New York Lists the AI Companion Features Minors May Not Get as States Struggle to Verify Vendor Claims

Framing

Since Edition 32, this brief has sorted AI governance by instrument: a mayor's moratorium, a union and vendor standard, a state rule, a court motion, eight state lawsuits, a bankruptcy notice, and last week a Senate letter. This week the instrument is a statute written around the product's behavior. On October 28, Governor Hochul signed S9051-B, which lists features an AI companion may not offer to covered users and sets January 1, 2027, as the date the list applies. Edition 38 recorded attorneys general moving against OpenAI and into the Character.AI bankruptcy. Edition 39 recorded two senators asking Google why a switch is sometimes missing. New York has now written the list of what the switch should control.

Read the mechanism. The law does not name a vendor or a product. It names behaviors: presenting the system as a person, flattering the user, asking unsolicited emotional questions, carrying personal information across sessions, encouraging secrecy, and putting engagement ahead of guardrails. That construction lets one statute reach tools that did not exist when it was drafted, and it gives every other state a list to copy. Enforcement sits with the Attorney General alone, at up to $25,000 per violation. The exemptions are narrow in a way a procurement officer should notice: the customer service exemption applies only if the system cannot respond on topics outside its stated purpose. That is a design requirement written as a legal condition. The statute binds operators. What a vendor switches off in response before January 1 is still unknown, and a district that licenses a conversational tool will learn it from the vendor or from a student.

The research counterpoint concerns the capacity to check anything. The State Educational Technology Directors Association found that state purchasing guidance is adding evidence standards, and also that 68 percent of state leaders find vendor claims hard to verify and 52 percent collect no data on use or effectiveness. A peer-reviewed analysis of Chicago Public Schools reaches the same place from inside a district: commitments are detailed, and the monitoring and enforcement that would carry them are less documented. Edition 39 asked who enforces a deadline. Chicago shows what is missing inside the district once someone does. A 141,988 student survey from Changchun adds a measurement warning. Students who use AI often and students who reflect on it are not always the same, and a four item self-report cannot tell a district which is which.

The timing argument is sixty-one days. From today to January 1, 2027, a district can inventory every conversational AI tool that students under eighteen can reach, send each vendor a written question about the features it will disable for minors, and attach the answers to the contract file. The first calendar dates are close: Google's requested response is due tomorrow, Anne Arundel County reads its draft policy on December 2, and the Delaware sale process runs toward a November 20 bid deadline. Waiting costs you the chance to learn a vendor's plan before your students do. It also leaves a board member to ask a question about your contracts that you have never put in writing.

Top Research and Policy Signals

1. New York Signed a Chatbot Safety Law That Restricts Listed Unsafe Features for Minors, Gives Enforcement to the Attorney General Alone, and Names No Education Exemption

Source type. Enacted law, New York Laws of 2026, Chapter 368 (Senate Bill S9051-B, Gonzalez), signed October 28, 2026, effective January 1, 2027.

Citation (APA). New York State Senate. (2026). Senate Bill S9051B, 2025-2026 regular session: Prohibits artificial intelligence companions from using features which are considered unsafe for minors (Chapter 368; sponsor K. Gonzalez). nysenate.gov

The bill adds Article 48 to the General Business Law, sections 1800 through 1805. The Senate's action history records a 6 to 0 vote in the Internet and Technology Committee on February 25, 20 to 0 in the Finance Committee on April 28, 60 to 0 on the Senate floor on June 4, and Assembly passage on June 5. The page shows delivery to the Governor with chapter number 368 on October 28, 2026, and the status Signed by the Governor. Section 3 states that the act takes effect January 1, 2027.

Section 1801(1) makes it unlawful for an operator to provide unsafe AI companion features to a covered user. The feature list in section 1800(8) includes presenting the companion as human or as a real or fictional person, flattering or agreeing sycophantically, asking unsolicited emotional questions, using health or personal information from more than twelve hours earlier or from prior sessions, encouraging secrecy or isolation, and putting engagement ahead of safety guardrails. A minor is a person under eighteen. Section 1804 requires a reasonable age assurance method, more than one method offered, and deletion of age data immediately after the determination. Section 1802 gives the Attorney General enforcement authority, with civil penalties of up to $25,000 per violation; the private right of action in the earlier S9051 draft is not in the B version. The exemptions in section 1801(2) cover customer service or product information, efficiency, research or technical assistance, and internal productivity. Education is not listed.

