The provenance, misuse-monitoring and evidence-preservation controls AI systems need so child-safety is built into the model — not bolted on. Score a vendor against it.
A standard, not a filter. Conventional 'red-team the model and inspect the outputs' breaks down here, because legitimate researchers can't lawfully possess the material. These requirements move child-safety into the engineering itself — provenance, misuse monitoring and lawful evidence handling. Score each one for a vendor.
Compliance
0%
Set a status on each requirement to see the compliance picture.
Requirements
Training data
Documented data provenance
Maintain a manifest of training-data sources; no ingestion of known exploitative sources.
Known-CSAM screening
Screen training data against established hash lists (e.g. NCMEC/industry) before training; log results.
No scraping of abuse repositories
Exclude sources known to host or index abuse material.
Model behavior & evaluation
Refuse sexualized-minor requests
The model refuses to produce sexual content involving minors under any framing.
No synthetic-CSAM generation
Generation of minor sexual content is blocked at output, not only discouraged.
Evaluation without possessing illegal material
Red-team via lawful proxies and classifiers so researchers never hold illegal content — the central constraint the research flags.
Deployment & monitoring
Misuse monitoring
Detect patterns of attempted misuse (grooming, CSAM solicitation) at the account level.
Account linkage
Link repeated attempts across accounts/devices so a banned actor can't simply re-register.
Reporting pathway to NCMEC
A defined, tested pipeline to report apparent CSAM to NCMEC and cooperate with law enforcement.
Evidence & governance
Lawful evidence preservation
Preserve the minimum lawful evidence for reporting without redistributing illegal material.
Independent audit
Outside verification of these controls — not self-attestation.
Published policy & appeals
A public child-safety policy and a route to contest wrong enforcement.
Hand-off
Governance requirements only. This standard deliberately contains no operational detail about abuse material; it defines controls and reporting duties. Not legal advice. Apparent CSAM is reported to the NCMEC CyberTipline 1-800-843-5678.
CHILD-SAFETY AI STANDARD · a DAYBREAK response app · Born Between 2 Generals, LLC for America's Future, Inc. · working title · child-safety content is never paywalled
