Who this is for: Compliance officers, AI product teams, legal counsel, and engineering teams operating generative AI systems with 1M+ monthly users that are publicly accessible to Utah consumers.
Effective 2027. Utah HB 276 ("Artificial Intelligence Modifications") was signed by Governor Spencer Cox on March 24, 2026, with most provisions taking effect in 2027. The law requires covered providers to embed provenance data in AI-generated content, enable users to determine whether content was AI-generated or materially altered, and prohibits non-consensual deepfake intimate images. Enforcement is through the Utah Attorney General's office.
Contents
1. Overview 2. Key Obligations 3. Obligation-to-Procedure Mapping 4. Detailed Procedure Cards 5. How This Connects to Other Provenance Laws 6. Quick Reference 7. Quick Start 8. References1. Overview
Utah HB 276 -- the Digital Content Provenance Standards Act / Digital Voyeurism Prevention Act -- addresses two converging risks in generative AI: the inability to distinguish AI-generated content from human-created content, and the weaponization of AI to produce non-consensual intimate imagery. The law targets covered providers -- operators of generative AI systems with 1 million or more monthly users that are publicly accessible in Utah.
The evidence gap is straightforward. When an auditor or the Utah AG asks "How do you prove provenance data is embedded in every piece of AI-generated content?", most organizations point to a product feature or a policy document. Neither is evidence. A watermark that exists but cannot be independently verified is not provenance -- it is a claim. SWT3 witness procedures close this gap by producing cryptographically anchored records that prove provenance embedding occurred, disclosure mechanisms are operational, and consent controls are enforced.
Key characteristics:
- Scope: Generative AI systems with 1M+ monthly users, publicly accessible in Utah -- regardless of where the provider is headquartered
- Provenance requirement: Embed provenance data in AI-generated content (watermarking or embedded metadata satisfies this)
- Disclosure requirement: Allow users to determine if content was AI-generated or materially altered by AI
- Deepfake prohibition: Bans creation and distribution of non-consensual deepfake intimate images ("counterfeit intimate images")
- Enforcement: Utah Attorney General's office
- Alignment: Designed to align with California SB 942 and Washington HB 1170 provenance frameworks
2. Key Obligations
| Obligation | Requirement | Timeline |
|---|---|---|
| Provenance embedding | Embed provenance data (watermark or metadata) in AI-generated content that persists through reasonable transformations | At content generation |
| Disclosure | Provide mechanisms allowing users to determine whether content was AI-generated or materially altered by AI | At point of access |
| Integrity | Provenance data must be tamper-resistant and survive common content transformations (cropping, compression, re-encoding) | Throughout content lifecycle |
| Deepfake consent | Prohibit generation of non-consensual intimate images depicting identifiable individuals without their explicit consent | Pre-generation gate |
| Training data | Maintain records of data sources used to train generative models, sufficient to respond to AG inquiries | Ongoing |
| Audit trail | Maintain records sufficient to demonstrate compliance with provenance and disclosure obligations upon AG request | Ongoing |
3. Obligation-to-Procedure Mapping
| HB 276 Obligation | SWT3 Procedure | What It Witnesses | Evidence Produced |
|---|---|---|---|
| Provenance embedding | AI-WATERMARK.1 | Watermark or metadata embedding at content generation time | Anchor with watermark type, embedding method, content hash, generation timestamp |
| Disclosure | AI-TRANS.1 | Transparency disclosure mechanism availability and activation | Anchor with disclosure type (AI-generated content), delivery method, recipient context |
| Integrity | AI-MARK.1 | Content marking integrity verification across transformations | Anchor with marking method, integrity check result, transformation type survived |
| Deepfake consent | AI-CONSENT.1 | Consent verification gate before generating content depicting identifiable individuals | Anchor with consent status, subject identifier (hashed), verification method |
| Training data | AI-DATA.1 | Training data provenance and source documentation | Anchor with data source identifier, collection method, licensing status |
| Inference provenance | AI-INF.1 | Model inference event with input/output provenance | Anchor with model identifier, input hash, output hash, latency, clearing level |
| Audit trail | AI-AUDIT.1 | Compliance audit trail integrity and completeness | Anchor with audit scope, record count, integrity hash, retention period |
4. Detailed Procedure Cards
Content Provenance Watermarking
HB 276 requires: Covered providers must embed provenance data in AI-generated content. This can be satisfied through digital watermarking, embedded metadata (such as C2PA manifests), or equivalent techniques that allow downstream verification of AI origin.