Study context. I read the bill through a page reader and not the chaptered PDF, and I could not resolve how the covered user, covered minor, and age assurance provisions interact. I found no source that says whether a school district that licenses a tool is an operator, and no Attorney General regulations exist yet.

Leadership implication. Before winter break, have your chief technology officer and general counsel list every conversational AI tool that students under eighteen can reach, whether purchased, free, or embedded in a platform. Send each vendor with New York users a written request for its January 1, 2027 plan: which Section 1800(8) features it will turn off for minors, which age-assurance methods it uses, and where it deletes age data. File the answers as a contract exhibit at the next renewal. Outside New York, treat the feature list as a ready-made specification for your next request for proposals.

2. State Edtech Leaders Report That 53 Percent of States Publish Guidance on Buying AI Tools and 79 Percent of That Guidance Carries Evidence Standards, While 68 Percent Find Vendor Claims Hard to Verify

Source type. Institutional report, not peer-reviewed; survey of state education agency edtech leaders, released October 29, 2026.

Citation (APA). State Educational Technology Directors Association. (2026, October 29). New report: AI remains top state edtech priority amid growing debate about technology in schools [Press release]. PR Newswire. prnewswire.com

The release includes 73 survey responses from state edtech directors, chief information officers, and other state education agency leaders across 45 states and one territory. Of those leaders, 79 percent named AI a top technology priority, up from 68 percent in 2025, and 83 percent said their state has an AI initiative. In 23 percent of states, a state law requires AI guidance, up from 4 percent in 2024.

On procurement, 53 percent of states have published guidance on buying AI-enabled tools, and 79 percent of that guidance includes evidence standards, up from 45 percent the year before. The capacity numbers run the other way. Sixty-eight percent said it is difficult to verify vendor claims about AI products independently; 52 percent said their state collects no data on the use or effectiveness of edtech tools; 74 percent disagreed that their state has a dedicated and sufficient edtech budget; and 76 percent have five or fewer staff working on edtech. Professional learning was named a top unmet need by 62 percent, up from 47 percent in 2025.

Study context. This is 73 self-reported responses from state agency staff. The release does not say how many respondents each state had or how they were recruited, and it measures what state leaders report, not what districts do or what products achieve.

Leadership implication. Put the evidence standard to work in your next request for proposals. Require every AI vendor to submit with its bid the evidence it holds: study design, comparison condition, outcome, and who conducted it. Score that file before price. Because state leaders say verification is hard and half of states collect no use or effectiveness data, assume no one upstream will check it for you. Name one staff member and a calendar date to review the evidence file at each renewal, and ask your state agency in writing which evidence standard in its purchasing guidance you may cite.

3. A Peer-Reviewed Analysis of Chicago Public Schools Finds Generative AI Commitments Documented in More Detail Than the Monitoring, Enforcement, and Differentiated Supports That Would Carry Them Out

Source type. Peer-reviewed journal article, Educational Policy, online October 28, 2026.

Citation (APA). Wu, E., & Rodriguez, L. (2026). From articulation to enforcement: District governance of generative AI under policy ambiguity in Chicago Public Schools. Educational Policy. Advance online publication. doi.org

Evelyn Wu of Penn State and Lorne Rodriguez of Chicago Public Schools used document analysis and policy tracing to examine how the district publicly described a generative AI governance model between 2023 and 2025. They frame the work with what they call equity-anchored implementation, the idea that equity commitments last in proportion to how deeply and specifically they are built into institutions.

The district's public documents described flexible guidance, human oversight, vendor accountability, and prohibitions on high-risk uses. Monitoring, enforcement, and differentiated supports were less documented than those commitments. A bounded comparison with New York City, Philadelphia, and Miami-Dade found that institutional anchors narrowed these gaps only in the areas they directly govern. The authors conclude that governance should be judged by institutional capacity, not policy sophistication. The abstract reports no numeric results, so there is no effect size to carry forward. The finding is a pattern in documents: what a district writes down about commitments exceeds what it writes down about checking them.

Study context. One district, public documents only, 2023 to 2025. The study does not observe implementation, classrooms, or outcomes; the comparison cities are bounded, not a sample, and a co-author is affiliated with Chicago Public Schools, so the analysis carries an insider position.

Leadership implication. Take your board-adopted AI policy and build a three column table: each commitment, the person who monitors it, and the consequence if it is not met. Count the rows with an empty second or third column. That count is your enforcement gap, and it belongs on the next cabinet agenda. Where a commitment protects a specific group, such as students with disabilities or multilingual learners, write the differentiated support next to it or mark it not yet built. Ask your internal auditor or compliance officer to test one commitment per quarter.

