Eight States Sue OpenAI and 40 Attorneys General Guard Children's Chat Data in a Bankruptcy Sale

Framing

Since Edition 32, this brief has sorted AI governance by instrument. A mayor imposed a moratorium. A union and a vendor wrote a standard that a district can invoke only if it asks. Florida amended a rule every district already keeps. In Edition 36, one Attorney General asked a Florida court to restrict a vendor's product for minors. In Edition 37, a Maryland school board banned the category because it could not verify what a vendor's control would do. This week the instrument changes again. Eight attorneys general filed their own suits against the same vendor that day, and 40 more asked a bankruptcy court to attach child-protection conditions to the sale of another vendor's data. The actors are no longer a single state. They are a coalition using existing law.

Read the mechanism, not the headline. A consumer protection claim does not need an education statute, a new agency, or a district to have done anything wrong. It needs an allegation that a company told parents something untrue about a product their children used. Every state has such a statute, which is why one complaint template can be filed in eight courts in a day. In the petitions I could read for Iowa, Kansas, Montana, and West Virginia, several sources and phrases recur, including a Brookings study on academic performance and a line about students hiding intoxication at school. The bankruptcy filing works differently and travels just as well. A notice of interest costs a state little and puts conditions on the record before it chooses a buyer. Both instruments reach your vendor without passing through your board.

The research counterpoint explains why courts, not districts, are being asked. A district that wants to verify a vendor's claim has little to go on. Cohn and colleagues show that when three models from different vendors agree about why a student erred, the agreement says little about whether the diagnosis is correct. Saha, Yu, and Wang reviewed 100 empirical studies of youth AI safety and found that safeguards are proposed far more often than they are implemented or evaluated, and that evaluations often measure technical performance rather than protection from harm. The one peer-reviewed K-12 result this week is encouraging and carefully hedged: third graders in Taiwan did better on computational thinking with a Socratic tutor, and the authors say the study cannot separate which part of the bundle did the work.

The timing argument is the calendar. The bankruptcy court's objection deadline on bidding procedures is October 23, and the hearing is October 30, per a docket summary I could not check against the docket itself. Florida's amended Rule 6A-1.0957 takes effect October 29, with the district deadline of July 1, 2027 that Edition 34 reported. Colorado's attorney general takes comment on rules that would make a school a deployer of automated decision tools on October 26. Waiting costs you the one thing you can still do cheaply: read your own contracts before a court, a regulator, or a buyer reads them for you. Pull every agreement under which a vendor holds student conversations, and find the clause that says what happens to that data if the vendor is sold, fails, or changes its terms.

Top Research and Policy Signals

1. Eight State Attorneys General Each Filed a Consumer Protection Suit Against OpenAI on the Same Day, and the Complaints I Could Read Cite Students but None Reach School-Purchased Accounts

Source type. Eight state court complaints and state attorney general press releases, October 14, 2026. Complaints are allegations, not findings. Mixed rungs: attorney general releases for Iowa, Arizona, Idaho, and Montana at rung one; hosted complaint copies for Iowa, Kansas, Montana, and Arkansas at rung two; Nebraska and West Virginia complaints from the attorney general sites at rung two.

Citation (APA). Iowa Attorney General. (2026, October 14). Attorney General Brenna Bird announces lawsuit against OpenAI for deceiving Iowa parents and Iowans [Press release]. iowaattorneygeneral.gov Arizona Attorney General. (2026, October 14). Attorney General Mayes announces lawsuit against OpenAI for deceiving Arizonans [Press release]. azag.gov Idaho Attorney General. (2026, October 14). AG Labrador announces lawsuit against OpenAI for deceiving Idaho parents about the safety of ChatGPT [Press release]. ag.idaho.gov Montana Department of Justice. (2026, October 14). Attorney General Knudsen files lawsuit against OpenAI for endangering children and teens [Press release]. dojmt.gov State of Nebraska. (2026, October 14). Complaint, District Court of Buffalo County, Case No. D09CI260000488. ago.nebraska.gov State of West Virginia. (2026, October 14). Complaint, Circuit Court of Mason County, Case No. CC-26-2026-C-70. ago.wv.gov Kansas, Iowa, Montana, and Arkansas complaints, as hosted by Just Security (Nagle, M., 2026, October 15).