How SWT3 addresses it: The witnessWatermark() call mints an anchor each time provenance data is embedded in generated content. The anchor records the watermark type (steganographic, metadata, C2PA), the embedding method, and a hash of the content at the point of embedding. This creates a verifiable chain from the generation event to the provenance mark.
Query AI-WATERMARK.1 anchors and correlate with AI-INF.1 anchors for the same session. Every inference that produces user-facing content should have a corresponding watermark anchor. A content hash mismatch between the INF and WATERMARK anchors indicates the watermark was applied to different content than what was generated -- a compliance gap.
AI-Generated Content Disclosure
HB 276 requires: Users must be able to determine whether content was AI-generated or materially altered by AI. The disclosure mechanism must be accessible and understandable to a reasonable user.
How SWT3 addresses it: The witnessTransparency() call records each disclosure event -- when a user is informed that content is AI-generated, the delivery method (label, metadata viewer, API response header), and the content identifier. This proves the disclosure mechanism is not just designed but actively delivered.
AI-TRANS.1 anchors should exist for every content delivery path. Factor A (disclosure type) should indicate "AI-generated content" or "materially altered." Cross-reference with AI-INF.1 to verify every inference that reaches end users has a corresponding disclosure event. Gaps indicate content paths that bypass the disclosure mechanism.
Content Marking Integrity
HB 276 requires: Provenance data must persist through reasonable content transformations -- cropping, compression, format conversion, re-encoding. A watermark that disappears when a user screenshots the content does not satisfy the requirement.
How SWT3 addresses it: The witnessContentMarking() call records integrity verification events -- periodic checks that provenance marks survive transformations. The anchor captures the marking method, the transformation type tested, and the verification result (detected or lost).
AI-MARK.1 anchors demonstrate ongoing integrity testing, not just initial embedding. Show anchors across multiple transformation types (compression, crop, re-encode). A pattern of PASS results across transformations proves robustness. Any FAIL results should have corresponding remediation records or design change documentation.
Training Data Provenance
HB 276 requires: Covered providers must maintain records of training data sources sufficient to respond to Attorney General inquiries. This includes the ability to describe data origins, collection methods, and licensing arrangements.
How SWT3 addresses it: The witnessDataProvenance() call records training data source documentation events -- when data lineage records are created, updated, or verified. The anchor captures the data source identifier, collection method, and licensing status, creating an auditable trail of data governance.
AI-DATA.1 anchors prove training data documentation is maintained, not just claimed. Factor A (data source identifier) should cover all major training data sources. Look for regular anchor creation corresponding to data pipeline updates. Gaps in AI-DATA.1 anchors after known training data changes indicate documentation lag.
Deepfake Consent Verification
HB 276 requires: Non-consensual generation of intimate images depicting identifiable individuals is prohibited. This requires a pre-generation gate that prevents the creation of "counterfeit intimate images" without explicit consent from the depicted individual.
How SWT3 addresses it: The witnessConsent() call records consent verification events at the pre-generation gate. The anchor captures the consent status (granted, denied, not applicable), a hashed subject identifier (never the raw identity), and the verification method. Denied consent triggers a generation block, and the anchor proves the block occurred.
AI-CONSENT.1 anchors with "denied" status prove the pre-generation gate is actively blocking prohibited content. Anchors with "not applicable" prove the gate evaluated the request and determined consent was not required (non-intimate content). The absence of AI-CONSENT.1 anchors for content generation requests involving identifiable individuals is a compliance gap.
Inference Provenance
HB 276 requires: Provenance must be traceable back to the generation event. The inference itself -- which model produced which content, when, and in what context -- is the foundation of the provenance chain.
How SWT3 addresses it: The witnessInference() call (or the wrap() proxy) mints an anchor for every model inference. The anchor records the model identifier, input hash, output hash, latency, and clearing level. This is the base layer that all other provenance procedures reference.
AI-INF.1 anchors are the foundation. Every AI-WATERMARK.1, AI-TRANS.1, and AI-MARK.1 anchor should trace back to a corresponding AI-INF.1 anchor via shared session or content identifiers. Orphaned provenance anchors (watermark without inference) or orphaned inferences (inference without watermark) both indicate gaps in the provenance chain.
