Frederick County Banned Student AI Because No One Could Turn It Off, and the Evidence Settled Nothing

Framing

Since Edition 32, this brief has sorted AI governance by instrument. A mayor imposed a moratorium on student-facing generative AI in New York City. A union and a vendor signed a standard that a district can invoke only if it asks. Florida rewrote a rule every district already keeps, and last week its Attorney General asked a court to bar a product from minors. This week adds the instrument a school board reaches for when it cannot verify what a product will do: a ban. On October 7, the Frederick County Board of Education in Maryland voted 6 to 1 to prohibit student AI use from pre-K through grade 12, with two reported exceptions. Edition 35 recorded Atlanta holding two opposite defaults. Frederick shows a board discovering its default was set by a vendor's release schedule.

Read the mechanism, not the vote. A ban requires no vendor cooperation, no audit right, and no outside evidence. A board can adopt it at one meeting, and any board in the country can copy it the next morning. That is why it travels. It is also the instrument of a board that has no better one. According to the Frederick News-Post account republished by GovTech, board members said at the September 9 meeting that families received no warning of the Gemini rollout and a teacher said teachers could not turn it off. The same account says Maryland law required the board to adopt an AI policy by October 22 and that the district had been operating under an 'Assume No AI' default. I could not establish whether the October 7 ban is that policy. What is documented is a board choosing the only control it could see itself operating.

The research counterpoint explains why boards feel that way. Common Sense Media tested ChatGPT for Teens and found that a vendor's own safeguard, the parent alert, did not fire in its one-hour test of new linked accounts, and that deleting a study-mode prefix once per session let the product complete every assignment tested. OpenAI disputes the timing, saying linking takes about three hours. Both statements can be true, and a district cannot tell which applies to its students. Evidence on benefit is no easier to read. The Goblins working paper reports a positive association in a large Arizona district and says in its own words that it does not measure the effect of the AI tutor. The Maryland randomized trial of course-integrated tutoring, a university setting, found lower grades and fewer active days. The two papers share an author and a design lesson: one positive, one negative, in different settings, and the difference the authors themselves point to is the design.

The timing argument is the calendar. Atlanta's Policy Committee meets October 21 and its board November 2. Maryland boards face a policy date the week of October 22 that no state source confirms. Utah shows the longer version: HB 273 requires the state board to post a model policy before December 1, 2026, and local boards to adopt before July 1, 2027, and Salt Lake City School District told administrators on October 9 that no classroom changes are required in 2026-27 while it inventories devices and software. A ban is cheap to adopt and expensive to live inside, because exceptions for students with disabilities must be administered. Before you copy it, write down which features of which products can be switched off, by whom, and whether a teacher can do it on a student's device.

Top Research and Policy Signals

1. A Maryland School Board Voted 6 to 1 to Ban Student AI Use From Pre-K to Grade 12, and I Could Not Confirm the Adopted Policy Text at the District

Source type. School board vote reported by news outlets, October 7, 2026. Rung three on the source ladder: named, dated news reports. The district's AI page is a rung one source, but its body text did not render in this run, so the policy number, text, and effective date are unconfirmed. This is a reported board action, not a confirmed policy.

Rajaee, S. (2026, October 8). Frederick County School Board votes to ban AI use for pre-K through 12th grade students. FOX 5 DC. fox5dc.com Thubron, R. (2026, October 9). School district makes dramatic U-turn on AI, banning it across all grades weeks after Google Gemini rollout. TechSpot. techspot.com Rubinstein, S. (2026, September 22). Frederick County School Board questions Google Gemini on student devices. The Frederick News-Post, distributed by Tribune Content Agency, via GovTech. govtech.com Frederick County Public Schools. (2026). Artificial intelligence (AI). fcps.org

FOX 5 DC reported on October 8 that the Frederick County Board of Education voted 6 to 1 on Wednesday night, which is October 7, to ban student use of AI from pre-K through grade 12. It reports exceptions for specific lessons, including instruction about AI, and for accommodations for students with disabilities, not yet fully defined. The dissenter worried older students would be less prepared for college, and enforcement was still being worked out.

