Latin America's first binding AI law. Law No. 31814 and Supreme Decree 115-2025-PCM establish risk-based classification, prohibited uses, and audit requirements mapped to SWT3 witness procedures.
Who this is for: Peruvian AI developers and deployers, LATAM technology companies operating across the Andean Community, firms subject to OECD AI Principles, compliance teams preparing for Peru's phased implementation timeline, and organizations managing multi-jurisdiction compliance across Latin America.
Latin America's first binding AI law published. Risk-based regime with prohibited uses in force. Peru published the Regulation of Law No. 31814 via Supreme Decree No. 115-2025-PCM on September 9, 2025. This is the first enforceable AI law in Latin America. The regulation classifies AI use into prohibited, high-risk, and acceptable risk categories. Prohibited uses include manipulative AI, autonomous lethal systems, mass surveillance without legal basis, biometric profiling for social scoring, and predictive policing. Governance falls under the PCM's Secretariat of Government and Digital Transformation (SGTD). Organizations deploying AI in Peru face specific transparency, privacy, cybersecurity, and audit obligations with phased implementation timelines.
Peru is the first Latin American country with a binding AI law. The SGTD (Secretariat of Government and Digital Transformation) under the PCM (Presidency of the Council of Ministers) leads governance and oversight of AI deployment across both public and private sectors.
The law aligns with OECD AI Principles, which Peru adopted alongside Argentina, Brazil, Chile, Colombia, Costa Rica, and Mexico. This alignment means organizations already compliant with OECD frameworks have a head start, but Peru's regulation adds enforceable specifics that go beyond voluntary principles.
Phased implementation means some obligations activate immediately upon publication while others follow a defined timeline set by SGTD. Organizations should begin evidence collection now, even for obligations not yet in force, to avoid last-minute compliance gaps.
| Instrument | Authority | Scope | Status |
|---|---|---|---|
| Law No. 31814 | Congress of Peru | AI development and use framework | Enacted |
| Supreme Decree 115-2025-PCM | PCM / SGTD | Regulation implementing Law 31814 | Published Sep 2025 |
| Peru Data Protection Law (29733) | ANPD | Personal data processing | Enacted |
| OECD AI Principles | OECD | Intergovernmental AI standards | Adopted by Peru |
Supreme Decree 115-2025-PCM establishes three tiers of AI risk classification. Each tier carries distinct compliance obligations and maps to different SWT3 procedure sets.
These AI uses are banned outright: manipulative AI, autonomous lethal systems, mass surveillance without legal basis, biometric profiling for social scoring, and predictive policing. Organizations must demonstrate that guardrails actively prevent these uses. SWT3 procedure AI-GRD.1 provides cryptographic evidence that prohibited uses are blocked.
Subject to strict audits, registration, transparency obligations, and mandatory human oversight. High-risk AI systems require full procedure coverage including AI-AUDIT.1, AI-HITL.1, AI-TRANS.1, AI-EXPL.1, AI-LOG.1, and AI-DPIA.1.
Basic transparency duties apply. Organizations must maintain minimum evidence through AI-TRANS.1 and AI-LOG.1 to demonstrate responsible deployment.
| Peru Obligation | SWT3 Procedure | Evidence Produced |
|---|---|---|
| Risk classification documentation | AI-RISK.1 | Risk register anchor with classification tier |
| Prohibited use guardrails | AI-GRD.1 | Guardrail activation anchor proving prohibited uses blocked |
| Transparency obligations | AI-TRANS.1 | Transparency disclosure anchor |
| Human oversight for high-risk | AI-HITL.1 | Human review outcome anchor |
| Privacy impact assessment | AI-DPIA.1 | DPIA anchor with impact findings |
| Audit trail maintenance | AI-AUDIT.1 | Audit integrity anchor |
| Cybersecurity measures | AI-SEC.1 | Security scan anchor |
| Explainability for affected persons | AI-EXPL.1 | Explanation anchor |
| Record-keeping | AI-LOG.1 | Log completeness attestation |
| Data protection compliance | AI-DATA.1 | Data provenance anchor |
| Bias and fairness assessment | AI-FAIR.1 | Fairness assessment anchor |
| Incident reporting | AI-IR.1 | Incident response anchor |
Witnesses that content filters, safety guardrails, and policy enforcement mechanisms are active and functioning. Each guardrail trigger mints an immutable anchor recording the guardrail type, trigger condition, and action taken.