On October 14, 2026, the attorneys general of Arizona, Arkansas, Idaho, Iowa, Kansas, Montana, Nebraska, and West Virginia each filed a suit against OpenAI in their own state court. They were separate actions, not a joint suit. The Associated Press reported that at least eight states were suing.

Courts and statutes, as far as I could read them: Iowa, Polk County (Iowa Code 714.16 and 715D.4(2)); Arizona, Maricopa County Superior Court (Consumer Fraud Act); Idaho, Ada County (Consumer Protection Act); Montana, Lewis and Clark County (Consumer Protection Act and Consumer Data Privacy Act); Nebraska, Buffalo County, Case No. D09CI260000488; Kansas, Shawnee County, Case No. SN-2026-CV-000981; Arkansas, Faulkner County, Case No. 23CV-26-2984; West Virginia, Mason County, Case No. CC-26-2026-C-70. The Iowa and Montana copies had no case number.

Arkansas also pleads negligence, strict liability, and public nuisance and asks for an injunction on collecting minors' data. Nebraska seeks civil penalties, an injunction, restitution, and disgorgement without stating amounts. Iowa's Brenna Bird said OpenAI is "hurting kids and they are lying to Iowans." The Arkansas release also names Utah as a filer. No other source I opened does, so this brief does not count Utah. Where the complaints mention schools, it is as context: student reasoning, academic performance, and dropping out.

Study context. These are allegations that have not been tested. I read four complaints as hosted copies, two on attorney general sites, and the others as releases only, and I did not find a Kansas attorney general release. Nothing I opened mentions ChatGPT Edu or a district-managed account, so whether any theory reaches school-purchased use is an open question and my inference, not a finding.

Leadership implication. Ask your general counsel and your chief technology officer for a one-page list of every OpenAI product in use in the district: consumer accounts students use on their own, any district-managed accounts, and any vendor that embeds OpenAI models in a tool you bought. Add a notice clause to your next renewal requiring the vendor to tell you within 30 days of any state or federal enforcement action involving minors, and bring the list to your October cabinet meeting.

2. Forty Attorneys General Told a Bankruptcy Court That Character.AI's Under-18 Chat Restriction Is Voluntary and Could Be Rolled Back by a Buyer, and Asked for Nine Conditions on Any Sale

Source type. Court filing (Notice of Interest of the States) in a pending federal bankruptcy case, dated October 14, 2026, and a state attorney general press release. Rung two for the filing, rung one for the New York release. Sale terms and dates are rung three and unverified against the docket.

Citation (APA). Attorneys General of 39 states and the District of Columbia. (2026, October 14). Notice of Interest of the States, In re Character Technologies, Inc., Case No. 26-11603-BLS (Bankr. D. Del.). content.govdelivery.com New York State Attorney General. (2026, October 14). Attorney General James urges court to protect children and consumers in Character.AI bankruptcy [Press release]. ag.ny.gov [Headline wording of the New York release taken from the URL slug; page opened.] Sale terms: CBS News. (2026, October 13). Character.AI files for bankruptcy as Disney licenses its technology. cbsnews.com

Character Technologies, the company behind the Character.AI chatbot app, filed a Chapter 11 petition in the U.S. Bankruptcy Court for the District of Delaware. The petition date, October 12, comes from news reports, not the docket. On October 14, 40 jurisdictions, 39 states and the District of Columbia, filed a Notice of Interest in Case No. 26-11603-BLS. New York Attorney General Letitia James said the company "should not be sold without strong protections for the millions of people whose private conversations it holds."

Paragraph 24 says the under-18 restriction on open-ended chat applies only to users the company accurately identifies as minors, can be circumvented, does not reach a feature called Stories, and "is voluntary and could be rolled back at any time, including by a purchaser." Paragraph 25 lists nine conditions for any buyer, among them age assurance that does not rely solely on a self-declared birthdate, crisis safeguards, no model training on chat content beyond what users validly consented to, and deletion rights before any data sale.

News reports describe Disney as the lead bidder with a $20 million offer for a license to certain intellectual property, not the app or the user data. I found no source saying how user data will be handled in a sale. The states' conditions are the only stated proposal.

