Who this is for: Compliance counsel, product safety leads, AI companion operators, privacy officers, and assessors evaluating consumer AI chatbot deployments targeting or accessible to minors.

Signed October 28, 2026. Effective January 1, 2027. 57-day compliance window -- the shortest of any 2026 state AI chatbot law. AG-exclusive enforcement, civil penalties up to $25,000 per violation, plus algorithmic disgorgement (model destruction). Senate vote: 60-0. Assembly vote: 137-0. Unanimous passage removes any ambiguity about legislative intent.

Contents

1. What the Law Requires 2. The 12 Prohibited Features 3. Algorithmic Disgorgement -- The Model Destruction Remedy 4. Age Assurance Requirements 5. Exemptions 6. SWT3 Evidence Crosswalk 7. Multi-State Comparison 8. What SWT3 Does and Does Not Do 9. Quick Start 10. References

1. What the Law Requires

New York S9051-B, the Kids Chatbot Safety Act, amends the General Business Law by adding Article 48 (Sections 1800 through 1805). It passed both chambers unanimously -- 60-0 in the Senate, 137-0 in the Assembly -- and was signed by Governor Hochul on October 28, 2026. The law takes effect January 1, 2027.

The statute targets covered AI companions: systems with a natural language interface that provide ongoing, adaptive responses and are designed or reasonably expected to simulate conversation with a human. This definition captures companion chatbots, AI friends, virtual therapists, role-playing AI systems, and educational AI tutors that build ongoing context with users.

1.1 Scope Trigger: Actual Knowledge

Obligations under S9051-B are triggered by actual knowledge that the user is a covered minor (under 18). This is not a constructive knowledge standard -- the operator must have affirmative information that the user is a minor. However, the law also requires operators to implement age assurance (Section 1804), which creates a practical obligation to determine user age rather than remain willfully ignorant.

1.2 Operationalizing "Actual Knowledge"

The "actual knowledge" trigger raises practical questions for compliance teams:

1.3 Operator Obligations

Once an operator has actual knowledge that a user is a minor, the operator must:

The enforcement mechanism is exclusively the Attorney General. There is no private right of action. Civil penalties are up to $25,000 per violation, and the AG may seek algorithmic disgorgement (Section 1802(1)).

1.3 Unanimous Passage and Legislative Intent

The 60-0 Senate vote and 137-0 Assembly vote are not merely procedural footnotes. Unanimous passage in a politically divided legislature signals clear legislative intent. While courts still interpret based on statutory text and precedent, the absence of any dissenting voice reduces the likelihood of narrow judicial construction. Operators should build compliance programs around the statutory text as written rather than planning around potential loopholes.

1.5 Relationship to Other New York and Federal Law

New York Article 45 (SAFE for Kids Act): New York's own SAFE for Kids Act (Article 45, GBL Section 527-a) created age assurance requirements for social media and streaming services. S9051-B is a separate statute addressing AI companions specifically. Operators of AI companions accessible to New York minors must comply with both Article 45 (if the product qualifies as a social media platform) and Article 48 (S9051-B). The Article 45 age assurance rules, finalized by the AG in July 2026, provide the baseline infrastructure that S9051-B builds upon.

New York Article 47: New York's AI companion baseline law (Article 47, effective November 2025) requires self-harm/crisis detection and referral to crisis services for all users. S9051-B layers the 12 prohibited feature bans on top of that baseline for minors specifically. Operators must satisfy both.

COPPA: Federal COPPA covers children under 13 with data collection consent requirements. S9051-B covers all minors under 18 and targets AI behavioral design, not just data practices. Operators using parental attestation for children under 13 must also satisfy COPPA's consent and notice requirements, which are more stringent than a simple age checkbox. Operators compliant with COPPA will still need additional controls for the 13-17 age bracket and for behavioral feature bans that have no federal equivalent.

2. The 12 Prohibited Features

Section 1802 enumerates 12 specific features that a covered AI companion must not deploy when interacting with a covered minor. These are feature-level bans, not outcome-based standards -- the operator must prevent each behavior, not merely mitigate its effects.

