back
AI in schools cannot wait for legislation to catch up — but children cannot be left unprotected while it does. Israel's regulatory framework solves this with four simultaneous layers: usage guidelines, technical security standards, a reverse sandbox for emerging risks, and an exemptive sandbox that removes regulatory barriers to AI-enabled pedagogy. Protection and innovation, in parallel.
The Challenge
Governing AI in education requires addressing three simultaneous demands that conventional regulation sometimes handles poorly. First, AI evolves faster than legislative timelines. Its exponential pace of development is testing the ability of regulatory systems to respond in real time. Second, primary users are children at different emotional, cognitive, and psychosocial developmental stages, and existing safeguarding standards do not yet address the risks AI introduces at each stage. Third, many of the professional and pedagogical mandates governing schools, specifically on teacher roles, curriculum content, and assessment methods, may need revision: the rules most in need of reform are the ones blocking the pedagogy that AI makes possible.
The Solution
Israel’s response is a multi-tier modular regulatory framework: two conventional safety guardrail layers and two innovation governance mechanisms, all advancing simultaneously. This architecture holds protection and enablement in the same structure without sacrificing one for the other, addressing the multidimensional challenge: the speed, the diversity and current mandates. Keeping it flexible enough is what makes this solution stronger: strict prohibition protects students but strangles pedagogical potential, while unmanaged adoption enables experimentation, but leaves students without a safety floor. The modular framework makes that choice unnecessary.
The Framework
-
National Guidelines for responsible GenAI integration in education, published January 2025:
- Age-based access: Grades 1–3 prohibited; Grade 4 Ministry AI-bot only; Grades 5–6 teacher-supervised use; Grade 7+ independent use of approved tools
- All tools must pass central Ministry vetting before any classroom use is permitted
- Baseline AI competency, as defined in the national AI Competency Framework, required before independent access
- Parental consent mandatory from Grade 4
-
The Ministry’s central vetting system sets minimum technical requirements for all approved tools. Requirements address key risks introduced by AI:
- Input/output testing for bias, anomalies, code injections, and inappropriate content; continuous monitoring of system outputs
- Student data may not be used to train AI models; data deletion requirements are established
- Full encryption of all user interactions; users’ right to delete all interaction history
- LLM temperature parameters are calibrated to balance creativity with response safety and predictability
-
The Reverse Sandbox tests emerging risks in a controlled environment and generates recommendations for national regulation: the inverse of a standard sandbox. Six categories are currently active:
- Privacy in personalized learning: mass data collection combined with AI black-box models creates protection challenges existing law does not fully resolve
- Transparency and disclosure: age-appropriate mechanisms to prevent anthropomorphism and over-reliance on AI systems
- Hallucination mitigation: minimum error-reduction standards for child-facing tools, including mechanisms such as LLM-as-a-Judge
- Human-in-the-loop: technical and pedagogical requirements to ensure AI supports rather than replaces educator judgment
- Emotional wellbeing and safeguarding: cognitive, social, and developmental risks arising from AI dependency and over-reliance
- Algorithmic bias and equality: monitoring for disparate outcomes across student populations by background, language, and socioeconomic context
-
Allows temporary exemption from existing professional and pedagogical restrictions that constrain AI-supported learning. Three areas are currently active:
- Teacher role: adapting professional scope and training requirements to enable the transition from content provider to AI-supported pedagogical mentor
- Dynamic content creation: structured exemptions from static, centrally pre-approved textbook requirements, enabling AI-generated personalized learning materials under controlled conditions
- Assessment and evaluation: adapting assessment regulations to evaluate process, higher-order thinking, and AI-collaboration skills alongside academic integrity requirements
What This Framework Makes Possible
The four-layer structure achieves what neither prohibition nor unmanaged adoption can: a school system that is simultaneously protected and capable of generating the evidence it needs to improve its own governance. Safety standards and pedagogical innovation advance in parallel, each informing the other. Evidence from sandbox experiments refines the guardrails; the guardrails define the conditions under which experiments run safely.