For prohibited use compliance, verify that AI-GRD.1 anchors exist for each of the five prohibited categories in Supreme Decree 115-2025-PCM. The anchor should confirm the guardrail was present, triggered during testing, and what blocking action it performed.
Anchors the risk classification decision for each AI system, including the tier assignment (prohibited, high-risk, or acceptable) and the rationale supporting that classification.
Peru's three-tier classification mirrors the EU AI Act structure. Confirm each AI system has a classification anchor dated before deployment. SGTD may request classification evidence during registration.
Witnesses that a human reviewer evaluated AI output before consequential action. Records reviewer identity, decision outcome, review latency, and override status.
High-risk AI systems under Law 31814 require human oversight. Verify that AI-HITL.1 anchors show consistent human review for high-risk decisions, with reviewer binding that ties the review to a specific individual.
Anchors the completion and findings of a privacy impact assessment. Records scope, identified risks, mitigation measures, and residual risk acceptance.
Peru's Data Protection Law (29733) applies alongside Law 31814. Verify that AI-DPIA.1 anchors reference both instruments and that the DPIA covers personal data processed by the AI system.
Witnesses that audit logs are complete, tamper-evident, and available for inspection. Each anchor attests to log coverage period, record count, and integrity hash.
Supreme Decree 115-2025-PCM requires audit trail maintenance for high-risk systems. Verify AI-AUDIT.1 anchors cover the full operational period with no gaps in log coverage.
Anchors that an AI decision was accompanied by a meaningful explanation accessible to the affected person. Records explanation type, target audience, and delivery method.
Law 31814 grants affected persons the right to explanation. Verify that AI-EXPL.1 anchors demonstrate explanations were generated and delivered in Spanish, with language and accessibility appropriate for the intended audience.
Peru's five prohibited AI uses require organizations to demonstrate that guardrails are in place, tested, and enforced. The AI-GRD.1 procedure provides cryptographic proof for each category:
Each guardrail activation mints an immutable SWT3 Witness Anchor proving the guardrail was present, triggered, and what action it took. This evidence chain is available for SGTD audit at any time.
| Question | Answer |
|---|---|
| Governing law | Law No. 31814, Supreme Decree 115-2025-PCM |
| Regulator | SGTD under PCM |
| Risk tiers | Prohibited, High-risk, Acceptable risk |
| Prohibited uses | 5 categories (manipulative, lethal, surveillance, social scoring, predictive policing) |
| SWT3 procedures mapped | 12 (GRD, RISK, TRANS, HITL, DPIA, AUDIT, SEC, EXPL, LOG, DATA, FAIR, IR) |
| Jurisdiction code | PE (ISO 3166-1) |
| Data protection law | Law 29733 (ANPD) |
| International alignment | OECD AI Principles |
Install the SDK and begin collecting Peru-specific compliance evidence:
pip install swt3-ai
from swt3_ai import Witness
witness = Witness(
api_key="your-api-key",
jurisdiction="PE",
legal_basis="Law 31814"
)
# Risk classification anchor
witness.witness(
procedure="AI-RISK.1",
factor_a="credit-scoring-v3",
factor_b="high-risk",
factor_c="Affects credit access",
)
# Guardrail enforcement for prohibited uses
witness.witness_guardrail(
guardrail_type="manipulative_ai_block",
action="blocked",
model_id="content-filter-v2",
)
# Human oversight for high-risk decisions
witness.witness_human_review(
review_outcome="approved",
reviewer_id_hash="analyst-42",
items_reviewed=1,
latency_ms=3200,
)
# Flush all anchors
witness.flush()