Study context. Character.AI is a consumer chatbot, and no filing or release I opened discusses schools. This is a pending proceeding; the states filed a notice, not a ruling, and the Notice reserves the right to seek a consumer privacy ombudsman. The Notice includes the case number. The judge's name was not confirmed.

Leadership implication. Open the data processing agreement of every student-facing vendor and find three clauses: what the vendor may do with conversation content, what happens to your data on a sale, merger, or insolvency, and your deletion right and its deadline. If a clause is missing, add a rider at renewal that bars transfer of student data to a successor unless the successor assumes the contract in writing, and calendar October 30 to read the court's order on bidding procedures.

3. Third Graders in Taiwan Scored Higher on Computational Thinking After a Socratic AI Tutor, Motivation Did Not Differ After Correction, and the Authors Say the Bundle Cannot Be Taken Apart

Source type. Peer-reviewed journal article, Journal of Educational Computing Research, first published online October 10, 2026 (OnlineFirst; volume and issue not assigned).

Citation (APA). Huang, Y.-C., & Li, C.-H. (2026). Socratic, artifact-gated generative AI scaffolding in a 6E inquiry unit: A clustered quasi-experiment in elementary students' computational thinking. Journal of Educational Computing Research. Advance online publication. doi.org

Huang and Li studied 229 third graders in 10 intact classes in Taiwan over a six-lesson computational thinking unit. Both conditions used the same curriculum, tasks, teacher, platform, and instructional time. They differed in a bundled scaffolding configuration, which the SAGE summary lists as source, delivery modality, initiation, privacy, and artifact requirements. The treatment was a generative AI tutor the authors call TALPer, which gives student-initiated, artifact-contingent Socratic prompts.

In a two-level model controlling for pretest computational thinking, the adjusted posttest favored the AI tutor condition: B = 4.05, SE = 1.58, 95% CI [0.94, 7.15], p = .011, Hedges' g = 0.52. A sensitivity analysis gave p = .034. The condition-by-pretest interaction was not significant, so the authors did not find that the effect depended on prior skill.

On motivation, measured with the Intrinsic Motivation Inventory, no subscale differed after Holm correction for multiple comparisons. Perceived Choice had the largest positive estimate (p = .061, Holm-adjusted p = .306). The authors also descriptively coded 101 matched student responses about AI-use orientations, with no inferential test. They conclude that the findings "provide preliminary evidence concerning the bundled intervention but do not isolate component contributions." Pupil-level data are available only on request because they involve minors.

Study context. One country, one grade, one unit, 10 classes. The abstract does not say how classes were assigned or what the comparison classes received, and I did not read the full text. A cluster-level quasi-experiment with 10 classes cannot establish that the tutor, rather than the other bundled features, caused the difference.

Leadership implication. Use this as a model for what to ask a vendor, not as a warrant to buy. In your next evaluation plan or RFP scoring rubric, require that any claimed learning effect include the design (random assignment or not), the number of classes (not just students), the comparison condition in plain words, and results on pre-specified outcomes, including the ones that came out null. Give your curriculum director a one-page template with those four fields before the next pilot is approved.

4. Three AI Models Agreed With Each Other Far More Than With Human Labelers on Why K-12 Students Missed Math Problems, and the Authors Say Agreement Is Not Validity

Source type. Preprint, not peer-reviewed. arXiv:2610.08703, cs.CL. No venue or acceptance stated.

Citation (APA). Cohn, C., Fonteles, J., Vanacore, K., Mazza, G., Crawford, C., Hooper, T., Biswas, G., & Kizilcec, R. (2026). Agreement is not validity: Cross-model LLM consensus in diagnosing student failure modes in K-12 math tutoring dialogue. arXiv. arxiv.org

Cohn and eight co-authors asked three large language models from different vendors, plus a fourth used as a robustness check, to label why a student went wrong in a K-12 math tutoring exchange. The data came from the Third Space Learning tutoring platform, which serves students in the United States and United Kingdom: 22,821 sessions between November 2024 and July 2025, 35,072 graded responses, and a pool of 3,854 items. The authors drew 100 adjudicated tutoring moments and used five failure categories.

