Quellenbibliothek

Publikation 2026; im Register zuletzt geprüft am 25.08.2026. Finds a modest positive overall effect of AI-assisted feedback in L2 writing, with stronger effects on emotional engagement than on cognitive gains and a clear need for pedagogical integration. Methodische Einordnung: moderat; Large synthesis size supports cautious feedback claims, while domain specificity and heterogeneity limit transfer to general AI-didactic recommendations. Wichtigste Grenze: The evidence concerns L2 writing and cannot be transferred automatically to all subjects or summative assessment contexts.

Publikation 2026; im Register zuletzt geprüft am 24.08.2026. Reports a moderately positive pooled effect on learning outcomes, with very high heterogeneity and only a minority of included studies from primary or secondary education. Methodische Einordnung: moderat; The transparent meta-analysis and 4,193-participant corpus are informative, but extreme heterogeneity and limited school representation constrain transfer to Sek I/II. Wichtigste Grenze: The pooled estimate is highly heterogeneous (I² 91.444%), so an average effect does not establish a stable effect across contexts.

Publikation 2026; im Register zuletzt geprüft am 24.08.2026. Under the conservative missing-as-violation coding, only 9.52% and 6.52% of comparisons met all three controls; under benefit-of-doubt coding, 68.25% and 63.04% did. The large range reflects pervasive missing methodological information and does not establish that every incompletely reported comparison was confounded. Methodische Einordnung: moderat; The audit identifies severe comparability and reporting uncertainty, but its sensitivity range shows that missing information—not proven confounding in every comparison—drives much of the conclusion; it does not estimate a corrected pooled effect. Wichtigste Grenze: This is an audit of prior study comparisons rather than a new intervention effect estimate.

Publikation 2026; im Register zuletzt geprüft am 24.08.2026. Among 14-15-year-olds, note-taking alone and note-taking combined with an LLM produced better comprehension and three-day retention than LLM use alone. Methodische Einordnung: hoch; The preregistered randomized secondary-school design includes a delayed retention test and directly compares cognitively different study strategies. Wichtigste Grenze: The intervention used two reading passages over a short period and analysed 344 of 405 recruited students.

Publikation 2026; im Register zuletzt geprüft am 24.08.2026. Defines 19 primary- and secondary-level competencies across engaging with, creating with, managing and shaping AI, integrating knowledge, skills and attitudes. Methodische Einordnung: nicht bewertet; Final OECD-European Commission authority and transparent scope support framework claims; causal-method grading does not apply. Wichtigste Grenze: The framework defines intended competencies and examples but does not estimate causal learning impact.

Publikation 2026; im Register zuletzt geprüft am 24.08.2026. Utility-value prompting produced a more favorable perceived-utility trajectory than cognitive-strategy prompting, but neither structured GenAI condition significantly improved effort, domain knowledge or elaboration over standard ChatGPT use in grades 7-9. Methodische Einordnung: moderat; Preregistration and randomization make the null comparison valuable, while attrition, short duration and an active-AI comparator limit broader conclusions. Wichtigste Grenze: Approximately 35% dropout reduces precision and may affect generalizability.

Publikation 2026; im Register zuletzt geprüft am 24.08.2026. Synthesises emerging evidence that general-purpose GenAI can improve task performance without independent learning, while pedagogically designed uses can support learning. Methodische Einordnung: nicht bewertet; This authoritative synthesis is useful for scope and policy interpretation, but its method is not rated as a causal primary study. Wichtigste Grenze: The report synthesises heterogeneous studies and does not provide its own causal effect estimate.

Publikation 2026; im Register zuletzt geprüft am 24.08.2026. Maps reported K-12 risks to wellbeing, intellectual agency and educational ecology and identifies process assessment, scaffolding and critical AI literacy as mitigation directions. Methodische Einordnung: moderat; The review is directly K-12 and empirically bounded, but scoping methods and heterogeneous risk definitions preclude causal or prevalence claims. Wichtigste Grenze: A scoping review maps reported risks but does not estimate causal risk magnitude or prevalence.

Publikation 2026; im Register zuletzt geprüft am 24.08.2026. Finds that teacher monitoring, judgment, intervention and orchestration condition how AI information becomes classroom action across 29 K-12 studies. Methodische Einordnung: moderat; Direct K-12 scope and quality appraisal support the teacher-role synthesis, while limited GenAI-specific evidence prevents stronger causal claims. Wichtigste Grenze: Only six included studies concerned generative AI or chatbots; the remainder involved broader educational AI.

Publikation 2026; im Register zuletzt geprüft am 24.08.2026. Reports a large positive average effect of K-12 AI-education interventions on AI-literacy measures, with substantial heterogeneity across a small evidence base. Methodische Einordnung: moderat; The K-12 meta-analysis supports intervention promise, but a small heterogeneous corpus and varied measures leave developmental sequencing open. Wichtigste Grenze: Only 16 source articles contributed 57 effects, limiting moderator power and robustness.