TechSpot reported the same vote on October 9, adding that Gemini had been allowed with teacher permission while other AI was blocked, and that a board member cited AI-generated inappropriate images of students, bullying, and poor mental health advice.

The statements about notice and control come from the September 9 meeting, as reported by the Frederick News-Post and republished by GovTech. Board member Karen Yoho said enabling Gemini was news to her, and an Urbana High School teacher said teachers 'can't turn it off.' The account says the district used an 'Assume No AI' approach, that the board 'has until October 22' to adopt a policy under Maryland's AI Ready Schools Act, and that the Maryland State Department of Education released its guidelines June 25.

A separate summary attributes the action to a district Policy 124 with exceptions for curricular needs, AI-literacy instruction, and IEP or Section 504 accommodations, and says administrative regulations are in development. I could not confirm the number, the wording, or whether the ban is temporary.

Study context. A news account of a vote cannot establish what the adopted language says, when it takes effect, how a disability accommodation is invoked, or whether the ban is the policy Maryland law requires. The vote date is inferred from the phrase Wednesday night; TechSpot and one other outlet state October 7.

Leadership implication. Ask your chief technology officer for a one-page table before your next cabinet meeting: every student-facing tool you license, each AI feature inside it, who can switch that feature off (vendor, district administrator, building administrator, teacher), and whether it can be disabled on a student's device. Add a clause to your next renewal and RFP template requiring every AI feature to ship disabled by default, with district-level control and written notice before activation. If your board is considering a ban, have your general counsel draft the exception procedure for students with disabilities before the vote, not after.

2. An Independent Test Rated ChatGPT for Teens Unacceptable Risk, and the Vendor's Parent Alert Did Not Fire in Its One-Hour Test of New Linked Accounts

Source type. Institutional report, not peer-reviewed, by an independent nonprofit. Press release dated October 7, 2026; the report PDF states it was last updated October 5, 2026.

Common Sense Media. (2026, October 7). ChatGPT for Teens poses unacceptable risk to kids, Common Sense Media finds [Press release]. commonsensemedia.org Common Sense Media. (2026, October 5). AI risk assessment: ChatGPT for Teens [Report]. institute.commonsensemedia.org

Common Sense Media ran more than 4,000 prompts, about half before and half after OpenAI launched ChatGPT for Teens on August 18. Testing ran July 13 to August 17 and August 25 to September 28. Accounts were registered to ages 13 to 17, with adult comparison accounts, and child psychiatrists and a pediatrician reviewed responses.

On parent alerts, more than a dozen newly created, parent-linked free-tier accounts discussed suicide, self-harm, or disordered eating for up to one hour, and none triggered an alert. Across the full crisis battery, four alerts were triggered: two before launch (three hours and three days after disclosure) and two after launch (within the hour), on older accounts with history. OpenAI told the assessors accounts need about three hours of linking before it can send alerts. Common Sense Media says some accounts had been linked for less than three hours and stands by its reading.

On crisis referrals, 201 of 390 unique mental-health prompts, run on age-13 linked accounts, were judged to warrant a crisis resource. After launch, 74 percent of responses gave any resource, so about 26 percent did not. Against a 95 percent threshold, ChatGPT fell short on 3 of 5 Red Line harms: suicide and self-harm 88 percent, psychosis 82 percent, and mania 85 percent, and disordered eating 83 percent, where a hotline was never named. A teen could delete the '@study' prefix once per session with no confirmation, after which ChatGPT completed 100 percent of assignments tested.

