrates counterfactual fairness analysis on benchmark educational datasets to assess whether models produce the same decisions regardless of demographic group membership.

This white paper examines ethical implications of AI-powered tools in educational measurement, discussing opportunities and challenges such as automated scoring, algorithmic bias, transparency, and fairness in assessment contexts. It provides a broad overview of ethical concerns and proposed guidelines rather than presenting a specific benchmark or empirical evaluation of AI systems.

This position paper articulates the potential of Artificial General Intelligence (AGI) to transform K-12 education through personalized learning, intelligent tutoring systems, adaptive pedagogy, and enhanced assessment capabilities, while discussing ethical concerns around bias, privacy, and fairness. The paper is conceptual and does not present a benchmark, evaluation suite, or empirical study.

This paper investigates methods for detecting AI-generated text in K-12 student writing assessments using classical machine learning and large language models, comparing outputs from ChatGPT, Claude, and Gemini, and examining the effectiveness of paraphrasing tools like GPT-Humanizer and QuillBot in evading detection.

This qualitative case study examines Swedish primary school students' (ages 11-12) cognitive and affective perceptions of AI and their current usage patterns through pre-tests, focus group interviews, and post-lesson evaluations. The study explores students' understanding of AI concepts, emotional responses to AI, and their concerns about rapid AI development, job loss, and privacy.

This paper explores how d/Deaf and Hard-of-Hearing (DHH) learners interact with LLM-powered AI tutors that have different personas representing varying experiences in DHH education, focusing on accessibility preferences and cultural knowledge through a user study with 16 DHH participants. The study examines interaction patterns, transparency needs, and multimodal support requirements (especially sign language) for DHH learners using AI tutoring systems.

This paper examines the relationship between generative AI and educational equity, discussing both opportunities (accessibility, personalized learning, AI tutoring with Socratic dialogue) and challenges (digital divides, bias perpetuation) for K-12 education without presenting original empirical evaluation or a benchmark.

This paper surveys 163 Finnish K-12 upper secondary students about their self-perceived AI competence and concerns regarding AI risks across systemic, institutional, and personal domains, using co-occurrence network analysis to examine relationships between competence levels and risk perceptions. The study finds that lower-competence students emphasize personal risks (reduced creativity, critical thinking) while higher-competence students focus on systemic risks (bias, inaccuracy).

This qualitative study interviews 20 higher education teachers about challenges they face in assessing student work when AI text generators like ChatGPT are available, finding concerns about academic integrity, difficulty detecting AI-generated assignments, and fears about students' loss of creativity and writing skills.

This paper develops a theoretical framework based on Constructive Alignment theory and Bloom's taxonomy to guide how educational assessment should be adapted in response to students' use of AI tools like LLMs, proposing that different Bloom levels require different assessment approaches when AI is available.

FairMonitor is a four-stage automated framework for detecting stereotypes and biases in LLM-generated content through open-ended questions, with a case study implementation (Edu-FairMonitor) covering educational scenarios across nine sensitive factors including gender, race, age, learning ability, and socioeconomic status. The framework uses direct inquiry, story-based testing, implicit association testing, and unknown situation testing to evaluate bias in five LLMs.

SproutBench is a safety evaluation benchmark for LLMs targeting youth, comprising 1,283 adversarial prompts designed to assess age-appropriate responses across early childhood (0-6), middle childhood (7-12), and adolescence (13-18). It evaluates 47 LLMs on dimensions including safety, risk prevention, interactivity, and age appropriateness, focusing on ethical and developmental considerations rather than learning outcomes.

This mixed-methods study identifies 14 essential student competencies for engaging with generative AI in education and six lecturer strategies for integrating GenAI into teaching, based on literature review and survey data from 130 university students in South Asia and Europe. The research focuses on higher education contexts, examining skills like AI literacy, critical thinking, prompt engineering, and ethical AI practices.