Agreement among the models was high. Model-to-model kappa ranged from .755 to .781, and Krippendorff alpha was .769. Agreement between each model and the human reference was lower: kappa of .524 (GPT), .559 (Claude), and .597 (Gemini). A majority vote across the models reached a kappa of .568, no better than the best single model.

The most telling category was procedural slip, a careless error rather than a conceptual gap. The models recalled 22 percent of true procedural slips, and 61 percent of them were labeled conceptual gaps. That category has only six items, so the authors present it as a pattern and not an estimate. Their stated conclusion is that "cross-model agreement can create a misleading appearance of correctness." The abstract does not break out the tutoring platform or the K-12 grade levels.

Study context. 100 deliberately sampled moments from one commercial platform; transcripts only; three tutors interviewed and no students; no learning outcomes; some categories as small as six items. It tests diagnosis labels, not whether students learned.

Leadership implication. Add a validation requirement to your AI procurement standard: any vendor that claims its system diagnoses student errors, misconceptions, or needs must show agreement with trained human raters on a sample from your own student population, and must say whether a second AI model ran the check. Assign your assessment director to ask this question for the next renewal that includes adaptive or diagnostic features, and add the answer to the file before signing the contract.

5. A Review of 100 Youth AI Safety Studies Found Safeguards Are Proposed Far More Often Than Implemented or Evaluated, and Evaluations Often Measure Technical Performance Rather Than Protection From Harm

Source type. Preprint, not peer-reviewed. arXiv, cs.HC, cs.AI, cs.LG; version one dated October 6, 2026. Systematic review (Systemization of Knowledge). No venue stated.

Citation (APA). Saha, P., Yu, Y., & Wang, Y. (2026). Systemization of knowledge (SoK): Human-centered AI safety for youth. arXiv:2610.08554. arxiv.org

Saha, Yu, and Wang reviewed 100 empirical human-computer interaction studies published from 2019 to 2026 on AI risks and safeguards for children and young people, screened from 735 records after removing 15 duplicates. Seventy-one of the 100 studies came from the CHI conference. Schools are one setting among several, along with homes, care settings, and public services, and the abstract does not say what share of studies are school-based.

The authors traced each identified risk to a safeguard and to the furthest stage that safeguard reached. In their Table 5, of 80 risk-to-safeguard links, 50 percent were proposed only, 31 percent were implemented, and 18 percent were evaluated. Counting by risk, 31 percent had a safeguard that was only proposed, 25 percent had one implemented, 28 percent had one evaluated, and 14 percent had no safeguard. The text gives 34 risks in one place and 36 in another, so the by-risk percentages carry that uncertainty. Secondary use of children's data was the one risk with no proposed safeguard.

The authors write that "few countermeasures have been implemented, and fewer still evaluated," and that "existing evaluations often measure technical performance rather than protection from harm." In their words, "the existence of a safeguard says little about whether it protects young people." The abstract also says most risks are matched only with proposed or ideated countermeasures; the table supports that for links (50 percent) more clearly than for risks (31 percent).

Study context. A review of human-computer interaction research, mostly from one conference, not a K-12 school sample. It describes the research literature, not any vendor's actual safeguards. The authors' limitations section was not readable in the version I opened.

Leadership implication. When a vendor presents a safeguard (a filter, a parent alert, an age gate, a crisis referral), ask your technology and student services leaders to sort it into three bins in your vendor file: described, implemented in the build you license, or independently evaluated. Do not accept a described safeguard as a contract term. Require that the specification name the safeguard, the version, and who tests it, and list those requirements in the security addendum you attach to every student-facing AI renewal.

Emerging Strategic Themes

Theme 1. The regulator of your vendor may be a consumer fraud statute, not an education code. Edition 36 recorded one Attorney General's motion in Florida. This week eight states filed their own suits on the same day, citing statutes that exist in every state, and 40 attorneys general used a bankruptcy notice to set conditions. None of these instruments ask whether your district approved the tool. Review which vendors you use that a state could name in a complaint, and write the notice clause that tells you first.