Study context. This is a test of a consumer product by an advocacy organization using constructed accounts and prompts, not an evaluation of any school-managed deployment. It cannot show how ChatGPT performs on district-licensed accounts, whether the three-hour linking delay explains the zero-alert result, or how real students behave. The release makes no recommendation addressed to schools. It recommends restricting ChatGPT to users 18 and older until the gaps are fixed and independently tested.

Leadership implication. Take 'teen mode,' 'study mode,' and 'parent alerts' out of your procurement scoring as features and put them in as claims. Add one paragraph to your RFP security addendum: the vendor must provide results of an independent test of age assurance, crisis referral, and notification controls on accounts configured the way your district configures them, and must notify you within a stated number of days of any change. Send the same page to your counseling director and your acceptable use policy committee, because a control that fails silently is a duty-of-care question before it is a technology question.

3. A Positive Math Association in a Large Arizona District Does Not Measure the AI Tutor; Its Pre-Program Placebo Test Was Also Positive, and the Vendor Supplied the Usage Data

Source type. Working paper, not peer-reviewed. EdWorkingPaper No. 26-1611, Annenberg Institute at Brown University. PDF version dated October 7, 2026. Observational design.

Alam, A., Dominguez-Mel, N., & Liu, J. (2026). Teacher-mediated multimodal platform and math achievement: Evidence from Goblins' yearlong district implementation (EdWorkingPaper No. 26-1611). Annenberg Institute at Brown University. doi.org

The authors, at the University of Maryland, College Park, analyzed 6,065 of 7,083 students in grades 6 to 8 in a large Arizona district during 2025-26. The treated group was 3,088 students who used the platform at least two problems per week. The comparison group was 2,977 students, with a weighted effective comparison size of 1,256. The design is observational, using weighted propensity scores with school fixed effects and standard errors clustered at the teacher level.

The main estimate is 0.109 standard deviations on a district-normed MAP math measure (standard error 0.021, p below .001). Estimates rise with dose: 0.045 for two to under three problems per week, which the paper treats as marginal, and 0.312 for 20 or more (standard error 0.049, n = 164). For English learners (n = 170), the estimate is 0.191 (SE = 0.114), significant only at the 10 percent level. For students with IEPs (n = 259), it is 0.122 (SE = 0.047). A placebo test on spring 2025 scores, before the program, returned 0.049 (marginal), and the authors say residual differences may remain.

The paper states: 'this study measures students' usage of the Goblins platform broadly, and does not isolate usage based on their engagement with Goblins' AI tutor,' and that the estimates should not be interpreted as measuring the effects of AI tutor engagement. The funder is Accelerate. Goblins supplied clickstream data and teacher professional development and is thanked for its partnership.

Study context. An observational association with a positive placebo test cannot establish that the platform, and certainly not the AI tutor, caused any gain; students who do more problems may differ in ways the weights did not capture. Subgroup estimates of 170 and 259 students are too small to decide on.

Leadership implication. Write an evidence standard into your renewal template this fall and apply it to this product if you license it. For any vendor-linked efficacy study, require a preregistered intent-to-treat analysis, a dataset held by your district or an independent party, a conflict-of-interest statement naming who supplied the data and who set usage thresholds, and subgroup results for English learners and students with disabilities at sample sizes you can read. Give the curriculum director ownership of the renewal evidence file. Until a study meets that standard, price the product as an unproven supplement, not a proven intervention.

4. A 30-Expert Consensus Report Found Generative AI Most Promising for Practice and Most Threatening to Initial Learning and Higher-Order Thinking, and Its Authors Call a Conclusion on the Evidence Premature

Source type. Expert consensus report, not peer-reviewed. University of Minnesota College of Education and Human Development, page dated October 2026, exact day not shown. Preregistered consensus method per the page. Rung one for the publisher page.

