Compliance mapping

Emil enforces the MLCommons / Llama Guard safety taxonomy plus PII, secrets, and prompt-injection detectors. Here’s exactly how that maps to the major content and data regimes — what the screening layer can act on, and what it can’t.

Read this first

This mapping is tooling guidance, not legal advice. A content filter reduces risk against these obligations — it does not by itself make a deployment compliant. Many duties (consent, recordkeeping, risk assessments, contracts) are organizational and sit outside any screening layer. Review presets with qualified counsel for your jurisdiction and use case.

EU AI Act

European Union

Preset: eu-ai-act

Regulation 2024/1689. Bans a set of AI practices outright (Art 5) and imposes transparency duties (Art 50). A screening layer can block content tied to prohibited practices and flag for the disclosure duties; the risk-tier/registration machinery is organizational.

Emil enforces

Flag/block content exhibiting manipulative or deceptive patterns (the only Art 5 practices with a text nexus a filter can screen)

S2 Non-Violent CrimesS10 Hateinjection

flagArt 5(1)(a),(b)

Flag direct human-AI interaction and AI-generated/synthetic content to support the transparency disclosures

flag

flagArt 50(1),(2),(4)

Outside the filter (your responsibility)

  • Most Art 5 prohibitions are deployment/architecture, not text: social scoring (5(1)(c)), facial-image scraping (5(1)(e)), biometric categorisation of protected attributes (5(1)(g)), emotion recognition in work/education (5(1)(f)), real-time remote biometric ID (5(1)(h)) — a content filter cannot determine system intent or these design facts.
  • Risk-tier classification of your system. Note: listing in Annex III does NOT automatically mean high-risk — Art 6(3) provides exceptions (verified: the 'Annex III = auto high-risk' claim is false).
  • Machine-readable watermarking of generative outputs (Art 50(2)) — a marking duty on the generator, not the filter.
  • High-risk obligations: conformity assessment, technical documentation, human oversight, EU database registration.
Primary source

GDPR

European Union / EEA

Preset: gdpr

Regulation 2016/679. Data minimisation (Art 5) and special categories (Art 9) translate cleanly to PII detection + redaction before text leaves your environment — Emil's strongest fit. Note: redaction reduces exposure but is not GDPR 'anonymisation' (Recital 26 demands irreversibility).

Emil enforces

Redact personal data (Art 4) before it is sent to a third-party model, supporting data-minimisation (Art 5(1)(c))

pii

redactArt 4(1), Art 5(1)(c)

Redact/block special categories of personal data (health, race, religion, sex life, political opinion, biometrics)

piiS7 Privacy

redactArt 9(1)

Outside the filter (your responsibility)

  • Lawful basis, consent, and data-subject rights (access/erasure) — record-keeping and process duties.
  • True anonymisation (Recital 26) requires irreversible transformation; redaction-in-transit is a safeguard, not anonymisation.
  • Cross-border transfer mechanisms (SCCs, adequacy) and DPAs with sub-processors.
Primary source

Digital Services Act

European Union

Preset: nsfw-block

Regulation 2022/2065. Platforms must act on illegal content and protect minors. Emil can pre-screen user/AI content against illegal-content categories and gate age-inappropriate material; the notice-and-action and reporting machinery is process.

Emil enforces

Detect and block illegal content categories before publication/display

S1 Violent CrimesS3 Sex-Related CrimesS4 Child Sexual ExploitationS10 Hate

blockDSA Art 16 (notice & action context)

Reduce minors' exposure to age-inappropriate content (e.g. pornography, gambling promotion)

S12 Sexual ContentS3 Sex-Related Crimes

blockDSA Art 28 (protection of minors)

Outside the filter (your responsibility)

  • User-facing notice-and-action reporting mechanism and statement-of-reasons (Art 16, 17).
  • Ban on targeting ads to minors (Art 28(2)) — an ad-system control, not content screening.
  • Trusted-flagger handling, transparency reporting, appeals.
Primary source

UK Online Safety Act 2023

United Kingdom

Preset: nsfw-block

Services must take proportionate measures against priority illegal content and content harmful to children. Emil screens for the illegal-content categories and gates harmful-to-children content; 'highly effective' age assurance is a separate system.

Emil enforces

Detect and minimise priority illegal content (terrorism, CSEA) so users do not encounter it

S1 Violent CrimesS4 Child Sexual Exploitation

blockOSA s.10 (illegal content duties)

Prevent children encountering primary priority content harmful to children (e.g. pornography)

S12 Sexual ContentS11 Suicide & Self-HarmS3 Sex-Related Crimes

blockOSA s.12 (children's safety duties)

Outside the filter (your responsibility)

  • 'Highly effective' age verification / age estimation (s.12) — a distinct identity/age-assurance system.
  • Illegal-content risk assessments and record-keeping.
  • Ofcom transparency reporting and complaints processes.
Primary source

US — FTC Act §5

United States (federal)

There is no omnibus federal AI statute. The FTC enforces against unfair/deceptive AI practices under existing law (§5) — 'there is no AI exemption from the laws on the books'. Emil reduces deception/harm risk but cannot determine what is legally 'deceptive'.

