# Frozen search protocol: KI didactics pilot **Snapshot date / last search:** 2026-08-24 **Scope:** three pilot dossiers for secondary education, with Swiss authority context and explicitly labelled indirect evidence. **Review status:** KI-Didaktik editorial consistency review completed; no named independent or external specialist review. This is not a registered systematic-review protocol and has not been externally peer reviewed. ## Research questions 1. Under which conditions does generative AI improve or impair learning and transfer for secondary-school learners? 2. Which assessment practices preserve valid inferences about student learning when generative AI is available? 3. Which AI-literacy competencies are supported by final authoritative frameworks and empirical school research? ## Search window and eligibility The main publication window was **2022-01-01 through 2026-08-24**. Older foundational work was eligible only when an included recent source explicitly relied on it; such work informed interpretation and citation trails but was not automatically entered into the pilot registry. Included source types were peer-reviewed empirical studies, rigorous systematic/scoping reviews and meta-analyses, final official competency frameworks, and current Swiss legal or education-authority guidance. Eligible empirical learning records had to report an educational outcome rather than only attitudes, adoption, model accuracy, or product performance. Higher-education evidence was retained only as indirect context and is marked with an explicit transfer limit to Sekundarstufe I/II. Excluded were unsourced commentary, product marketing, duplicate or superseded reports, retracted items, draft frameworks when a final version existed, studies without educational outcomes for the learning dossier, and detector benchmarks presented without error analysis. A source was also deferred when the accessible original record was insufficient to verify the metadata or disclosure needed for responsible extraction. ## Boolean cores The following cores were used unchanged at concept level; only database field syntax and quotation handling were adapted: ```text ("generative AI" OR "large language model" OR ChatGPT) AND (learning OR achievement OR retention OR transfer) AND (secondary OR "high school" OR K-12) ("generative AI" OR "large language model" OR ChatGPT) AND (assessment OR "academic integrity" OR authenticity OR grading) AND (secondary OR "high school" OR K-12) ("AI literacy" OR "artificial intelligence literacy") AND (secondary OR "high school" OR K-12 OR teacher*) ``` ## Sources and execution record Search and verification combined discovery searches with original-source checking. The protocol names all channels considered; the table distinguishes direct use from citation-trail use so it does not imply database access that did not occur. | Source | How it was used in this snapshot | | --------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | ERIC | Targeted web-index queries of ERIC records using the Boolean concepts; no complete ERIC export was retained. | | Crossref | DOI metadata/citation resolution was used for known records; no exhaustive Crossref result set was exported. | | OpenAlex | Citation and update leads were checked where surfaced through DOI trails; no exhaustive OpenAlex result set was exported. | | Google Scholar | Used only as non-exhaustive backward/forward citation-trail visibility. No claim of complete result coverage is made. | | OECD | Direct search and verification on official publication pages and PDFs. | | UNESCO | Direct search and verification in UNESCO/UNESDOC official records. | | European Commission/JRC | Direct search of European Commission education pages for the final AILit release and related authority context. No separate JRC empirical record was included. | | Educa | Direct search and verification of Educa/official Swiss report records. | | EDK | Direct authority-site discovery/context checks; no separate EDK record met the pilot inclusion need. | | EDÖB | Direct search and verification on the official authority site. | | Swiss cantonal authorities | Direct targeted search, including the Canton Zurich Volksschulamt living guidance. | | Publisher/DOI landing pages | Final metadata, status, methods, results, limitations, funding and conflict disclosures were checked against the original article or official full text where accessible. | `found_in` in the screening log identifies the actual discovery/verification route for each record rather than presenting a fictional database count. ## Screening and citation-trail procedure 1. Mandatory seeds were resolved at their DOI, publisher, institutional repository, or authority page and screened before inclusion. 2. Included reviews were used for backward citation trails to school-level empirical studies and for method limitations. 3. DOI/publisher citation trails and 2025–2026 update searches were used for forward searching. 