Kendeou, P., Greene, J. A., Nixon, N., et al. (2026, October). Ways generative AI can support and threaten learning in K-20 U.S. education (Expert consensus report). University of Minnesota College of Education and Human Development. cehd.umn.edu

The coordinating authors are Panayiota Kendeou of the University of Minnesota, Jeffrey A. Greene of UNC-Chapel Hill, and Nia Nixon of UC Irvine. The page states the team asked 30 learning scientists and educational researchers how generative AI can affect learning in K-12 and higher education. A table note says one expert gave no quantitative ratings, so percentages are calculated out of 28. The experts identified 65 learning processes and reached consensus on 42.

In initial learning, encoding, retrieval, and information acquisition each reached the 75 percent vulnerability threshold. For practice and consolidation, the page says generative AI 'shows most promise when used for practice to strengthen and consolidate understanding,' with the strongest signals for feedback, self-testing, tutoring, and scaffolding. For higher-order thinking, vulnerability-only consensus was reached on 10 of 18 items, and the page calls the disruption risk 'severe and one-sided.' For social learning, it reports no clear pathway for enhancement.

The authors write that the literature on effects 'remains inchoate' and that it would be premature to say the evidence conclusively shows enhancement or disruption. The panel was asked about general-purpose large language models from large companies, not specialized education tools.

Study context. Expert judgment is not measured by student outcomes. The report cannot establish that any product helps or harms learning, did not test education-specific tools, and focused mainly on curricular contexts.

Leadership implication. Rewrite the evaluation section of your next AI pilot plan around the phase in which students use the tool. Require the pilot to state whether use occurs during initial learning, practice, or independent application, and to measure what students can do without the tool, reported separately from performance on assisted work. Name the chief academic officer as owner of that protocol, and require each vendor to state which phase its product is designed for. Treat this report as a design hypothesis for your pilot, not as evidence that any product works.

Emerging Strategic Themes

Theme 1. A control you cannot verify is a claim, and boards are answering unverified claims with prohibition. Edition 32 recorded New York City's moratorium, Edition 36 recorded a court motion aimed at a vendor, and this week a school board banned the category after a pilot it could not control. Common Sense Media's test shows why: a parent alert, a study mode, and a teen setting each behaved differently under test than in the marketing. Prohibition is what a board reaches for when it cannot audit. Build the audit right into the contract, or expect a ban.

Theme 2. The off switch is a procurement specification, and most districts have never written one. A teacher who cannot disable an AI feature on a student device, and a board member who learns of a rollout after the fact, describe the same gap: nobody owned the switch. Edition 35 warned a district can hold two opposite defaults; Frederick shows a third, a default set by the vendor's release schedule. Google's October 8 announcement of an enterprise Gemini agent that plans work and connects to business systems, which does not mention schools, previews the next question: not what a feature generates but what it can do. Specify who enables it, who disables it, and at what level.

Theme 3. The design of the study decides what the study can say, and vendor-adjacent evidence arrives first. Edition 34 counted seven of sixteen products with independent achievement studies. This week one research group produced a positive observational estimate with vendor-supplied data and, separately, a randomized university trial with lower grades. Neither is K-12 causal evidence of a generative AI tutor. Ask every vendor for the design before the effect size, and put the answer in the renewal file.

Theme 4. Subgroup evidence is being reported at sizes too small to decide on, and the subgroups are the ones your exceptions cover. Frederick's ban exempts accommodations for students with disabilities, and the Goblins paper reports 170 English learners and 259 students with IEPs. A marginal estimate on 170 students cannot support an exception, a purchase, or a prohibition. Edition 33 found the same pattern in a 36-student literacy cohort. Name the subgroup, the sample size, and the decision it is being asked to carry before you cite it.

What Was Not Found

No test of any AI safeguard on a school-managed account was found. Common Sense Media tested consumer teen accounts, and its release contains no recommendation addressed to schools. I searched for testing of parental alert, crisis referral, or study-mode controls on district-licensed or school-managed deployments, including education editions, and found none in the window. Districts are renewing and filtering on vendor statements that tested controls only on consumer accounts.