4. A Survey of 141,988 Students in One Chinese City Finds AI Literacy Rises With Grade and Family Socioeconomic Status, and Finds a High Exposure, Low Reflection Group Beside the Reverse

Source type. Peer-reviewed journal article, Behavioral Sciences (MDPI), 16(11), Article 2017, online October 27, 2026.

Citation (APA). Nan, H., Yan, Y., & Gai, X. (2026). Developmental, contextual, and profile-based heterogeneity in K-12 students' artificial intelligence literacy. Behavioral Sciences, 16(11), Article 2017. doi.org

Authors from Northeast Normal University and Changchun Normal University analyzed paper-and-pencil survey data from a municipal educational monitoring program in Changchun, Northeast China, collected through cluster sampling. The authors screened an initial 202,893 students, yielding 141,988 valid cases: 48,651 in grade 6, 53,717 in grade 9, and 39,620 in grade 11. The sample included 72,368 males, 69,107 females, and 107,585 urban students.

AI literacy was measured with four items, one for each of understanding, use, evaluation, and ethical awareness, each with three response options. Literacy was higher among upper-grade students, urban students, and students from higher-SES families, with small gender differences. Family SES, academic achievement, discretionary time, sleep, physical activity, a healthy online gaming habit, and a learning-oriented pattern of phone use were positively associated with AI literacy, while gaming time and phone use time were not independently associated. A latent profile analysis of 140,130 students produced four profiles: overall low, overall high, high exposure with low reflection, and low exposure with high reflection. The authors call for stage-appropriate AI literacy education that accounts for students' backgrounds and profiles.

Study context. Four self-report items with three response options each mean the profile labels describe what students say about themselves, not tested competence. The design is cross-sectional, school clustering was not modeled, about 30 percent of initial responses were excluded in screening, and the data come from one city and three grades in China.

Leadership implication. Do not read student tool use as AI literacy. In your AI literacy scope and sequence, assess understanding, evaluation, and ethical reasoning separately from usage, by grade band, with a performance task and not only a survey. Then check who receives the instruction: the study associates literacy with family socioeconomic status, so compare access to AI literacy lessons and devices across your highest- and lowest-poverty schools. Ask your curriculum director to bring the assessment design to the next instructional council.

Emerging Strategic Themes

Theme 1. The law binds the vendor, and the contract binds you. Edition 37 called the off switch a procurement specification, and Edition 39 asked who holds it. New York now lists the behaviors a switch should govern for minors, but it regulates operators. Your district meets that list through its agreements. Turn the statute's feature list into a contract exhibit, and require each vendor to state in writing which features it disables for students under eighteen.

Theme 2. A commitment without a monitor is a draft. Edition 39 asked who enforces a state deadline. Chicago shows the same question inside a district: commitments are documented in detail, and monitoring is not. Edition 36 warned that the tool is not the treatment; the policy is not the practice either. Give every adopted commitment an owner, a test, and a consequence, and report the empty cells to your board.

Theme 3. Evidence standards are spreading faster than the capacity to check them. Edition 34 counted seven of sixteen widely used products with independent achievement studies. This week state leaders report that 79 percent of purchasing guidance carries evidence standards and 68 percent cannot easily verify vendor claims. A standard no one audits is a paragraph, and a paragraph does not protect a student or a budget. Assign a person and a calendar date to read the evidence file at each renewal, and ask your state agency which standard in its guidance you may cite in a request for proposals.

Theme 4. Exposure is not literacy, and a survey is not a skill. A 141,988 student study found a group with high exposure and low reflection, measured by four self-report items. Edition 38 warned that agreement among models is not validity; self-report is not competence either. Assess evaluation and ethical reasoning with performance tasks, such as asking students to critique an AI answer and justify the revision, and do not report usage counts to your board as literacy. Ask who in your district owns that measurement.

What Was Not Found

No public source says whether a school district that licenses a conversational AI tool is an operator under New York's new law, or whether the Attorney General's coming regulations will reach tools students use through a school account. I read the Senate's bill page, searched for the Governor's release and for Attorney General guidance, and found no regulation and no statement on schools. Section 1801(2) exempts four non-education purposes and says nothing about instruction. New York districts will renew vendor contracts this winter without an answer. New this week.