Theme 2. Data disposition is the unwritten clause. The Character.AI notice treats a child's chat history as an asset that can be sold, and asks for deletion rights before that happens. A vendor contract that addresses collection and use but is silent on a sale or failure leaves that question to a bankruptcy judge. Add successor, insolvency, and deletion-deadline language to your standard rider before your next renewal, not after a vendor's lawyers file.

Theme 3. Agreement is not validation. Edition 34 counted seven of sixteen products with independent achievement studies. This week, Cohn and colleagues showed that models can agree with each other and still be wrong about the same student in the same way, and Saha and colleagues found safeguards rarely evaluated. A second AI check is still an AI check. Require human-rater agreement and independent evaluation in your specifications, and make the vendor say which was done.

Theme 4. Time saved is the metric vendors can win, and it is not learning. Gallup reports that 64 percent of AI-using educators say they gained time, all self-reported. The evidence base on student outcomes remains thin: Stanford SCALE reviewed 818 papers and found 20 with strong causal evidence. Treat teacher time as a workload benefit that needs its own measure, and do not let it stand in for an outcome measure when a renewal comes up.

What Was Not Found

No complaint, release, or filing I opened addresses school-purchased AI accounts. I opened a complaint or a release for each of the eight states and the Notice of Interest in the Character.AI case, and searched them for schools, students, education, and ChatGPT Edu. Several complaints mention students and school performance as context, and none mentions ChatGPT Edu or a district-managed account. Districts are deciding whether the consumer-minor theories apply to their contracts, without any statement from a court or attorney general.

No randomized K-12 trial of a generative AI tool with student achievement outcomes appeared in the window. Among the sources I could read, the strongest causal design was Huang and Li's quasi-experiment: 10 classes, one unit, third grade, computational thinking. Searches of the education indexes and the publisher pages I could reach (several large publishers blocked access) found nothing stronger dated October 5 to 18. Districts are renewing tools on evidence of this size.

No peer-reviewed K-12 study of AI text detectors, remote proctoring, weapons detection, or emotion and engagement recognition in schools appeared in the window. I ran the counter-evidence sweep for these systems. I found one adjacent journal article on engagement that I could not read past the metadata, plus higher education and non-peer-reviewed material. These systems report detection accuracy and sell it. A district cannot find independent evidence of what they do to student outcomes or privacy this week.

No outcome evidence in the window for English learners or students with disabilities. The only special education item I found was the introduction to a journal special issue, and I found no peer-reviewed K-12 study of English learners and AI dated in the window. This persistent gap carries over from earlier editions. New York City's moratorium (Edition 32) and Frederick's reported ban both carve out exceptions for these students, but no outcome evidence shows what those exceptions should include.

No Georgia district published an adopted AI policy or a first read on an agenda I could read this week. Lane 7 covered all ten Tier 1 districts. The Cobb County October 15 agenda was read in full and has no AI item. Fulton's October 14 bulletin lists no AI item. Henry's October 12 and Forsyth's October 13 agendas could not be read; Gwinnett's and Savannah-Chatham's portals blocked access, and the Atlanta Public Schools committee does not meet until October 21. This is both a coverage limit and a finding.

No independent measure of student learning sits behind the time savings educators report. Gallup's survey of 1,952 educators (fielded July 28 to August 14, margin of error plus or minus 2.7 points) reports 13 percent of users gained a lot more time and 51 percent somewhat more, and it contains no measure of instructional quality or student outcomes. The Alliance for Learning Innovation asks Congress to fund a stronger evidence base. Districts are scaling teacher-facing tools based on self-reported benefits.

The pattern is that enforcement and adoption are both outrunning evidence. States are filing suits on what a vendor told parents; courts have ruled on none of it, and the research on whether the safeguards work, or whether the tools raise learning, is thin and mostly vendor-adjacent. The correct response is not delay. It is monitoring, documentation, and exit ramps: a vendor file for each product, a notice clause in each contract, and a date by which each pilot either produces its own evidence or ends.