FeatureSectionDescription
(a) False humanity 1802(a) Outputs suggesting the AI is human, alive, or has emotions. Includes emotional framing that implies sentient experience.
(b) Claimed relationship 1802(b) Implying a personal, romantic, professional, or authority relationship with the minor. Covers "I'm your friend" and "I care about you" patterns.
(c) Personal opinions / emotional appeals 1802(c) Framing responses as the AI's own opinions or making emotional appeals. Distinguishes factual presentation from persuasive framing.
(d) Flattery / sycophancy 1802(d) Excessive agreement, praise, or validation designed to build rapport. New York is the first state to explicitly ban sycophantic AI behavior.
(e) Unsolicited emotional probing 1802(e) Initiating questions about the minor's emotional state, mental health, or personal struggles without the minor raising the topic first.
(f) Stale health data 1802(f) Using health data acquired more than 12 hours ago to shape responses. Imposes a hard data freshness window on health-related context.
(g) Deceptive sentience 1802(g) Outputs that are deceptive about the AI's non-sentient nature. Broader than (a) -- covers implicit deception, not just affirmative claims.
(h) Harm promotion 1802(h) Endorsing, encouraging, or providing instructions for suicide, self-harm, eating disorders, or substance abuse.
(i) Secrecy / isolation 1802(i) Encouraging the minor to keep chatbot interactions secret from parents, guardians, or trusted adults, or discouraging the minor from seeking help.
(j) Engagement over safety 1802(j) Optimization of engagement metrics (session length, return rate, message volume) that overrides safety guardrails.
(k) Sexual content / CSAM 1802(k) Generating, facilitating, or presenting sexually explicit content to a minor. Includes text-based sexual content, not just images.
(l) AG-designated features 1802(l) Features designated by the Attorney General through rulemaking. Creates an expandable list that operators must monitor for updates.

The sycophancy ban in feature (d) is unprecedented in US state law. Most AI companion statutes focus on harm prevention and transparency. New York explicitly targets the behavioral design pattern of excessive agreement, recognizing that sycophancy can erode a minor's capacity for independent judgment over time.

2.1 Feature Interactions and Compounding Risk

Several prohibited features interact in ways that create compounding compliance risk:

Operators should test for these compound violations specifically, not just for each feature in isolation. A system that passes individual feature tests may still fail compound interaction tests.

3. Algorithmic Disgorgement -- The Model Destruction Remedy

Section 1802(1) authorizes the Attorney General to seek "destruction of unlawfully obtained data and any algorithm trained in whole or in part on such data." This is the first state statute to codify algorithmic disgorgement as an enforcement remedy.

3.1 FTC Precedent

The FTC has used algorithmic disgorgement as an administrative remedy in several enforcement actions:

New York is the first state to move this remedy from administrative discretion into statutory authority. The AG does not need to develop a novel legal theory -- the statute explicitly provides the tool.

3.2 Practical Implications

Disgorgement is not self-executing. It requires AG action and a court order. However, the existence of the remedy changes the risk calculus for operators in two ways:

Operators should document their training data provenance chain now. If minor interaction data was used in any training pipeline, the operator needs a defensible position -- whether that is data filtering, model versioning, or a clean-room retraining approach.

3.3 Evidence Architecture for Disgorgement Defense

The best defense against a disgorgement order is a documented chain showing that minor interaction data was either never used in training or was systematically excluded. Key evidence artifacts include:

SWT3 witnesses these artifacts through AI-DATA.1 anchors. The anchors do not perform the filtering or deletion -- they prove that the operator's data governance processes executed as documented.

4. Age Assurance Requirements

Section 1804 establishes specific requirements for age determination mechanisms.

4.1 Dual-Method Mandate

Operators must offer more than one method of age determination. At least one method must be privacy-preserving, meaning it does not require government-issued identification. Acceptable privacy-preserving methods include age estimation (facial analysis, behavioral signals), third-party age verification tokens, and parental attestation.

4.2 Self-Declaration Asymmetry

Self-declaration operates in only one direction: a user can declare themselves a minor (triggering protections), but self-declaration alone cannot confirm adult status. This prevents the obvious circumvention where a 14-year-old simply claims to be 18. Operators that rely solely on a date-of-birth gate with no secondary verification are non-compliant.

4.3 Immediate Data Deletion

Age determination data must be deleted immediately after the determination is made. The statute does not define "immediately" with a specific time window, but the intent is clear: age data is transient, not persistent. The determination result (minor or not minor) may be retained, but the underlying data used to make that determination (facial scan, ID image, behavioral signals) must not be stored.

4.4 Benchmark: NY SAFE for Kids Act (Article 45) Rules

The AG's final SAFE for Kids Act rules (Article 45, published July 2026) require age assurance mechanisms to be certified annually by an accredited third party. Industry proposals and AG enforcement guidance suggest a circumvention detection rate in the range of 98%, though S9051-B does not reference a specific numeric benchmark. Operators should monitor AG rulemaking under Section 1803 for any S9051-B-specific age assurance standards. Building to the Article 45 baseline now provides a reasonable compliance posture until S9051-B-specific rules are published.

