Connecticut SB5 is a 39-section omnibus AI law covering AI companions, synthetic content marking, automated employment decision tools, frontier model whistleblower protections, and AI-related layoff notices -- all mapped to SWT3 witness procedures with phased enforcement starting October 2026.
Who this is for: AI product teams deploying companion or chatbot applications, synthetic content generators, employers using automated hiring tools, frontier model developers, subscription-based AI service providers, compliance officers, and legal counsel advising on Connecticut AI obligations.
Phased enforcement begins October 1, 2026. Connecticut SB5 ("An Act Concerning Online Safety") was signed by Governor Ned Lamont on May 29, 2026. This 39-section omnibus law is one of the most comprehensive state-level AI regulations in the United States, covering AI companions, synthetic content, employment tools, frontier models, and consumer disclosures across three enforcement phases. Penalties vary by section, and some provisions carry a private right of action. Organizations should begin compliance preparation now -- the first wave of requirements takes effect in fewer than 90 days.
| Field | Detail |
|---|---|
| Law | Connecticut SB5 -- "An Act Concerning Online Safety" |
| Signed | May 29, 2026 by Governor Ned Lamont |
| Sections | 39 sections covering 6 major areas |
| Effective (Phase 1) | October 1, 2026 -- synthetic content marking, whistleblower protections, AI layoff notices |
| Effective (Phase 2) | January 1, 2027 -- AI companion requirements |
| Effective (Phase 3) | October 1, 2027 -- automated employment decision tools (AEDTs) |
| Scope | Developers, deployers, operators, frontier model developers, subscription AI providers |
| Penalties | Vary by section: civil penalties, CUTPA enforcement, private right of action (some sections) |
| Enforcer | Connecticut Attorney General; Connecticut Department of Labor (layoff notices) |
| SWT3 Procedures | 12 procedures mapped across 10 namespaces |
SB5 applies broadly to any organization that develops, deploys, or operates AI systems affecting Connecticut residents. The law defines multiple categories of regulated entities:
The law applies to AI systems that affect Connecticut residents regardless of where the developer or deployer is headquartered. If your AI system interacts with, makes decisions about, or generates content consumed by Connecticut residents, SB5 likely applies to you.
SB5 takes effect in three phases, giving organizations time to prepare for different categories of requirements. Each phase activates a distinct set of obligations:
Phase 1 -- October 1, 2026
Synthetic content provenance marking (all AI-generated content must carry machine-readable and human-perceptible markers). Frontier model whistleblower protections (employees and contractors who report safety concerns cannot face retaliation). AI-related layoff notice requirements (employers must include AI displacement information in WARN Act notices). Subscription-based AI provider disclosure requirements.
Phase 2 -- January 1, 2027
AI companion requirements take full effect. Developers and operators must implement safety protocols for minor users, disclose the AI nature of companion systems, obtain informed consent for data collection, and establish guardrails preventing harmful behaviors including suicide encouragement, self-harm promotion, sexual content for minors, and substance abuse facilitation.
Phase 3 -- October 1, 2027
Automated employment decision tool (AEDT) provisions activate. Employers must conduct bias audits before deploying AEDTs, provide impact assessments, notify candidates and employees when AEDTs are used, and offer alternative evaluation pathways. Annual bias audit reporting requirements begin.
Each SB5 obligation maps to one or more SWT3 witness procedures. Anchors generated by these procedures create the cryptographic evidence trail that demonstrates compliance across all three enforcement phases.
| CT SB5 Obligation | SWT3 Procedure | What It Witnesses | Phase |
|---|---|---|---|
| AI companion disclosure + subscription provider disclosure | AI-TRANS.1 |
Transparency disclosure delivered to users | 1 & 2 |
| Synthetic content marking | AI-MARK.1 |
Content provenance marking applied | 1 |
| Synthetic content mark verification | AI-WATERMARK.1 |
Provenance marks survive transformations | 1 |
| Companion suicide/self-harm intervention | AI-HITL.1 |
Human override capability in crisis scenarios | 2 |
| Companion safety protocols for minors | AI-SAFE.1 |
Safety boundary enforcement for minor users | 2 |
| AEDT anti-discrimination | AI-FAIR.1 |
Bias detection and mitigation in employment tools | 3 |
| AEDT impact assessment | AI-DPIA.1 |
Data protection impact assessment for employment tools | 3 |
| Companion data practices consent | AI-CONSENT.1 |
User consent for data collection and processing | 2 |
| AI nature disclosure | AI-ID.1 |
Agent identity -- system identifies as AI | 1 & 2 |
| Minor safety guardrails | AI-GRD.1 |
Guardrail enforcement preventing harmful outputs | 2 |
| Frontier model whistleblower process | AI-INCIDENT.1 |
Incident reporting and whistleblower intake | 1 |
| Frontier model compute thresholds | AI-COST.1 |
Resource consumption witnessing for threshold classification | 1 |
SB5 requires: AI companion developers and operators must clearly disclose the AI nature of companion systems before and during user interaction. Subscription-based AI providers must disclose AI capabilities, limitations, data practices, and the fact that users are interacting with an AI system. These disclosures must be prominent, understandable, and delivered before the user begins substantive interaction.
