SafeGradeAI
Methodology & Standards · Version 2.4
Monitoring activeMethodology v2.5Last audited May 2026Standards Board sign-off

How SafeGradeAI evaluates AI tools

Every certification is the product of a proprietary, evidence-driven review framework built specifically for schools, districts, and families. We evaluate what matters for children — privacy, moderation, age appropriateness, governance, and educational fit — using a structured internal methodology, and re-verify our findings on an ongoing basis. Certain internal review criteria remain confidential to preserve scoring integrity.

  • COPPA aligned
  • FERPA aligned
  • GDPR-K aligned
  • UK AADC informed
  • Independent review
  • Documentation verified
  • Human evaluated
See the five dimensions Badge standards

The Framework

Five evaluation dimensions, weighted for educational use

SafeGradeAI uses a proprietary multi-factor evaluation framework. Each certified product is reviewed against five public dimensions. Relative weighting reflects what most affects safety, suitability, and institutional risk when AI is used by minors — exact weightings and internal escalation thresholds are intentionally not disclosed to preserve framework integrity.

Privacy & Data Handling

Weighting · Primary

How a tool collects, retains, shares, and uses data generated by minors and educational accounts.

Relative weightingPrimary

What we review

  • Data retention windows and deletion controls
  • Third-party data sharing and sub-processor disclosure
  • Student data usage and model-training opt-outs
  • Account creation requirements (email, age gate, parental consent)
  • Telemetry, analytics, and advertising signals collected
  • Regional handling for COPPA, FERPA, GDPR-K, and UK AADC

Moderation & Safety Controls

Weighting · Primary

How effectively the system prevents harmful outputs and resists adversarial attempts to bypass safeguards.

Relative weightingPrimary

What we review

  • Harmful content prevention across CSAM, self-harm, and violence categories
  • Jailbreak and prompt-injection resistance under repeated probes
  • Behavior on unsafe, leading, or manipulative prompts
  • Layered filtering: model-level, system-level, and post-response review
  • User reporting, escalation, and human-in-the-loop pathways
  • Recovery behavior after a filter breach or unsafe output

Age Appropriateness

Weighting · Substantial

Whether the experience is developmentally suitable for the audience the vendor claims to serve.

Relative weightingSubstantial

What we review

  • Reading level and language complexity versus stated audience
  • Emotional safety: tone, persona behavior, and parasocial risk
  • Developmental suitability of topics, imagery, and interaction patterns
  • Required supervision level and parental visibility
  • Consistency of age gating across web, mobile, and embedded surfaces
  • Behavior when a user self-reports as under the stated minimum age

Transparency & Governance

Weighting · Material

How clearly the vendor publishes, maintains, and stands behind its safety and data commitments.

Relative weightingMaterial

What we review

  • Published privacy policy, DPA, and child-safety statement
  • Vendor responsiveness to safety inquiries within 10 business days
  • Auditability of decisions: logs, model versioning, and change history
  • Explainability of refusals and content decisions to end users
  • Disclosure clarity around AI generation, sources, and limitations
  • Public incident history and remediation track record

Educational Alignment

Weighting · Material

How well the tool fits classroom workflows, district procurement requirements, and policy frameworks.

Relative weightingMaterial

What we review

  • Classroom usability for teachers and students under realistic load
  • Compatibility with common AUPs and district AI policies
  • District readiness: SSO, rostering, admin console, audit logs
  • Teacher controls over student-facing features and prompts
  • Implementation safeguards: pilot guidance, training, change management
  • Alignment with state and national digital learning standards

Scoring System

Four Certification Badges

Overall scores are calculated from the five weighted dimensions and mapped to one of four bands. Bands determine which audiences a tool is appropriate for and what level of oversight we recommend.

KidSafe AI (5–8)

SafeGradeAI Approved

85–100

Strong privacy protections and robust safety safeguards.

Suitable for unsupervised use within the stated age range. Vendor demonstrates mature privacy controls, layered moderation, transparent governance, and education-ready deployment options.

Examples of what lowers a score in this band

  • Ambiguous opt-out flows for model training
  • Partial admin console coverage
  • Inconsistent disclosure between marketing and policy pages
Tween AI Approved (9–12)

Approved With Guidance

75–84

Suitable with parental or institutional supervision.

Acceptable when paired with classroom supervision, a managed account, or active parental oversight. Some controls are present but not uniformly enforced across surfaces.

Examples of what lowers a score in this band

  • Long-context safety drift on extended sessions
  • Limited teacher controls over student-facing features
  • Sub-processor list missing or incomplete
Teen AI Approved (13–17)

Restricted Use

65–74

May require additional oversight or classroom controls.

Use only inside a managed environment with explicit institutional controls. Significant gaps in moderation, transparency, or age handling reduce suitability for unsupervised minors.

Examples of what lowers a score in this band

  • Weak age gating across mobile and embedded surfaces
  • No published incident history
  • Inconsistent refusal behavior under repeated probes
Not Recommended

Not Recommended

Below 65

Significant moderation, privacy, or governance concerns.

Not appropriate for use by minors in our review. Concerns may include reported safety incidents, unbounded persona behavior, opaque data practices, or absent governance commitments.

Examples of what lowers a score in this band

  • Conversations stored and used for training without clear opt-out
  • Documented safety incidents involving minors
  • No reliable separation between adult and minor user experiences

Overall scores are generated through a structured internal review methodology that combines the five public dimensions with proprietary risk, escalation, and reassessment criteria. Exact dimension weights, escalation thresholds, and red-flag triggers are intentionally not disclosed to preserve framework integrity and reduce manipulation risk.

Tier assignments refresh whenever monitoring evidence, vendor policy changes, or reassessment review warrant a change.

Review Process

Seven stages, from intake to ongoing monitoring

Every certification follows the same procedural framework. Reviewers, timelines, and evidence requirements are defined in advance so the outcome is reproducible from the case file.

The vendor submits a structured intake covering product surface area, intended audience, data flows, and existing safety documentation. Intake is scoped before any review hours are spent.

Reviewer actions

  • Reviewer assigns a case ID and conflict-of-interest screen
  • Vendor receives an intake checklist within one business day
  • Scope agreement signed (surfaces, regions, languages)

Evidence & criteria

  • Product surface map (web, mobile, embedded, API)
  • Stated audience and minimum age
  • Regions where the product is offered

Ongoing Monitoring

A certification is a commitment, not a snapshot

AI products change quickly. SafeGradeAI treats every certification as a living record with continuous monitoring, scheduled re-evaluation, and the authority to act between cycles.

Monitoring Active

Live

Certified products are tracked continuously. Privacy policies, DPAs, and child-safety statements are diffed automatically.

Annual Re-evaluation

12-month cycle

Every certified product is fully re-reviewed against the current methodology version on a 12-month cycle.

Policy Updated

10-day SLA

Material policy changes are surfaced on the verification page within 10 business days, with reviewer notes attached.

Status Can Change

Always revocable

Certifications can be paused, downgraded, or revoked between cycles when incidents, policy changes, or test results warrant it.

Use the framework

Trusted by schools, families, and procurement teams

Browse certified tools, learn what each badge means, or submit a product for independent evaluation under the current methodology.

Explore certified tools Badge standardsSubmit a tool

Methodology metadata

Current version
2.4 · published Mar 2026
Next scheduled revision
Sep 2026 (semi-annual)
Public change log
Maintained per certification
Standards committee
Independent review panel