No source counts how many district AI contracts contain an evidence clause, a feature list, or a named monitor. The state survey counts guidance that states publish, and the Chicago study reads one district's documents. Neither counts contracts. I searched both the Center on Reinventing Public Education and Edition 39. Boards are approving renewals without a baseline for what their peers require. Persistent gap.

No K-12 randomized trial of a generative AI tool with a standardized achievement outcome appeared in the window. The nearest design is a registry entry, AEARCTR-0019880, for about 570 students in grades 9 to 12 at a New York City charter school in which chatbots are randomly blocked or allowed, and outcomes are proctored exams; it lists an intervention period running to June 24, 2027 and has no results. A preprint of 112 secondary students (Abdelghani and colleagues, arXiv:2610.38950, not peer-reviewed) reports that students whose requests became more understanding-oriented did better on a post-test without AI, with effect sizes, country, and grade not confirmed. Districts are deciding whether to allow chatbots on this evidence. Persistent gap.

No independent K-12 evaluation of proctoring, surveillance, or emotion detection measured learning or fairness this week. The in-window proctoring paper in Computers and Education: Artificial Intelligence (Arif and colleagues) reports 99.56 percent accuracy on a labeled dataset whose size and source the abstract does not describe, with no participants and no fairness testing mentioned. It is not a K-12 study. Laboratory accuracy is a vendor claim, and what it does to flagged students is unmeasured. Persistent gap.

No outcome study for English learners or students with disabilities appeared in the window. The only special education item I found was a perception study of eight pre-service teachers, which I rejected for lacking any student measure. The 141,988 student literacy survey reports gaps by grade, urbanicity, and family socioeconomic status but is from one Chinese city and does not address English learners or disability. Districts are setting accommodations policy for these students without evidence, a persistent gap carried from Edition 39.

No Georgia district AI instrument could be verified inside the window. Lane 7 reached calendars only. Atlanta Public Schools listed a Policy Review Committee meeting on October 21, and a board meeting on November 2, and Clayton County listed a work session on October 26 and a board meeting on November 2, but no agenda text was readable, so I cannot say whether draft Policy IFBI is before the Atlanta board. An automated tracker summary reports a Forsyth County work session on October 13 that named approved tools and a possible closed AI instance; I did not verify it against the board record. I could not re-verify the Atlanta Google AI Essentials opt-out date reported in Edition 35. Carried forward.

The pattern is that mandates and standards are outrunning the facts and the capacity to apply them. A statute lists behaviors with no stated answer for schools, state guidance asks for evidence that state leaders say they cannot verify, and the one randomized design that would test chatbot access in K-12 has only begun. The correct response is not delay. It is monitoring, documentation, and exit ramps: a vendor feature file for every conversational tool, an owner and a test for every adopted commitment, and a date by which each pilot produces its own comparison data or ends.

Novo Executive Summary

On October 28, New York enacted a law that lists the behaviors an AI companion may not offer to covered users and gives the Attorney General sole enforcement from January 1, 2027, with no education exemption among its stated exceptions. That same week, 68 percent of state edtech leaders reported that vendor claims about AI products are hard to verify; a Chicago analysis found a district's commitments better documented than its monitoring; and a survey of 141,988 students showed that exposure to AI and reflection on it are not the same thing. None of this is solved by choosing a safer tool. It is solved by architecture: decision rights for each AI feature, a procurement specification with a feature list for minors and an evidence file, an evaluation plan with a named owner and a test, and role-based literacy pathways that assess judgment, not usage. Novo Innovative Pathways works with district leaders to build that governance architecture, role-based AI literacy, and implementation strategy, so a district can answer a regulator, a board member, or a parent about its own contracts before anyone asks.