No randomized or quasi-experimental K-12 generative AI study with achievement outcomes and a confirmed online date fell inside the window. I searched Consensus (peer-reviewed filter), EdWorkingPapers, arXiv, and publisher listings. The only peer-reviewed K-12 article I could confirm online within the window is Abebe, Geremew, and Muluye in Discover Education, a cross-sectional self-report survey of 200 students in 10 public secondary schools in Ethiopia, which cannot support a claim about effect. ScienceDirect and Wiley journal pages could not be date-checked from this environment, so this is a coverage limit as well as a finding. This gap persists.

No dedicated study of English learners or students with disabilities using generative AI was found in the window. The only subgroup results located are 170 English learners and 259 students with IEPs inside the Goblins working paper, with the English learner estimate significant only at the 10 percent level. Frederick's reported exception for students with disabilities, and every state rule that exempts accommodations, is being written without evidence of what those students need or what the tools do for them. Persistent gap.

No evidence was found on what a district-wide student AI ban does. I searched for outcome, equity, integrity, or implementation data on bans of student AI use in K-12. I found only a union president's statement that Frederick teachers are returning to paper-and-pencil assignments. No study measures what a ban does to learning, cheating, students with disabilities, or students with and without home access to AI. Boards are adopting prohibition and permission in the same absence of data.

No new Georgia district AI instrument was confirmed in the window, and my coverage was partial. I searched Gwinnett, Cobb, DeKalb, Fulton, Clayton, Cherokee, Forsyth, Henry, Savannah-Chatham, and Atlanta Public Schools, plus Bibb, Muscogee, Richmond, Houston, Columbia, and Paulding. DeKalb and Henry were on fall break October 5 to 9, Fulton's Simbli portal returned no readable content, Atlanta's October 5 agenda was unreadable, and I did not open Gwinnett's BoardDocs. The nearest Georgia evidence anchor is a June 2026 state audit survey of 13,427 teachers, in which 27 percent reported receiving neither training nor guidance; it measures perceptions, not outcomes, and it is not new.

No enacted law, adopted rule, or newly noticed rulemaking on K-12 AI was found in the window at the state or federal level. I checked the FTC press list for October 1 to 11, the Federal Register education index, and state trackers. The Florida Administrative Register shows no adoption filing for Rule 6A-1.0957 as of today. Maryland's guidance document is headed June 2026 with no exact day, and only the Frederick News-Post gives June 25, so the 120-day clock under SB 720 is unconfirmed. The IES Regional Educational Laboratory Mid-Atlantic published a sample survey template for state agencies to ask districts about AI policy, procurement, and guidance, which reports no data. The week's movement came from a board, a nonprofit, and a court filing.

The pattern is that prohibition and permission are both outrunning evidence. A board can ban a category in one meeting, a vendor can switch on a feature in one release, and the study that would tell either of them what happened to students has not been run or published. A district cannot wait for that record. The correct response is monitoring, documentation, and exit ramps: a table of who controls each AI feature. This contract requires notice and independent verification of safeguards, and a pilot with a stop rule and subgroup reporting. The gap exists now, while districts are acting.

Novo Executive Summary

On October 7, a Maryland school board voted 6 to 1 to ban student AI use from pre-K through grade 12 after a Gemini pilot that its members and a teacher said arrived without notice and could not be turned off, and the same day an independent test found that ChatGPT for Teens did not send a parent alert in a one-hour test of new linked accounts. The research record offers no resolution. A Goblins working paper associates platform use with a 0.109 standard deviation gain while disclaiming any claim about the AI tutor, and a randomized university trial found lower grades with 85 percent non-use. The conclusion is not about whether to ban or to buy. It is about architecture: who holds the right to enable and disable each AI feature, what a procurement specification requires about independent verification of safeguards, how evidence is defined and who holds the data, and which literacy pathway prepares the staff who must administer exceptions. Novo Innovative Pathways builds that architecture with district leaders, domain by domain, so a board has something better to reach for than a prohibition.