Novo Executive Summary

On October 14, eight state attorneys general filed separate consumer protection suits against OpenAI over ChatGPT and minors, and 40 attorneys general told a Delaware bankruptcy court that Character.AI's restriction on children's open-ended chat is voluntary and a buyer could reverse it. The research record explains why the question is landing in courts. A preprint shows three AI models agreeing with one another far more than with human labelers, a review of 100 youth AI safety studies finds safeguards proposed far more often than evaluated, and the one peer-reviewed K-12 result this week, a Taiwanese quasi-experiment in third grade, is promising and cannot separate its parts. The conclusion is not which vendor to trust. It is architecture: who holds decision rights over each AI feature and each child's data, what the procurement specification requires about independent evaluation and data on sale or failure, how evidence is defined before a pilot starts, and what training lets staff ask these questions. Novo Innovative Pathways works with district leaders to build that governance architecture, role-based AI literacy, and implementation strategy, so a district reads its own contracts before anyone else does.

Watch This Week

• Thursday, October 22, 2026, Maryland. This is the date WBOC reported Assistant State Superintendent Richard Kincaid gave on July 28 for local boards to approve AI policies under Senate Bill 720, which follows 120 days from the Department's guidance release; I could not confirm the release date at the Department. I found no Frederick County policy text, number, or effective date, and the Baltimore Sun's October 7 headline that Anne Arundel will miss the date was behind a robots block. Carried forward from Editions 35 through 37.
• Tuesday, October 20 and Thursday, October 29, 2026, Florida. The Florida Rules gateway shows new Rule 6A-14.0719, for Florida College System institutions, effective October 20, and amended Rule 6A-1.0957, Internet Safety Policy, effective October 29, notice ID 31488060. This resolves the register question carried from Editions 34 through 37; the K-12 district deadline of July 1, 2027, is as reported in Edition 34 and was not re-read this week. No hearing date, response, or order was found in the Highlands County case against OpenAI.
• Friday, October 23, and Friday, October 30, 2026, U.S. Bankruptcy Court, District of Delaware, In re Character Technologies, Inc., Case No. 26-11603-BLS. Per news reports, the objection deadline on bidding procedures is October 23, the hearing is October 30, the proposed bid deadline is November 20, and the outside closing date in the Disney agreement is December 17. These are rung three and not checked against the docket. Watch whether the court adopts any of the states' nine conditions.
• Monday, October 26, 2026, Colorado Attorney General. Comments close at 11:59 p.m. Mountain, and a hearing is scheduled for 10 a.m. on the interim draft rules under SB 26-189 and HB 26-1263, effective January 1, 2027. The draft calls a school the deployer of an automated decision tool in an example. I could not confirm whether the rules reach public K-12 districts.
• Wednesday, October 21 and Friday, November 13, 2026, Atlanta Public Schools and Los Angeles Unified. The Atlanta Policy Committee meets October 21, 12:15 to 1:45 p.m., with the board on November 2; draft Policy IFBI is still pending, and I could not confirm an agenda. Los Angeles Unified's Generative AI Ad Hoc Committee, listed for October 21, has been postponed to Friday, November 13, per the board's own event page, with no agenda posted. Carried forward from Editions 32 through 37.
• Monday, November 9 and Tuesday, December 1, 2026, Kansas and Utah. The Kansas State Board received Item 11, the instructional technology policy, on October 13 as a received item with no vote, and the KSDE release says a vote on the final version is expected in November; the packet lists the next meeting as November 9. Utah's board sent its AI model policy back to its Standards and Assessment Committee on October 1, per CitizenPortal; HB 273 requires the model policy before December 1.
• Sunday, November 1 and Tuesday and Wednesday, November 10 and 11, 2026. Microsoft's stated date for the National AI Safety and Privacy Standard to apply to its school contracts, and the Paris Peace Forum, where the Coalition for Children's Rights and Protection in the Age of AI plans its first voluntary commitments. I found no new signatory by OpenAI, Anthropic, or Google as of October 18. New York S9051B, the Kids Chatbot Safety Act, shows as passed by both houses with no delivery to the governor on the Senate page. Carried forward from Editions 33 through 37.