4. Contradiction searches combined the three Boolean cores with each of: `harm`, `overreliance`, `offloading`, `no effect`, `null`, `retention`, and `transfer`. 5. Learning evidence was separated into task performance, immediate learning, delayed retention, and transfer. An output-quality advantage while AI remained available was not coded as independent learning. 6. Assessment recommendations were coded as recommendations, not causal findings, unless supported by a comparative educational design. Detector studies were retained only with documented false-positive/error analysis. 7. Framework and legal/policy sources were rated `not-rated` for method confidence because causal study appraisal is inapplicable; their authority, version, finality, scope, and implementation limits were recorded instead. 8. Contradictory or null evidence and transfer boundaries were entered in the claim-source matrix, including when a recommendation is precautionary rather than directly tested. 9. `supporting_sources` and `contradicting_or_null_sources` contain only semicolon-separated stable source IDs. Interpretive caveats belong in `transfer_limit` or `editorial_decision`, which keeps the registry and matrix roles machine-checkable. ## Data extraction and confidence For each included source, extraction covered population, sample or corpus size, education level, geography, subject, intervention/comparison, duration, outcomes, main result, transfer measure, limitations, funding, conflicts, and an explicit confidence rationale. `n` is left as `not applicable` when a policy/framework has no participant sample and as `not reported/verified` when the accessible original does not support a defensible participant count. Absence of an accessible disclosure is recorded as unverified; it is never converted into “none.” Method confidence is an editorial appraisal for this evidence hub, not a formal GRADE score: - **high:** strong causal design with a relevant population/outcome and usable transfer or delayed measure; - **moderate:** useful empirical or synthesis evidence with material heterogeneity, indirectness, attrition, or design limitations; - **low:** descriptive/qualitative evidence, incomplete method visibility, confounded comparisons, or a review with major scope/quality limits; - **not-rated:** final framework, policy, legal or authority guidance where causal grading does not apply. Claim evidence status is separate from source method confidence: - **robust**: an empirical or causal claim is supported by multiple methodologically strong and sufficiently comparable sources, or a directly verifiable final framework-version, legal-duty or official-guidance fact is supported by a single authoritative primary source within its stated scope; - **mixed**: credible evidence points in different directions or effects depend materially on design, context, comparator or outcome; - **limited**: the available evidence is sparse, indirect, methodologically weak or insufficient for a broad inference; - **open**: an important empirical question remains unresolved and the corpus does not support a directional conclusion. Normative authority can make a framework-version, legal-duty or official-guidance claim robust within its jurisdiction, but it does not establish causal learning impact. A precaution recommendation can likewise be strongly supported by documented risk and authority guidance without being mislabelled as an intervention-effect finding; the status applies to the narrow recommendation, while `recommendationBasis: precaution` keeps its basis explicit. ## Limitations of this frozen snapshot - This is an auditable curated editorial evidence pack, not an exhaustive systematic review. No database-wide deduplicated export, independent dual screening, protocol registration, or external risk-of-bias adjudication was completed. - Google Scholar visibility was non-exhaustive and partly dependent on public indexing. Citation counts and “cited by” paths are unstable. - Paywalls and publisher rendering limits prevented full verification of some disclosures or methods; those records are labelled low confidence, deferred, or excluded rather than completed by inference. - Crossref and OpenAlex assisted DOI/citation resolution, but this snapshot does not claim a complete API harvest from either service. - The rapid publication cycle means articles published or corrected after 2026-08-24 are outside this frozen corpus. Retraction/correction status must be rechecked at the next review. - Direct Swiss secondary-school causal evidence and long-term transfer/retention evidence remain sparse. International or higher-education results cannot be assumed to transfer unchanged to Swiss Sekundarstufe I/II. - Authority guidance can establish current obligations and professional expectations, but it does not demonstrate causal learning effects. The files in this directory therefore represent a **frozen 2026-08-24 evidence snapshot**. Living pages may be updated by their owners; the verified date, not an invented publication year, is used for the undated Canton Zurich guidance.