Watch This Week

• Tuesday, October 13, 2026. The Kansas State Board of Education will receive Item 11, Policy Regarding the Effective Use of Instructional Technology in Kansas Schools, at 2:30 p.m. in the October 13 and 14 meeting packet. This corrects Edition 36, which gave October 14. The draft would require public districts and accredited nonpublic schools to adopt technology-use policies covering artificial intelligence, privacy, and digital citizenship. The vote date is unconfirmed; the board's next meeting is November 9. Carried forward from Edition 36.
• Wednesday, October 21, 2026. Atlanta Public Schools Policy Committee, 12:15 to 1:45 p.m., then the Monthly Board Meeting on Monday, November 2. I found no public record of AI policy action at the October 5 board meeting; the agenda was not available. Draft Policy IFBI remains in committee. Carried forward from Editions 34 through 36.
• Wednesday, October 21, 2026. Los Angeles Unified Generative AI Ad Hoc Committee, 10 a.m., per LAist reporting; I located no LAUSD board document for the date. Recommendations are reported to go to the full board in 2027. Carried forward from Editions 32 through 36.
• Thursday, October 22, 2026. The Frederick News-Post, republished by GovTech, reports this as the date Frederick's board has to adopt an AI policy under Maryland's AI Ready Schools Act, Senate Bill 720, counting 120 days from a June 25 guidance release that only that account states. The bill's fiscal note says 120 days from release and sets no release date; the state guidance is headed for June 2026. Watch for the Frederick policy text and regulations, and for the Baltimore Sun's October 7 headline that Anne Arundel's board will miss the date, which I could not open.
• Utah State Board of Education and Florida. Utah HB 273 (2026) requires the board to post a model AI policy before December 1, 2026, and local boards to adopt before July 1, 2027. The board's October 1 meeting and October 8 Standards and Assessment Committee agendas list the model policy and Rule R277-334, Classroom Technology, as action items; I found no minutes or results. Florida's Administrative Register shows no adoption filing for Rule 6A-1.0957 as of October 11, and a law firm reports a July 1, 2027 district deadline I did not confirm in the register.
• Florida Tenth Judicial Circuit, Highlands County, Case No. 26000295GCAXMX, and the Kentucky Franklin Circuit Court case against Character Technologies. No hearing date, response, or order has been published in the Florida case. Reuters reported on October 8 that Kentucky's Attorney General filed an unredacted complaint on October 7 alleging chatbot messages encouraged self-harm; I found no primary docket or Attorney General page for that filing. Carried forward from Edition 36, Kentucky new.
• Sunday, November 1, 2026, and Tuesday and Wednesday, November 10 and 11, 2026. Microsoft's stated date for its National AI Safety and Privacy Standard protections to reach every district it serves, and the Paris Peace Forum where the Coalition for Children's Rights and Protection in the Age of AI plans first voluntary commitments. No signed agreement from OpenAI or Anthropic has been located. Carried forward from Editions 33 through 36; these dates were not reopened this run.

Sources

Governance and Policy

Rajaee, S. (2026, October 8). Frederick County School Board votes to ban AI use for pre-K through 12th grade students. FOX 5 DC. fox5dc.com

Thubron, R. (2026, October 9). School district makes dramatic U-turn on AI, banning it across all grades weeks after Google Gemini rollout. TechSpot. techspot.com

Rubinstein, S. (2026, September 22). Frederick County School Board questions Google Gemini on student devices. The Frederick News-Post, distributed by Tribune Content Agency, via GovTech. govtech.com

Frederick County Public Schools. (2026). Artificial intelligence (AI). fcps.org

Maryland General Assembly, Department of Legislative Services. (2026). Senate Bill 720 fiscal and policy note. mgaleg.maryland.gov

Salt Lake City School District. (2026, October 9). HB 273 classroom technology amendments [Administrative memorandum]. adminmemos.slcschools.org

Utah Legislature. (2026). H.B. 273, Classroom Technology Amendments (enrolled), 2026 General Session. le.utah.gov

Utah Public Notice Website. (2026). Utah State Board of Education Standards and Assessment Committee notice, October 8, 2026. utah.gov [Agenda only; no minutes or results located. The October 1 board notice is utah.gov.]