Emil enforces

Flag/block AI outputs that make unsupported guarantees or impersonation/fraud-style claims

S2 Non-Violent CrimesS5 DefamationS6 Specialized Advice

flagFTC Act §5(a) (unfair/deceptive)

Block secrets/credentials leaking through AI flows (security-by-design posture)

secrets

blockFTC §5 (data-security enforcement)

Outside the filter (your responsibility)

  • Whether a specific claim is legally 'deceptive' — a fact-specific legal judgment.
  • Substantiation of advertising claims and overall business practices.
Primary source

CCPA / CPRA

United States — California

Preset: gdpr

CPRA created a protected class of 'sensitive personal information' (SPI). Emil's PII detector targets the SPI identifiers directly — redact before egress. State analogues (Colorado CPA, Virginia VCDPA, etc.) follow similar SPI logic.

Emil enforces

Redact 'sensitive personal information' — SSN, driver's licence/state ID/passport, financial account + access code, precise geolocation, race/religion/union, contents of private communications, genetic/biometric, health, sex life

piiS7 Privacy

redactCal. Civ. Code §1798.140(ae) (SPI)

Outside the filter (your responsibility)

  • Consumer rights handling (opt-out of SPI use, deletion, access) — process duties.
  • 'Notice at collection' and privacy-policy disclosures.
Primary source

COPPA

United States — children under 13

Preset: coppa

Protects personal information of children under 13. Emil can redact COPPA's enumerated identifiers and block age-inappropriate content, but verifiable parental consent is the operator's process duty.

Emil enforces

Redact a child's personal information: name, address, online contact, phone, SSN, persistent identifiers, geolocation, photo/video/audio of the child

pii

redact16 CFR §312.2 (definition of personal information)

Block age-inappropriate content in services directed to under-13 users

S12 Sexual ContentS3 Sex-Related CrimesS1 Violent Crimes

block16 CFR Part 312

Outside the filter (your responsibility)

  • Verifiable parental consent before collection (16 CFR §312.5).
  • Determining a user's actual age / under-13 status.
  • Direct-notice and data-retention-limitation duties.
Primary source

HIPAA

United States — healthcare

Preset: hipaa

Emil can redact the 18 Safe Harbor identifiers before PHI reaches a third-party model. Caveat: Safe Harbor also requires no actual knowledge of residual re-identification risk, and indirect identifiers can still re-identify in combination — redaction is a strong safeguard, not certified de-identification.

Emil enforces

Redact the 18 Safe Harbor identifiers (names, geographic subdivisions, dates, contact info, SSN/MRN/account numbers, biometrics, IP, etc.)

piiS7 Privacy

redact45 CFR §164.514(b)(2) (Safe Harbor)

Block disclosure of PHI tied to a named individual + condition

piiS6 Specialized Advice

block45 CFR §164.502 (uses & disclosures)

Outside the filter (your responsibility)

  • Vendor contracts and agreements required for handling PHI (these are contractual arrangements outside the product).
  • Expert-determination de-identification and residual-risk attestation.
  • Breach notification, access controls, audit infrastructure beyond the screening log.
Primary source

FERPA

United States — education

Preset: education

Emil redacts personally identifiable information from education records before staff send it to an AI tool. Disclosure authorisation and the school-official exception are institutional process.

Emil enforces

Redact PII from education records: student name + grades/discipline/IEP, student ID, parent/guardian contacts tied to a named minor

pii

redact34 CFR §99.3 (PII), §99.30

Outside the filter (your responsibility)

  • Obtaining written consent for disclosure (34 CFR §99.30).
  • Designating the vendor a 'school official' with legitimate educational interest.
Primary source

GLBA

United States — financial

Preset: finance

Emil redacts nonpublic personal information (NPI) — account/routing numbers, balances, card numbers — before it reaches an AI tool, and flags unlicensed-advice patterns. The Safeguards Rule program is organizational.

Emil enforces

Redact nonpublic personal information (account, routing, card numbers, balances)

piisecrets

redact16 CFR Part 313 (Privacy Rule)

Flag content resembling individualized financial advice or guaranteed returns

S6 Specialized Advice

flagGLBA / FTC §5 advice context

Outside the filter (your responsibility)

  • Safeguards Rule security program (16 CFR Part 314).
  • Annual privacy notices and opt-out handling.
Primary source

IRS Circular 230 & Privilege

United States — tax / legal practice

Preset: legal

For tax and legal practitioners, Emil redacts client identifiers and flags privileged content before AI processing. Whether a duty of competence/confidentiality is met is a professional-responsibility judgment.

Emil enforces

Redact taxpayer/client identifiers (SSN, EIN, account numbers) before AI processing

pii

redactCircular 230 §10.51 (competence/confidentiality context)

Flag attorney-client privileged content and client-matter identifiers

S7 Privacy

flagAttorney-client privilege / work product

Outside the filter (your responsibility)

  • The practitioner's duty of competence in using AI and reviewing its output.
  • Engagement-specific confidentiality and conflict determinations.
Primary source

Regulation deep-dives

HIPAAGDPREU AI ActFERPAGLBASOC 2

Need a regime we haven’t mapped, or a counsel-reviewed preset for your jurisdiction? Talk to us.