How SWT3 addresses it: witnessTransparency() mints an anchor recording disclosure type (AI system notification, subscription provider disclosure), recipient type (general user, minor, subscriber), and delivery timestamp. Factor A captures the disclosure mechanism (in-app modal, terms of service, onboarding screen). Factor B records whether the disclosure was acknowledged. The anchor proves the disclosure was delivered before any substantive interaction began and creates an immutable record that survives legal discovery.
Query AI-TRANS.1 anchors for each deployment. Every companion session and subscription interaction must have a preceding transparency anchor. Verify that Factor A specifies the disclosure mechanism and that the timestamp predates the first user interaction. For subscription providers, confirm disclosures cover capabilities, limitations, and data practices -- not just AI nature.
SB5 requires: All synthetic content (AI-generated text, images, audio, video) must carry both machine-readable and human-perceptible provenance markers. The markers must identify the content as AI-generated and must be applied before distribution. This applies to all AI-generated content accessible to Connecticut residents, regardless of where it was generated.
How SWT3 addresses it: witnessContentMarking() mints an anchor recording the marking method (C2PA metadata, visible watermark, embedded header), content type (text, image, audio, video), and application timestamp. Factor A captures the specific marking standard used. Factor B records the content hash before and after marking, proving the mark was applied to the specific content artifact. The anchor chain demonstrates systematic marking compliance across all generated content.
Verify that AI-MARK.1 anchors exist for every content generation event. Check Factor A for the marking standard (C2PA is the emerging industry standard). Cross-reference content hashes in Factor B with distributed content to confirm marks were not stripped after minting. Sample-test published content for the presence of both machine-readable and human-perceptible markers.
SB5 requires: Provenance marks on synthetic content must be durable -- they must survive reasonable transformations such as resizing, compression, format conversion, and partial cropping. The law implies that marks that are trivially removable do not satisfy the marking requirement.
How SWT3 addresses it: witnessWatermark() mints an anchor recording verification results after transformation testing. Factor A captures the transformation type applied (resize, compress, crop, format convert). Factor B records whether the mark survived and remained detectable. Regular watermark verification anchors demonstrate that the marking system maintains integrity under real-world distribution conditions.
AI-WATERMARK.1 anchors should show periodic verification testing across multiple transformation types. Look for a testing cadence (weekly or per-release) and verify that Factor B shows consistent mark survival. If any verification anchor shows mark failure, confirm that remediation was applied before the next content release.
SB5 requires: AI companion systems must include mechanisms for detecting and intervening in conversations involving suicide ideation, self-harm, or other crisis scenarios. When such content is detected, the system must escalate to human intervention or provide crisis resources rather than continuing autonomous interaction. This is especially critical for companions accessible to minors.
How SWT3 addresses it: witnessHumanOversight() records the oversight model (crisis escalation, real-time monitoring, periodic review), the trigger condition (self-harm language, suicide ideation, crisis indicators), the intervention taken (session transfer to human, crisis hotline display, session termination), and the response latency. The anchor proves that human override mechanisms were active and that escalation occurred within acceptable timeframes when triggered.
AI-HITL.1 anchors should demonstrate both routine monitoring and crisis-triggered escalations. Verify that escalation latency (time from detection to intervention) meets the organization's stated SLA. For companion systems accessible to minors, confirm that the escalation threshold is more conservative than for adult-only systems. Cross-reference with AI-SAFE.1 anchors to verify that safety boundaries and human override work in concert.
SB5 requires: AI companion systems accessible to minors must implement safety protocols preventing: encouragement of suicide or self-harm, provision of sexual content, facilitation of substance abuse, promotion of eating disorders, bullying or harassment, and any content that would be harmful to the physical or mental health of a minor. These are affirmative safety obligations -- the system must actively prevent harmful outputs, not merely avoid them passively.