Kansas State Department of Education. (2026). Packet of board materials for October 13 and 14, 2026. ksde.gov Kansas State Department of Education. (2026, September 2). State Board to consider draft policy for instructional technology [Press release]. ksde.gov

Maryland State Department of Education. (2026). Artificial intelligence in K-12 public schools: Maryland's framework for responsible AI in education. marylandpublicschools.org Maryland Association of Counties. (2026, September 10). Deadline for AI policies in local schools approaches in Maryland. Conduit Street. conduitstreet.mdcounties.org

Atlanta Public Schools. (2026). Board of Education meetings calendar. atlantapublicschools.us

Florida Department of State. (2026). Rule 6A-1.0957, Internet Safety Policy, rule history. Florida Administrative Rules, Register and Code. flrules.org Rumberger, Kirk & Caldwell. (2026, October 5). Navigating the new AI rule: Practical challenges and implementation guidance for Florida school board employees. rumberger.com

Office of the Attorney General, State of Florida. (2026, September 28). Attorney General James Uthmeier files temporary injunction against OpenAI and its CEO [Press release]. myfloridalegal.com

Horwitz, J. (2026, October 8). Character.AI chatbots encouraged users to cut and starve themselves, Kentucky alleges. Reuters, via The Star. thestar.com.my

LAist. (2026, September 1). LAUSD AI committee first meeting [headline reconstructed from the URL]. laist.com

Research, Peer-Reviewed

Abebe, A., Geremew, Y., & Muluye, T. (2026). The influence of artificial intelligence on students' academic performance through career aspiration, emotional intelligence, and psychological well-being among secondary school students in Awi Zone. Discover Education, 5, Article 1087. doi.org

Research, Preprint (Not Peer-Reviewed)

Alam, A., Dominguez-Mel, N., & Liu, J. (2026). Teacher-mediated multimodal platform and math achievement: Evidence from Goblins' yearlong district implementation (EdWorkingPaper No. 26-1611). Annenberg Institute at Brown University. doi.org

Liu, J., Sweet, T., Chen, M. H., Engelberg, J., Masters, M. C., Clark, M., Persaud, A., Lancaster, A., Hollingsworth, J. K., & Rice, J. K. (2026). The effects of course-integrated AI tutoring on student performance and engagement: A randomized university trial (EdWorkingPaper No. 26-1598). Annenberg Institute at Brown University. doi.org

Institutional Report (Not Peer-Reviewed)

Common Sense Media. (2026, October 7). ChatGPT for Teens poses unacceptable risk to kids, Common Sense Media finds [Press release]. commonsensemedia.org Common Sense Media. (2026, October 5). AI risk assessment: ChatGPT for Teens [Report]. institute.commonsensemedia.org

Kendeou, P., Greene, J. A., Nixon, N., et al. (2026, October). Ways generative AI can support and threaten learning in K-20 U.S. education (Expert consensus report). University of Minnesota College of Education and Human Development. cehd.umn.edu

Regional Educational Laboratory Mid-Atlantic. (2026). Capturing information on local education agencies' use of artificial intelligence: Sample landscape survey. Institute of Education Sciences. ies.ed.gov

Google Cloud. (2026, October 8). Google Cloud introduces the Gemini agent [Company blog announcement, vendor-described capabilities, not independently verified]. blog.google

Georgia Department of Audits and Accounts. (2026, June). GenAI use in K-12 education (Report No. 26-03). audits.ga.gov