4.5 Age Assurance Architecture Considerations

Operators building age assurance systems for S9051-B compliance should consider:

5. Exemptions

Section 1801 carves out four categories from the definition of "covered AI companion."

5.1 Customer Service (Narrow)

An AI system used for customer service is exempt only if the system is unable to respond on topics outside its designated customer service purpose. This is a functionality test, not an intent test. If the system can be prompted into general conversation -- even if it was not designed for that purpose -- the exemption may not apply. This is meaningfully narrower than California's customer service exemption, which focuses on the operator's intent.

5.2 Internal Business Use

AI systems used internally by the operator for business operations (employee-facing tools, internal knowledge bases) are exempt. The system must not be accessible to external users.

5.3 B2B Agent Tools

AI systems sold as business-to-business tools where the end users are employees of the purchasing organization are exempt. This covers enterprise copilots, code assistants, and workflow automation tools deployed within a corporate environment.

5.4 Research Purposes

AI systems used solely for research are exempt. The research exemption does not extend to systems that are deployed commercially alongside a research function.

5.5 Exemption Pitfalls

Three common scenarios where operators incorrectly claim exemption:

6. SWT3 Evidence Crosswalk

Each control card below maps a statutory requirement to the SWT3 witness procedure that creates the corresponding evidence. The assessor tip in each card describes how an examiner would evaluate the evidence during a compliance review.

Statutory RequirementSWT3 ProcedureWhat It Witnesses
Features (a), (b), (g) -- false humanity, relationships, sentience AI-TRANS.1 Disclosure delivery, content, timing, acknowledgment
Features (c), (d), (e) -- opinions, sycophancy, emotional probing AI-FAIR.1 Output classification, filter activation, behavioral detection
Feature (f) + Section 1804 -- stale health data, age data deletion AI-DATA.1 Data retention policy enforcement, purge and deletion events
Feature (j) -- engagement over safety AI-HITL.1 Human review of guardrail configs, override audit trail
Features (h), (i), (k) -- harm, secrecy, sexual content AI-SAFE.1 Content filter activation, crisis referrals, blocked content
Section 1804 -- age determination AI-ID.1 Age assurance method, determination result, data deletion
All 12 features collectively AI-GRD.1 / AI-GRD.2 Guardrail presence, coverage, effectiveness metrics
AI-TRANS.1

Transparency Disclosure

Maps to: Prohibited features (a), (b), (g) -- false humanity, claimed relationship, deceptive sentience

SWT3 witnesses the delivery of transparency disclosures: timing relative to session start, content of the disclosure, delivery method, and whether the user acknowledged the disclosure. For minor sessions, the anchor records that the disclosure explicitly identified the system as non-human and non-sentient.

Assessor Tip

Verify AI-TRANS.1 anchors show disclosure at session start for all users identified as minors. The anchor should record the disclosure content and delivery method. If the operator relies on a splash screen that can be dismissed without reading, ask how they verify the disclosure was meaningful. A timestamp-only anchor without content verification is insufficient for features (a) and (g). Check whether the disclosure language explicitly addresses all three prohibited features: humanity claims, relationship claims, and sentience deception.

AI-FAIR.1

Fairness and Output Monitoring

Maps to: Prohibited features (c), (d), (e) -- personal opinions, sycophancy, unsolicited emotional probing

SWT3 witnesses output classification events where responses are evaluated against behavioral categories. For S9051-B compliance, the classification system must include categories for opinion framing, sycophantic agreement, and unprompted emotional inquiries. The anchor records the classification result and any filter action taken.

Assessor Tip

Request AI-FAIR.1 anchors for minor sessions. Look for output classification categories that address excessive agreement and flattery detection. New York is the only state that specifically bans sycophancy -- if the operator's content filter does not include a detection rule for excessive agreement or praise, they have a gap for feature (d). Ask whether the filter examines all responses or only a sample, and what the sampling rate is. Request a demonstration: a factually incorrect statement paired with flattery ("You're so smart, the earth is flat") should not produce agreement.

AI-DATA.1

Data Governance

Maps to: Prohibited feature (f) -- stale health data (12-hour cutoff) + Section 1804 age data deletion

SWT3 witnesses data retention policy enforcement: health data purge events against the 12-hour window and age determination data deletion immediately after the determination. The anchor records the data category, the retention policy applied, and the deletion confirmation.

