Does COPPA require a per-request AI audit log?

Part 312 addresses notice and consent, parent access and minimization, plus security and retention. It leaves the exact log design open. A per-request record is an operational way to prove what happened on a managed AI route. It can show the supplied identity context and classified data; destination and policy version; and decision. Counsel should map that evidence to the operator's facts.

What belongs in the audit population?

Include every in-scope routed AI attempt during the review period. That means permitted and redacted requests. It also means denied and failed requests, plus retried and timed-out requests. Reconcile that set against application and provider records. Embedded vendor inference and local models need separate populations. So do direct consumer tools, because an external HTTP gateway never receives those events.

Should the evidence package retain full prompts?

The Rule requires data minimization and purpose-bound retention. Store full content only when the approved purpose and retention analysis support it. A fingerprint and classification result may support testing with less duplication. A controlled content reference or redacted excerpt may do the same. The assessor should verify that retrieval remains possible for the evidence actually relied upon.

How should an auditor test deletion?

Use a synthetic record linked across the consent service and application; request evidence and vendor workflow; backups and exception register. Trigger the approved deletion process and document completion. Identify each retained artifact with its purpose and expiry. The retention provision requires deletion when information is no longer reasonably necessary for the collection purpose.

Can a vendor report replace operator evidence?

A vendor report supports the service-provider file. The operator still needs evidence connecting that provider to the specific child-facing service and disclosed purpose. It must also connect the provider to the routed destination and retention policy. The package should document observed control operation separately. AI vendor due diligence covers procurement evidence; the COPPA package adds live operating proof.

COPPA AI audit evidence for sampling, retrieval, and integrity

COPPA AI audit evidence starts with a reconstructable data flow. A child types a question into a tutor, the application adds profile context, and an HTTP request carries that payload to an LLM. The assessor needs records tying audience and consent decisions to notices and destinations; safeguards and retention; and parental-rights responses. Policy PDFs leave the live request unanswered. I would begin with the request population because a polished policy can hide an unsampled route.

TL;DR

Define the complete population of child-facing AI requests before selecting a sample. Include blocked and failed events.
Retrieve consent and notice through stable identifiers. Do the same for parent-request and vendor records. Apply that method to retention and security-testing records, then retrieve the request records.
Test integrity with policy versions and timestamps. Verify record integrity through signatures or hashes. Review custody records and reconcile them against independent systems.
Keep legal conclusions with privacy counsel. Gateway evidence covers only authenticated HTTP traffic deliberately routed to an LLM.

Start with scope and a complete population

The current COPPA Rule in 16 CFR Part 312 applies to covered operators of child-directed services and operators with the specified actual knowledge. The audit file should begin with the operator analysis and the child-directed or mixed-audience decision. Identify the products in scope and collection points. Add the third parties and review period.

For AI, build a population of routed requests using gateway events and application transaction IDs, then reconcile it with model account records and deployment inventories. Capture all permits and redaction decisions. Include denials and policy errors such as timeouts or retries. Totals should agree across independent sources, with every gap explained before sampling. Document the query window and clock source. State the deduplication rule and treatment of asynchronous callbacks so the reviewer can reproduce the denominator. The bright red denial line in a test console matters as much as a successful tutor response because it proves the prohibited route entered the population. Student data privacy in EdTech AI provides the broader vendor and school context.

Sampling must preserve meaningful risk strata

A random sample alone can miss the event classes that carry COPPA risk. Stratify the population by audience lane and data type; consent state and model destination; disclosure category and decision result; and policy version. Select routine requests alongside first collection, changed practices, parent revocation, deletion, a denied third-party route, and a security-control failure.

The sampling memo should identify the population query, selection method, reproducible criteria, exclusions, sample IDs, and execution date. The memo must name its preparer and reviewer. Preserve the query text and a result fingerprint so an assessor can rerun it without copying child content into the memo. For each selection, the reviewer should be able to trace the sampled ID back to its stratum and explain why the event entered the test. Stable event references and controlled retrieval reduce the personal information duplicated for audit. The AI governance audit framework supplies a broader control-sampling structure.

Retrieval proves that rights and records work

The Rule's parental-rights provision requires a reasonable means for a parent to review a child's personal information and refuse further use or collection. It also permits the parent to direct deletion. An audit should test the workflow rather than rely on a screenshot of the portal. Use a synthetic child record and authenticate the parent through the approved process. Request the record and revoke further collection. Initiate deletion, then preserve the timestamps and outcomes. The test should follow linked copies into downstream systems and identify exceptions plus failed jobs or manual handoffs that prevent completion.

Retrieval also applies to the assessor's package. A request ID should lead to the applicable notice and consent record; policy version and data classification; destination and vendor assurance; and retention disposition. Record failed lookups and orphaned events. The FTC's COPPA compliance guide describes parent access and deletion, notice and consent, plus security and retention. It is staff guidance, while Part 312 remains the operative rule.

Integrity separates evidence from a convenient log

The Rule's security provision requires a written information security program, risk assessment, safeguards, regular monitoring and testing, plus program evaluation. Evidence should identify the coordinator, dated assessment, safeguard design, test results, findings and remediation, including approved program changes. The retention provision adds the written policy and deletion process.

For request records, document the timestamp source and event identifier; policy version and destination; decision and integrity mechanism. A signature or hash chain can support tamper evidence, as can an append-only control or independently monitored store. COPPA prescribes duties rather than a specific logging design, so the architecture choice is an operational evidence pattern. Test it by altering a copied record and restricting deletion permissions, then trace custody through export. My preference is blunt: if the audited application can quietly rewrite its own history, the export is an activity log rather than dependable evidence.

Assemble the assessor package by claim

Build the package around the claims an assessor must test:

Coverage: Include the scope memo and product inventory. Add the audience analysis and data-flow diagram. Include the complete event population.
Notice and consent: Include effective notice versions and direct notices. Add consent records and the material-change workflow. Document third-party disclosure choices separately.
Parent rights: Include authenticated access and revocation records. Include the deletion test records.
Minimization and retention: Include the field inventory and necessity decision. Add the written retention policy and disposal job evidence. Document every approved retention exception.
Security: Include the program owner and risk assessment. Add vendor assurances and safeguard tests. Include monitoring results and remediation.

Index each artifact to the tested claim and control owner; system of record and review period; and sample ID. That index lets an assessor locate supporting records without browsing an oversized export or receiving unrelated child information.

The FTC's 2025 final rule amendments explain the amended definitions plus the notice and security provisions, along with the retention and third-party provisions. Keep the rule text beside the implementation interpretation so an auditor can distinguish obligation from chosen control.

DeepInspect

DeepInspect can supply a bounded portion of this evidence when an authenticated application deliberately routes HTTP AI traffic through its stateless proxy. The application must provide the originating user or agent identity, age or audience context and purpose, plus stable references. DeepInspect can evaluate those fields through prompt classification and destination checks under a versioned policy, then permit or deny the request while redacting content when policy requires it.

That record supports population construction and sampling, plus retrieval and integrity review for the routed request. Privacy and legal teams retain responsibility for consent validity and parent verification; notices and retention decisions; and legal conclusions. Application owners must supply evidence for their databases and embedded vendor AI. Local execution and direct unrouted calls require separate controls. Book a demo today.