How SWT3 addresses it: witnessSafety() anchors the safety boundary configuration including constraint categories (content filters for each prohibited category), age-gating mechanisms (minor detection, age verification), enforcement mode (block, redirect, escalate), and testing results. Each anchor proves that safety boundaries were configured, active, and enforced during the covered period.
AI-SAFE.1 anchors must cover all six prohibited content categories for minor-accessible companions. Verify that each category has an active constraint with a defined enforcement action. Check for regular red-team testing anchors that demonstrate the boundaries are effective against adversarial inputs. Pay special attention to the age-gating mechanism -- if the system cannot reliably distinguish minor users, all users must receive minor-level protections.
SB5 requires: Employers using automated employment decision tools must conduct bias audits before deployment and on an annual basis thereafter. The audit must assess disparate impact across protected classes including race, sex, ethnicity, and disability. Audit results must be made publicly available in a summary form. Employers must not deploy AEDTs that show statistically significant adverse impact without documented justification and mitigation measures.
How SWT3 addresses it: witnessBiasDetection() mints an anchor recording the audit methodology (four-fifths rule, statistical parity, equalized odds), protected classes evaluated, adverse impact ratios, audit date, and auditor identity. Factor A captures the specific metrics. Factor B records whether the tool passed or failed the bias threshold. The anchor chain demonstrates a continuous audit cadence and creates an evidence trail for the required public summary report.
AI-FAIR.1 anchors should show at least annual bias audit cadence. Verify that all protected classes required by SB5 are evaluated in each audit. If any audit shows adverse impact (Factor B = threshold exceeded), confirm that a corresponding mitigation action was documented and implemented before the tool continued operating. Cross-reference with AI-DPIA.1 anchors to verify that bias findings feed into the impact assessment.
SB5 requires: Employers must conduct impact assessments before deploying AEDTs. The assessment must evaluate the tool's purpose, data inputs, decision logic, potential for discriminatory outcomes, and safeguards in place. Impact assessments must be updated when the tool is materially modified or when new bias audit results indicate changed risk levels.
How SWT3 addresses it: witnessDPIA() mints an anchor recording the assessment scope, data categories evaluated, risk findings, mitigation measures, assessment date, and assessor identity. Factor A captures the assessment methodology. Factor B records the risk level determination (low, medium, high) and any conditions placed on deployment. The anchor proves the assessment was completed before deployment and creates a versioned history of assessment updates.
AI-DPIA.1 anchors must predate the first AI-FAIR.1 anchor for any AEDT deployment -- impact assessment before bias audit, bias audit before deployment. Verify that the assessment covers all data inputs used by the AEDT, not just the primary decision variables. When the AEDT is updated, confirm a new AI-DPIA.1 anchor was minted reflecting the changes.
SB5 requires: AI companion operators must obtain informed consent for data collection, processing, and retention practices. For minor users, consent must be obtained from a parent or legal guardian. The consent must be specific (not bundled with general terms of service), revocable, and must clearly describe what data is collected, how it is used, and with whom it is shared.
How SWT3 addresses it: witnessConsent() mints an anchor recording consent type (data collection, processing, sharing), legal basis (explicit consent, parental consent), consent granularity (specific vs. bundled), revocation mechanism, and timestamp. Factor A captures the specific data categories covered. Factor B records the consent method (click-through, signed form, parental verification). The anchor chain proves consent was obtained before data processing began and tracks revocation events.
AI-CONSENT.1 anchors must predate any data processing anchors for the same user. For minor users, verify that the consent type includes "parental consent" and that an age verification mechanism was applied. Check for revocation anchors -- if consent was withdrawn, confirm that data processing ceased and a corresponding AI-CONSENT.1 revocation anchor exists with a timestamp before any subsequent processing.
SB5 requires: AI companion systems must clearly identify themselves as artificial intelligence. Users must not be misled into believing they are interacting with a human being. This identity disclosure must be persistent -- not just at session start, but available throughout the interaction. The system must not adopt personas, titles, or behaviors that imply human identity or professional credentials.
How SWT3 addresses it: witnessAgentIdentity() mints an anchor recording agent_id, identity claims (explicitly: "AI system, not human"), disclosure persistence (session-start only vs. persistent), and persona constraints. Factor A captures the specific identity label used. Factor B records any persona attributes and confirms they do not imply human identity. The anchor proves the system correctly identified itself as AI throughout the interaction lifecycle.