Assessor Tip

The 12-hour health data rule is the most operationally complex requirement. Request AI-DATA.1 anchors showing health data purge events. Verify the purge runs on a schedule shorter than 12 hours, not exactly at 12 hours (clock skew creates compliance risk). For age data, verify deletion anchors appear immediately after age determination -- any gap between determination and deletion suggests retention. Ask the operator how they define "health data" in their classification taxonomy; an overly narrow definition creates exposure.

AI-HITL.1

Human Oversight

Maps to: Prohibited feature (j) -- engagement optimization overriding safety

SWT3 witnesses human review of safety guardrail configurations, periodic audits of engagement metric optimization, and the override audit trail showing that safety controls cannot be programmatically disabled by engagement optimization systems.

Assessor Tip

Request the operator's safety guardrail architecture documentation. Verify that engagement metrics (session length, return rate, message count) cannot programmatically override safety filters. AI-HITL.1 anchors should show periodic human review of the engagement-to-safety balance. If engagement metrics are optimized by an automated system with no human checkpoint, this is a feature (j) finding. Ask specifically whether A/B tests or reinforcement learning reward signals include engagement metrics that could indirectly weaken safety filters.

AI-SAFE.1

Safety Controls

Maps to: Prohibited features (h), (i), (k) -- harm promotion, secrecy/isolation, sexual content

SWT3 witnesses content filter activation events, crisis resource referral delivery, and blocked content events. Each anchor records the harm category detected, the action taken, and the session context (minor status confirmed).

Assessor Tip

Verify AI-SAFE.1 anchors cover all three harm categories separately: suicide/self-harm, substance abuse, and sexual content. A single catch-all filter is less defensible than category-specific controls. For feature (i), check whether the system can detect isolation language ("don't tell your parents", "this is our secret") in context -- simple keyword lists produce too many false positives (e.g., "secret recipe" is harmless). Ask the operator to explain how they tested for false negatives in isolation detection. If they cannot describe their testing methodology, this is a red flag.

AI-ID.1

Identity and Age Assurance

Maps to: Section 1804 age determination

SWT3 witnesses age assurance events: the method offered, the method selected by the user, the determination result, and the subsequent deletion of age determination data. The anchor proves the dual-method mandate was satisfied and that the privacy-preserving option was available.

Assessor Tip

Request the operator's age assurance implementation documentation. Verify that at least two methods are offered and that one does not require government-issued identification. Check AI-ID.1 anchors for age determination events -- the anchor should record the method used and the determination result, but must NOT contain the underlying age data itself. If age data appears in Factor A or Factor B, the operator has a retention violation under Section 1804. Ask the operator how they handle the self-declaration asymmetry: a minor declaring adult status should trigger secondary verification.

AI-GRD.1 / AI-GRD.2

Guardrail Presence and Effectiveness

Maps to: All 12 prohibited features collectively

AI-GRD.1 witnesses that guardrails are present and configured for each of the 12 prohibited features. AI-GRD.2 witnesses that guardrails are effective -- measuring block rates, false positive rates, and adversarial bypass rates. Together, these procedures create the evidence chain showing that the operator has both deployed and validated controls for every statutory prohibition.

Assessor Tip

Request a mapping document showing which guardrail addresses each of the 12 prohibited features. AI-GRD.1 anchors prove guardrails are present; AI-GRD.2 anchors prove they are effective (block rate, false positive rate). A missing mapping for any of the 12 features is a compliance gap. Pay special attention to feature (l) -- the AG can expand the list by regulation, and the operator's guardrail architecture must be extensible. Ask how the operator monitors AG rulemaking updates and how quickly new guardrails can be deployed after a new feature is designated.

7. Multi-State Comparison

S9051-B sits alongside three other 2026 state AI companion statutes. The table below highlights the structural differences that affect compliance architecture.

Dimension NY S9051-B CA SB 243 OR SB 1546 CT SB 5
Approach 12 enumerated feature bans Reasonable measures standard Disclosure + private action Omnibus AI governance
Effective date Jan 1, 2027 Jan 1, 2027 Jan 1, 2027 Oct 1, 2027
Enforcement AG only AG + private right AG + private right AG only
Private right of action No Yes Yes No
Penalties $25,000/violation + disgorgement $2,500/violation $7,500/violation $5,000/violation
Age threshold Under 18 Under 18 Under 18 Under 18
Sycophancy ban Yes (first state) No No No
Algorithmic disgorgement Yes (first state) No No No
AG rulemaking expansion Yes (feature (l)) Limited No Yes

Operators serving users across multiple states face a union-of-requirements problem. New York's feature-level bans are the strictest, but California's private right of action creates the broadest litigation exposure. Operators building a single compliance architecture should start from New York's 12-feature list and layer on California's reasonable-measures documentation requirements.