Publikation 2025; im Register zuletzt geprüft am 25.08.2026. Reports positive average effects of ChatGPT-supported learning on behavioural, cognitive and emotional engagement while also identifying risks of over-reliance and disengagement. Methodische Einordnung: moderat; The meta-analysis offers useful engagement evidence, but small corpus size and varied outcomes prevent treating engagement as a general learning guarantee. Wichtigste Grenze: The corpus contains only 17 empirical studies and combines varied disciplines, ages and implementation roles.

Publikation 2025; im Register zuletzt geprüft am 25.08.2026. Shows that GenAI learning effects depend on the alignment of technology, pedagogy and subject matter, and highlights pedagogical scaffolding as a recurring condition for productive student use. Methodische Einordnung: moderat; The systematic review is directly relevant to instructional design, but heterogeneity and higher-education dominance constrain strong school-level prescriptions. Wichtigste Grenze: The reviewed evidence is very recent and uneven across school levels, with higher education strongly represented.

Publikation 2025; im Register zuletzt geprüft am 24.08.2026. Organises GenAI literacy into knowing, applying, evaluating, ethics and attitudes, while showing that most included evidence came from higher education rather than schools. Methodische Einordnung: moderat; The review supports multidimensional GenAI-literacy scope, but higher-education dominance and an early evidence window constrain school recommendations. Wichtigste Grenze: Across 53 educational-level assignments in 51 articles, the counts were 39 higher education, 10 secondary and 4 primary; two articles were assigned to more than one level, so these are not exclusive article shares.

Publikation 2025; im Register zuletzt geprüft am 24.08.2026. Finds generally positive but comparison-sensitive effects for K-12 intelligent tutoring systems, most of which predate contemporary generative AI. Methodische Einordnung: moderat; The school-level systematic review is directly relevant, but its evidence mainly concerns earlier ITS and heterogeneous short quasi-experiments. Wichtigste Grenze: Most included systems were classical ITS rather than general-purpose generative AI.

Publikation 2025; im Register zuletzt geprüft am 24.08.2026. A carefully prompted and sequenced AI tutor produced larger immediate post-test gains than in-class active learning in two Harvard physics lessons. Methodische Einordnung: moderat; The randomized crossover design supports an immediate causal estimate, but higher-education indirectness and absence of retention or far transfer limit school recommendations. Wichtigste Grenze: The population was higher education, not Sekundarstufe I/II; school transfer is indirect.

Publikation 2025; im Register zuletzt geprüft am 24.08.2026. Sets out seven development lines and 51 development approaches for coherent, secure and ethically appropriate educational data use across the Swiss education space. Methodische Einordnung: nicht bewertet; The report is authoritative Swiss strategy context, not causal evidence or a substitute for applicable federal and cantonal law. Wichtigste Grenze: The report proposes development approaches and measures; it is not enacted law or a causal impact study.

Publikation 2025; im Register zuletzt geprüft am 24.08.2026. Synthesises 69 articles and 62 meta-analysed studies and reports positive average performance effects, alongside substantial heterogeneity and weak separation of assisted performance from independent transfer. Methodische Einordnung: moderat; Systematic methods support an average-effect estimate, but high heterogeneity and tool-available outcome testing materially limit the learning claim. Wichtigste Grenze: The performance estimate is highly heterogeneous (I² 91.789%).

Publikation 2025; im Register zuletzt geprüft am 24.08.2026. A six-week, teacher-led after-school programme using Microsoft Copilot in nine public secondary schools improved an endline learning composite and regular term examination performance. Methodische Einordnung: moderat; Random assignment and a regular-term outcome are valuable, but attrition and a bundled after-school treatment prevent attribution to GenAI alone. Wichtigste Grenze: Only 759 assigned students were observed at endline, so attrition and voluntary attendance affect interpretation.

Publikation 2025; im Register zuletzt geprüft am 24.08.2026. Across roughly 50 classes and nearly 1,000 grades 9-11 students, unrestricted GPT improved supported practice but reduced later no-tool exam performance; a tutor with safeguards increased practice performance and mitigated the harm. Methodische Einordnung: hoch; A preregistered randomized school study with a later no-tool assessment directly distinguishes assisted performance from independent learning. Wichtigste Grenze: The intervention covered four 90-minute mathematics sessions in one Turkish school context.

Publikation 2025; im Register zuletzt geprüft am 24.08.2026. Maps applications, opportunities and gaps in K-12 generative-AI research and is retained as a low-confidence gap map rather than causal evidence. Methodische Einordnung: niedrig; The publisher verifies a broad K-12 review, but method inconsistencies and absence of study-quality appraisal restrict it to gap mapping. Wichtigste Grenze: The reported study counts are not fully consistent across the article.

Publikation 2025; im Register zuletzt geprüft am 24.08.2026. Explains that the Federal Act on Data Protection applies directly to AI-supported processing within its scope and emphasizes transparency, safeguards, objection rights and human review of automated individual decisions. Methodische Einordnung: nicht bewertet; The source is rated for current Swiss legal authority and scope rather than causal research method. Wichtigste Grenze: This general legal guidance is not specific to school assessment practice and is not individualized legal advice.