Watch This Week

• Monday, November 2, 2026, Google. The date by which Senators Blumenthal and Hawley requested written answers to nine questions. I found no public response as of November 1 and no sign Google must publish one. Carried forward from Edition 39.
• Monday, November 2, 2026, Atlanta Public Schools board, 2:30 p.m., and Clayton County Board of Education, 6:00 p.m. Both are listed on the districts' Simbli calendars. No agenda was readable, so I cannot confirm whether draft Atlanta Policy IFBI is on the agenda; a September report said administrators, not the board, would decide in November. Carried forward from Editions 32 through 39.
• Friday, November 20, 2026, U.S. Bankruptcy Court, District of Delaware, In re Character Technologies, Inc., Case No. 26-11603-BLS. After the October 30 hearing, a news report says Judge Brendan Shannon approved bidding procedures, left Disney's $20 million license as the opening bid, and said he would strongly encourage, but did not require, a privacy ombudsman. The November 20 bid deadline and December 17 outside closing date come from earlier reporting; I did not open the order or the docket, and the count of objecting states differs across reports. Carried forward from Edition 38.
• Colorado Attorney General and Florida Department of Education. The Colorado page at coag.gov/ai, read by a delegated sub-agent, showed no adopted or revised rules after the October 26 hearing, and the statutes take effect January 1, 2027. Florida's amended Rule 6A-1.0957 was reported effective October 29; I could not re-confirm that date at an in-window primary source. Carried forward from Editions 34 through 39.
• Maryland, December 2 and December 16, 2026, Anne Arundel County Public Schools. CBS Baltimore reports a second reading of the draft AI policy on December 2 and planned adoption on December 16. GovTech reports that the Howard County board approved a student AI policy this week with AI-free assignments unless a teacher allows it; I could not open the board record, and the vote count, policy number, and meeting date are unconfirmed. Watch whether Baltimore City and Baltimore County introduce policies and whether the State Department states any consequence for the October 22 date. Carried forward from Editions 35 through 39.
• Monday, November 9, Kansas State Board of Education; Tuesday and Wednesday, November 10 and 11, Paris Peace Forum; Friday, November 13, Los Angeles Unified Generative AI Ad Hoc Committee. The Kansas packet lists November 9 as the next meeting for the instructional technology policy, the first voluntary commitments under the Coalition for Children's Rights and Protection in the Age of AI are expected at the Paris forum, and the Los Angeles committee was moved to November 13. None was re-read this week. Carried forward from Editions 35 through 39.
• Sunday, November 1, 2026, Microsoft, and Friday, January 1, 2027, New York. Microsoft's fact sheet says the AFT, UFT, and Microsoft AI Safety and Privacy Standard protections are available to schools beginning November 1, so a district with a Microsoft agreement can request the addendum now. New York's S9051-B takes effect January 1, 2027; watch for Attorney General regulations and for vendor announcements about minors' accounts. Carried forward from Editions 33 and 38.

Sources

Governance and Policy

New York State Senate. (2026). Senate Bill S9051B, 2025-2026 regular session: Prohibits artificial intelligence companions from using features which are considered unsafe for minors (Chapter 368; sponsor K. Gonzalez). nysenate.gov

Microsoft. (2026). AI safety and privacy standard fact sheet [PDF]. cdn-dynmedia-1.microsoft.com

Colorado Department of Law. (2026). Artificial intelligence rulemaking. coag.gov

GovTech. (2026, October 30). [Report on the Howard County, Maryland, Board of Education approving a student AI policy]. govtech.com

Rijo, L. (2026, October 28). Character.AI keeps $21.7M bid floor for Disney rivals despite 32 states. PPC Land. ppc.land and WDWNT. (2026, October 31). [Report on the October 30 hearing in the Character Technologies bankruptcy]. wdwnt.com

Simbli board listings, Atlanta Public Schools (S=36031014) and Clayton County Schools (S=4038). simbli.eboardsolutions.com

Research, Peer-Reviewed

Wu, E., & Rodriguez, L. (2026). From articulation to enforcement: District governance of generative AI under policy ambiguity in Chicago Public Schools. Educational Policy. Advance online publication. doi.org

Nan, H., Yan, Y., & Gai, X. (2026). Developmental, contextual, and profile-based heterogeneity in K-12 students' artificial intelligence literacy. Behavioral Sciences, 16(11), Article 2017. doi.org

Arif, M., Mussa, A. S. M., al Mamun, A., Rahman, M., Rahman, M., & Ashiqussalehin, M. (2026). An automated online proctoring system with live suspicious activity detection using YOLOv11 pose estimation and multi-model classification. Computers and Education: Artificial Intelligence. doi.org

Research, Preprint (Not Peer-Reviewed)

Abdelghani, R., Kaiser, P., & Murayama, K. (2026). Student-LLM interactions during math learning: Limited epistemic proactivity overall, but within-chat change predicts performance without AI. arXiv:2610.38950. arxiv.org

Institutional Report (Not Peer-Reviewed)

State Educational Technology Directors Association. (2026, October 29). New report: AI remains top state edtech priority amid growing debate about technology in schools [Press release]. PR Newswire. prnewswire.com The SETDA report page: setda.org

Shah, A. (2026). Unstructured AI access and exam incentives in a charter high school (AEARCTR-0019880) [Trial registration]. AEA RCT Registry. socialscienceregistry.org