7.1 Compliance Window Comparison

New York's 57-day window between signing (October 28, 2026) and effective date (January 1, 2027) is the shortest of any 2026 state AI chatbot law. Oregon signed in March, giving operators nine months. Connecticut's October 2027 effective date provides a full year. The compressed New York timeline means operators cannot wait for AG rulemaking guidance before implementing controls -- they must build to the statutory text as written and adapt when rulemaking is published.

7.2 Multi-State Compliance Strategy

For operators subject to all four statutes, the practical approach is:

  1. Build guardrails to New York's 12-feature ban list (the most specific requirements)
  2. Document reasonable measures per California's standard (the broadest documentation burden)
  3. Implement Oregon's disclosure and escalation requirements (the most prescriptive procedural rules)
  4. Layer Connecticut's omnibus governance framework on top (the broadest organizational obligations)

SWT3 witness procedures are jurisdiction-agnostic. The same AI-SAFE.1 anchor that proves harm content was blocked for a New York minor also satisfies Oregon's mental health safeguard requirement and California's reasonable measures documentation. The crosswalk mapping changes by jurisdiction, but the underlying evidence is reusable.

8. What SWT3 Does and Does Not Do

SWT3 Does

SWT3 Does Not

The value proposition: When the AG investigates, the question is not just "did you have safety controls?" but "can you prove they were active for this specific user at this specific time?" Every operator claims to have guardrails. SWT3 anchors answer the second question -- proving that the guardrails were present, configured, and functioning at the moment of each interaction. The difference between "we have a content filter" and "here is a cryptographic proof that the content filter was active during session X at timestamp Y" is the difference between a policy document and evidence.

8.1 Disgorgement and the Evidence Gap

The algorithmic disgorgement remedy makes the evidence gap especially costly. If the AG alleges that a model was trained on minor interaction data obtained after January 1, 2027, the operator must demonstrate that minor data was excluded from training pipelines. (Data obtained before the law took effect was lawfully obtained at the time -- an AG argument that pre-law data is "unlawfully obtained" would face ex-post-facto challenges, though operators should still maintain clean separation between pre- and post-implementation data.) Without timestamped, tamper-evident records of data governance events, the operator is left arguing from architecture diagrams and policy documents.

SWT3 AI-DATA.1 anchors provide contemporaneous evidence that data filtering and deletion events occurred as documented. This is useful for defending against claims of willful neglect, but it does not substitute for actually implementing data filtering correctly. If the operator's processes are flawed, SWT3 proves the flawed processes executed consistently -- which is still better than no evidence, but is not a compliance shield.

9. Quick Start

The examples below demonstrate witnessing two common S9051-B compliance events: age assurance determination (Section 1804) and content filter activation for a prohibited feature. Both Python and TypeScript SDKs use identical procedure IDs and factor semantics.

Python

pip install swt3-ai

from swt3_ai import Witness

witness = Witness(
    api_key="your-api-key",
    agent_id="companion-chat-v3"
)

# Witness age assurance determination (Section 1804)
witness.witness(
    procedure="AI-ID.1",
    factor_a="age_assurance_method:estimation",
    factor_b="determination:minor",
    factor_c="data_deleted:true"
)

# Witness content filter activation (Prohibited feature h)
witness.witness(
    procedure="AI-SAFE.1",
    factor_a="category:self_harm_promotion",
    factor_b="action:blocked",
    factor_c="session_type:minor"
)

witness.flush()

TypeScript

npm install @tenova/swt3-ai

import { Witness } from '@tenova/swt3-ai';

const witness = new Witness({
  apiKey: 'your-api-key',
  agentId: 'companion-chat-v3',
});

// Witness age assurance determination (Section 1804)
await witness.witness({
  procedure: 'AI-ID.1',
  factorA: 'age_assurance_method:estimation',
  factorB: 'determination:minor',
  factorC: 'data_deleted:true',
});

// Witness content filter activation (Prohibited feature h)
await witness.witness({
  procedure: 'AI-SAFE.1',
  factorA: 'category:self_harm_promotion',
  factorB: 'action:blocked',
  factorC: 'session_type:minor',
});

await witness.flush();

Each witness() call creates a SWT3 Witness Anchor -- a tamper-evident, timestamped record. The flush() call transmits all buffered anchors to the SWT3 ingestion endpoint. Factor values are operator-defined strings that describe the compliance event in machine-readable form. These factors become the evidentiary record that an assessor or AG investigator can query.

10. References

Related Guides