Publikation 2025; im Register zuletzt geprüft am 24.08.2026. Synthesises assessment-integrity challenges and recommends process evidence, oral defence, staged work, reflection, disclosure and human oversight, primarily for higher education. Methodische Einordnung: niedrig; The review offers useful design options, but higher-education indirectness and limited search and appraisal methods require cautious school transfer. Wichtigste Grenze: The review is predominantly higher education; recommendations transfer only indirectly to Sekundarstufe I/II.

Publikation 2025; im Register zuletzt geprüft am 24.08.2026. Five students aged 14-17 described organizational and cognitive-load benefits alongside weaker personal voice, reflective depth, overreliance concerns and AI guilt. Methodische Einordnung: niedrig; Rich process evidence is directly relevant to authenticity, but the five-student qualitative design cannot support broad effectiveness claims. Wichtigste Grenze: The very small purposive case cannot estimate prevalence or causal effects.

Publikation 2024; im Register zuletzt geprüft am 24.08.2026. Specifies 12 student competencies across human-centred mindset, ethics, AI techniques and applications, and AI system design, with Understand-Apply-Create progression levels. Methodische Einordnung: nicht bewertet; The official final UNESCO framework supports competency-scope claims but does not provide causal or validated-progression evidence. Wichtigste Grenze: The framework is normative and curricular; it is not an empirical impact evaluation.

Publikation 2024; im Register zuletzt geprüft am 24.08.2026. Defines 15 teacher competencies across a human-centred mindset, ethics, AI foundations and applications, AI pedagogy, and AI for professional learning. Methodische Einordnung: nicht bewertet; The final UNESCO authority supports teacher-competency scope; causal and psychometric claims are outside the framework's evidence design. Wichtigste Grenze: The framework guides professional development and policy but does not test effects on classroom learning.

Publikation 2024; im Register zuletzt geprüft am 24.08.2026. Senior-secondary students improved on course-aligned AI problem-solving, empowerment and ethical-understanding measures after a project-based AI-literacy course. Methodische Einordnung: niedrig; Direct senior-secondary feasibility evidence is valuable, but the uncontrolled single-group design cannot establish causal effectiveness. Wichtigste Grenze: Without a comparison group, maturation, selection and testing effects cannot be separated from course impact.

Publikation 2023; im Register zuletzt geprüft am 24.08.2026. Seven detectors produced substantially more false positives on 91 TOEFL essays than on 88 US eighth-grade essays, demonstrating consequential error disparity. Methodische Einordnung: moderat; The transparent error analysis demonstrates that detector outputs are not reliable sole evidence, while dataset confounding limits attribution of the disparity. Wichtigste Grenze: Language proficiency, dataset origin and essay genre are confounded, so the analysis does not isolate a single causal mechanism.

Publikation 2023; im Register zuletzt geprüft am 24.08.2026. Across 756 tool-text tests, all 14 detectors remained below 80% accuracy and were vulnerable to translation and paraphrasing, supporting caution against sole-evidence use. Methodische Einordnung: moderat; Multi-tool error testing supports the narrow reliability warning, but the small constructed corpus and rapid tool change limit numeric generalization. Wichtigste Grenze: The 54-text benchmark is not a representative sample of all school submissions.

undatierte Behördenquelle; im Register zuletzt geprüft am 25.08.2026. Explains that Private actors and Bundesorgane fall under the DSG and EDÖB remit, while data processing by cantonal and municipal authorities falls under cantonal or municipal data-protection oversight. Methodische Einordnung: nicht bewertet; The source is appraised for institutional authority, jurisdiction and currency; a causal-method rating is not applicable. Wichtigste Grenze: The page describes institutional competence at a general level; exact allocation for a school, provider or processing operation still depends on the actors' roles and the applicable law.

undatierte Behördenquelle; im Register zuletzt geprüft am 25.08.2026. Explains for Zurich Volksschulen that the Informations- und Datenschutzgesetz (IDG) and its Verordnung (IDV) apply and points school staff and authorities to cantonal data-protection guidance and advice. Methodische Einordnung: nicht bewertet; The source is appraised for authority, jurisdiction and currency; a causal-method rating is not applicable. Wichtigste Grenze: The guidance is cantonal, not national, and must not be transferred to another jurisdiction or school sector without checking the applicable law and competent authority; it is not individualized legal advice.

undatierte Behördenquelle; im Register zuletzt geprüft am 24.08.2026. The Zurich Volksschulamt calls on the public Volksschule to use transparent rules, process-aware and varied assessment, teacher professional judgment, verification of AI-supported feedback, data protection and caution with detector outputs. Methodische Einordnung: nicht bewertet; The source is appraised for authority, jurisdiction and currency; a causal-method rating is not applicable. Wichtigste Grenze: The page has no stated publication date and is designed to change with the discussion; 2026 in the reserved ID records verification context, not publication year.