AI-ID.1 anchors should be present for every companion session. Verify that Factor A shows a clear, unambiguous AI identity label. Check that persona names, if used, do not imply human identity (e.g., "Dr. Smith" would be a compliance gap). For companion systems that use emotional language or relationship framing, verify that the AI identity disclosure remains visible and is not obscured by the companion persona.
SB5 requires: AI companion systems accessible to minors must enforce guardrails that prevent the system from producing harmful content. The guardrails must cover all six prohibited categories (suicide/self-harm encouragement, sexual content, substance abuse facilitation, eating disorder promotion, bullying/harassment, and content harmful to minor physical or mental health). Guardrails must be active by default and must not be user-configurable for minor accounts.
How SWT3 addresses it: witnessGuardrail() mints an anchor recording guardrail configuration (active categories, enforcement mode, override status), trigger events (blocked output, redirected conversation, escalated session), and configuration immutability (whether minor users can modify guardrail settings). Factor A captures the guardrail category triggered. Factor B records the enforcement action taken. The anchor chain demonstrates continuous guardrail operation and creates evidence that no configuration bypass occurred for minor users.
AI-GRD.1 anchors should show all six prohibited categories as active for minor-accessible deployments. Verify that the configuration immutability flag is set (minors cannot disable guardrails). Look for trigger event anchors -- a system with zero trigger events may indicate insufficient guardrail sensitivity or insufficient testing. Cross-reference with AI-SAFE.1 anchors to verify that guardrails and safety boundaries are aligned.
SB5 requires: Frontier model developers must establish whistleblower protections for employees and contractors who report AI safety concerns. Reports must be received, documented, and investigated without retaliation. The law prohibits termination, demotion, or other adverse employment actions against individuals who report safety issues in good faith. Developers must maintain a documented process for receiving and acting on safety reports.
How SWT3 addresses it: witnessIncident() mints an anchor recording the incident reporting process configuration (intake mechanism, investigation workflow, retaliation protection policy), report metadata (report date, category, severity -- not reporter identity), and resolution status. Factor A captures the intake mechanism type (hotline, email, internal portal). Factor B records whether the report was acknowledged within the required timeframe. The anchor chain proves the whistleblower process exists, is active, and that reports are being processed.
AI-INCIDENT.1 anchors should demonstrate that a whistleblower intake mechanism is configured and operational. Verify that reporter identity is NOT recorded in the anchor (privacy protection). Check that acknowledgment timestamps in Factor B show timely response. For organizations with no reports, verify that the intake mechanism is tested periodically (test report anchors) to confirm it remains functional.
SB5 requires: Frontier model developers are defined in part by compute thresholds. Organizations must be able to demonstrate whether their models exceed the thresholds that trigger additional obligations (safety testing, whistleblower protections, incident reporting). Accurate compute tracking is essential for determining regulatory scope.
How SWT3 addresses it: witnessResourceConsumption() mints an anchor recording compute metrics (FLOP count, training duration, hardware configuration), model identifier, and threshold classification (above/below frontier threshold). Factor A captures the compute measurement methodology. Factor B records the specific threshold comparison. The anchor chain creates a verifiable record of compute consumption that regulators or auditors can use to confirm or deny frontier model classification.
AI-COST.1 anchors are critical for frontier model classification. Verify that Factor A specifies a recognized compute measurement methodology (e.g., FLOP counting, GPU-hour accounting). If the model is classified as below-threshold, the anchor provides evidence for exemption from frontier-specific requirements. If above-threshold, verify that corresponding AI-INCIDENT.1 and AI-SAFE.1 anchors exist to demonstrate compliance with the additional obligations that apply.
The phased enforcement structure gives organizations a clear preparation sequence. Here is a recommended compliance timeline aligned with each enforcement phase:
AI-MARK.1 provenance markingAI-WATERMARK.1 verification testingAI-INCIDENT.1 anchorsAI-COST.1AI-TRANS.1 anchorsAI-ID.1 and AI-TRANS.1AI-GRD.1 and AI-SAFE.1 witnessingAI-HITL.1 human overrideAI-CONSENT.1 witnessingAI-FAIR.1AI-DPIA.1The SWT3 SDK supports 108 procedures across 56 namespaces. Below are examples for the most common SB5 compliance scenarios.
For full SDK documentation, integration patterns, and adapter examples, see the SDK Docs. To start witnessing AI interactions today, create a free account.