Sources

Governance and Policy

Iowa Attorney General. (2026, October 14). Attorney General Brenna Bird announces lawsuit against OpenAI for deceiving Iowa parents and Iowans [Press release]. iowaattorneygeneral.gov

Arizona Attorney General. (2026, October 14). Attorney General Mayes announces lawsuit against OpenAI for deceiving Arizonans [Press release]. azag.gov

Idaho Attorney General. (2026, October 14). AG Labrador announces lawsuit against OpenAI for deceiving Idaho parents about the safety of ChatGPT [Press release]. ag.idaho.gov

Montana Department of Justice. (2026, October 14). Attorney General Knudsen files lawsuit against OpenAI for endangering children and teens [Press release]. dojmt.gov

State of Nebraska. (2026, October 14). Complaint, District Court of Buffalo County, Case No. D09CI260000488. ago.nebraska.gov

State of West Virginia. (2026, October 14). Complaint, Circuit Court of Mason County, Case No. CC-26-2026-C-70. ago.wv.gov

Nagle, M. (2026, October 15). Complaints in the state attorney general suits against OpenAI [Document collection]. Just Security. justsecurity.org

Attorneys General of 39 states and the District of Columbia. (2026, October 14). Notice of Interest of the States, In re Character Technologies, Inc., Case No. 26-11603-BLS (Bankr. D. Del.). content.govdelivery.com

New York State Attorney General. (2026, October 14). Attorney General James urges court to protect children and consumers in Character.AI bankruptcy [Press release]. ag.ny.gov

CBS News. (2026, October 13). Character.AI files for bankruptcy as Disney licenses its technology. cbsnews.com

Florida Department of State. (2026). Rule 6A-1.0957, Internet Safety Policy; Rule 6A-14.0719, Artificial Intelligence (AI) Policies in Florida College System Institutions. flrules.org and flrules.org

Los Angeles Unified School District Board of Education. (2026). Postponed to 11/13/26: Generative AI Ad Hoc Committee [Event listing]. boe.lausd.org

Kansas State Department of Education. (2026). Packet of board materials for October 13 and 14, 2026. ksde.gov Kansas State Department of Education. (2026, October 8). State Board to receive 2026 spring assessment data, 2025-26 KSDE annual report during October meeting [Press release]. ksde.gov

Colorado Department of Law. (2026). Automated decision-making technology act and chatbot safety act rulemaking. coag.gov

Maryland General Assembly. (2026). Senate Bill 720, Chapter 634. mgaleg.maryland.gov Evangelista, A. (2026, July 28). [Maryland school systems AI policy deadlines]. WBOC.

Microsoft. (2026, September 9). AFT, UFT and Microsoft announce national AI safety and privacy standard for schools. news.microsoft.com

Elysee. (2026, September 22). Launch of the roadmap of the Coalition for Children's Rights and Protection in the Age of AI. elysee.fr

Utah State Board of Education. (2026). Standards and Assessment Committee notice, October 8, 2026. utah.gov citizenportal.ai

Research, Peer-Reviewed

Huang, Y.-C., & Li, C.-H. (2026). Socratic, artifact-gated generative AI scaffolding in a 6E inquiry unit: A clustered quasi-experiment in elementary students' computational thinking. Journal of Educational Computing Research. Advance online publication. doi.org

Research, Preprint (Not Peer-Reviewed)

Cohn, C., Fonteles, J., Vanacore, K., Mazza, G., Crawford, C., Hooper, T., Biswas, G., & Kizilcec, R. (2026). Agreement is not validity: Cross-model LLM consensus in diagnosing student failure modes in K-12 math tutoring dialogue. arXiv:2610.08703. arxiv.org

Saha, P., Yu, Y., & Wang, Y. (2026). Systemization of knowledge (SoK): Human-centered AI safety for youth. arXiv:2610.08554. arxiv.org [Not peer-reviewed.]

Institutional Report (Not Peer-Reviewed)

Brenan, M. (2026, October 14). Teachers using AI say it frees time for what matters. Gallup, in partnership with the Walton Family Foundation and The Mattering Institute. news.gallup.com

Alliance for Learning Innovation. (2026, October). Investing in R&D for safe and effective AI in education: Federal policy recommendations. alicoalition.org

Fesler, L., Martinez Claeys, J., Agnew, S., & Loeb, S. (2026). The evidence base on AI in K-12: A 2026 review. Stanford SCALE Initiative. [Institutional report, not peer-reviewed. Cited through the ALI report and the verification pass; reviews 818 papers in the repository as of October 2025, of which 20 had strong causal evidence.]