Compliance Audit Trail
HB 276 requires: Covered providers must maintain records sufficient to demonstrate compliance with provenance and disclosure obligations. The Utah AG can request these records during an investigation.
How SWT3 addresses it: The witnessAuditIntegrity() call records audit trail completeness and integrity verification events. The anchor captures the audit scope (which obligations were checked), the record count, an integrity hash of the audit dataset, and the retention period. This proves the audit trail itself is monitored for completeness.
AI-AUDIT.1 anchors prove the audit system is self-monitoring. Regular anchors (daily or weekly) demonstrate continuous compliance monitoring, not just point-in-time snapshots. The integrity hash allows independent verification that audit records have not been tampered with. Cross-reference record counts with expected volumes from AI-INF.1 anchor counts.
5. How This Connects to Other Provenance Laws
Utah HB 276 is part of a growing wave of state-level AI provenance legislation. The same SWT3 witness procedures that satisfy HB 276 produce evidence applicable to parallel obligations in other jurisdictions.
| Law | Jurisdiction | Shared Obligations | Key Differences |
|---|---|---|---|
| California SB 942 | California | Provenance embedding, AI-generated content disclosure, watermarking | SB 942 includes manifest-based provenance (C2PA alignment) and covers a broader range of content types. Stronger technical specification for provenance standards. |
| Washington HB 1170 | Washington | Content provenance, disclosure to users, deepfake restrictions | HB 1170 includes election-related deepfake provisions and broader synthetic media definitions. Overlapping watermark and disclosure requirements. |
| EU AI Act Art. 50 | European Union | AI-generated content marking, transparency obligations, machine-readable provenance | Art. 50 requires machine-readable marking (not just human-readable disclosure). Applies to all providers placing AI systems on the EU market, regardless of size. Broader scope than HB 276's 1M+ user threshold. |
The operational advantage of procedure-based evidence is that a single AI-WATERMARK.1 anchor proves provenance embedding for Utah, California, Washington, and the EU simultaneously. The obligation differs by jurisdiction; the evidence does not.
6. Quick Reference
| Examiner Question | Where to Look |
|---|---|
| Do you embed provenance data in AI-generated content? | AI-WATERMARK.1 anchors. Each anchor records watermark type, embedding method, and content hash. Cross-reference with AI-INF.1 to verify coverage across all generation events. |
| Can users determine if content is AI-generated? | AI-TRANS.1 anchors with disclosure type "AI-generated content." Verify anchors exist for every content delivery path, not just the primary UI. |
| Does provenance survive content transformations? | AI-MARK.1 anchors showing integrity checks across transformation types (compression, crop, re-encode). PASS results prove robustness. FAIL results should have remediation documentation. |
| How do you prevent non-consensual deepfakes? | AI-CONSENT.1 anchors at the pre-generation gate. "Denied" status anchors prove active blocking. Absence of consent anchors for requests involving identifiable individuals is a gap. |
| What training data records do you maintain? | AI-DATA.1 anchors documenting data sources, collection methods, and licensing. Regular anchor creation proves ongoing documentation, not a one-time snapshot. |
| Can you produce compliance records on AG request? | AI-AUDIT.1 anchors with integrity hashes and record counts. Self-monitoring audit trail proves records are complete, tamper-evident, and exportable on demand. |
7. Quick Start
pip install swt3-ai
# Initialize with the AI provenance profile
swt3 init --profile nist-ai-rmf --tenant YOUR_TENANT
# Run the demo to see provenance and watermark anchors
python -m swt3_ai.demo
# Or use TypeScript
npm install @tenova/swt3-ai
npx swt3-init --profile nist-ai-rmf
Full SDK documentation: sovereign.tenova.io/docs
Create a free account: sovereign.tenova.io/signup
8. References
- Utah HB 276 -- Utah Legislature, signed March 24, 2026, effective 2027
- Utah Attorney General's Office -- Enforcement authority
- California SB 942 -- California AI Transparency Act (content provenance)
- Washington HB 1170 -- Washington synthetic media and AI provenance law
- EU AI Act Crosswalk -- Art. 50 transparency obligations mapped to SWT3 procedures
- SWT3 UCT Registry -- 271 procedures across 9 namespaces
- SWT3 Bidirectional Framework Crosswalks -- machine-readable JSON, 77 frameworks
- SDK Documentation -- Python, TypeScript, and 5 additional language SDKs