The Teacher and the Tutor

A student sits at a kitchen table with a math problem that will not yield. The teacher is home. The classmates are gone. A private tutor would be expensive, assuming one is available at all. The student opens an AI system and asks for help. The system does not get tired, does not become impatient, and can explain the same idea five different ways. It can slow down, translate, generate another example, or stay with the student until midnight.

For most of educational history, that level of individualized attention was scarce. Artificial intelligence can make it ordinary.

That possibility is genuinely exciting. It is also where the delegation problem becomes unusually subtle, because education is different from most of the work discussed earlier in this book. In a company, the purpose of a task is usually the output. In a classroom, the purpose of the task may be the capability formed while producing the output. A solved equation, finished essay, correct translation, or working program can therefore conceal failure if the student did not build the ability the assignment was designed to teach.

The assignment is not the product. The learning is.

One Tutor for Every Student

The promise of AI tutoring begins with a real shortage. Teachers work with groups. Human tutors cost money. Office hours end. A student who is embarrassed to ask a question in class may remain silent. A student who learns more slowly may need another explanation after everyone else has moved on. A student who learns quickly may need a harder problem rather than another repetition.

An AI tutor can adjust the pace, offer examples, ask questions, and give feedback when a student gets stuck. It can also help identify an earlier idea the student needs to understand before tackling the current problem. Software can make this kind of individual attention available to more students than human tutors alone can reach.

Recent experiments show that this is more than a theoretical possibility. A World Bank randomized trial in Nigeria used Microsoft Copilot, powered by GPT-4, in a six-week after-school program for first-year secondary students. The intervention combined AI tutoring with teacher support and produced gains of 0.31 standard deviations on a combined assessment of English, AI knowledge, and digital skills, with a 0.23 standard-deviation effect on English alone. The authors estimated the gains as comparable to roughly 1.5 to 2 years of business-as-usual schooling, although that comparison depends on the study context and should not be treated as a universal conversion.

At Harvard, a randomized crossover study in an undergraduate physics course found that students using a carefully designed AI tutor learned more in less time than students receiving an active-learning classroom lesson covering the same material. The AI-tutored students also reported greater engagement and motivation. The study involved one course, two lessons, and a custom system deliberately built around established teaching practices, so it does not prove that a generic chatbot is a superior teacher. It proves something more useful: structured AI tutoring can work extremely well when the pedagogy is designed into the system.

The opportunity is therefore not a fake one. AI can expand access to explanations, practice, and feedback. The harder question is what the tutor should refuse to do.

The Completion Trap

Most general-purpose AI systems are rewarded for being helpful. A student asks a question and the system tries to answer it. That design is sensible for many tasks. It can be destructive when the thing being optimized is learning.

A 2025 randomized field experiment involving nearly 1,000 high school mathematics students in Turkey makes the problem unusually clear. Students using a standard GPT-4-style interface performed 48% better on assisted practice problems than the control group. Students using a more structured AI tutor with teacher-designed safeguards performed 127% better during practice. Then the AI was removed and students took an unassisted exam. The standard-interface group scored 17% worse than students who had never received AI assistance. The guarded tutor largely eliminated that negative effect, although it did not produce a statistically significant exam advantage over the control group.

The result exposes what I will call the completion trap. If the system is judged by how many problems the student finishes correctly while the system is present, a tutor can look successful while weakening the very capability the lesson is meant to build. Correct answers during practice are not the same thing as learning.

The researchers found that students often used the standard system as a crutch, asking for and copying solutions, while the guarded tutor pushed them toward hints and more substantive engagement. The difference was not that one model knew mathematics and the other did not. The difference was the boundary around the help.

Principle 2 becomes especially important here. Capability is not permission. An AI system may be fully capable of solving the problem and still be the wrong party to solve it, because the student is supposed to acquire the capability. In education, a machine can be too helpful.

Productive Struggle

There is a temptation to romanticize difficulty once technology makes tasks easier. That would be a mistake. Chapter Six argued that suffering is not a curriculum, and the same rule applies in school. Confusion, boredom, and unnecessary frustration do not become virtuous merely because generations of students endured them.

Finishing without effort does not necessarily build understanding. Students need practice recalling an idea, explaining it, comparing approaches, making mistakes, and trying again. That useful mental work is productive struggle. A tutor should help students work through uncertainty while leaving enough of the thinking for students to do themselves.

The right educational boundary is therefore dynamic. Early in a problem, the tutor may ask a question instead of giving an answer. After a genuine attempt, it may supply a hint. If the student is missing prerequisite knowledge, it may stop the current problem and teach the missing concept. If the system is uncertain, it should say so. If the task is meant to measure independent capability, the correct tutoring behavior may be to stop helping altogether.

Chapter Ten described abstention as a form of competence in medicine. Education gives the idea another use. A capable tutor should know when not to help.

This is not a moral preference for struggle. It is a design requirement tied to the learning objective. The system should supply enough assistance to keep the student moving without doing so much that it removes the practice the student needs. A good tutor reduces unnecessary friction and preserves productive effort.

The Novice Cannot Be the Final Checker

AI literacy advice often tells students to verify what the model says. The advice is sensible and incomplete. A student is frequently using the tutor precisely because the student does not yet know enough to verify the answer.

A beginning algebra student may not recognize a subtle algebraic error. A student learning a foreign language may accept a fluent but unnatural translation. A novice programmer can run code without understanding why it works. A student researching an unfamiliar subject may have no basis for recognizing a fabricated source. The instruction to check the AI can become circular: first acquire the knowledge needed to verify the tool that is being used to acquire the knowledge.

The beginner cannot carry the whole burden of checking the tutor. The school must check the system before and during its use: choose suitable tools, set limits, test them against what students are meant to learn, and watch for mistakes. Teachers need to see enough of the exchanges to understand how the tutor is helping or misleading students.

The four reviewer conditions from Chapter Ten translate cleanly into the classroom. The teacher needs authority over the instructional use, enough evidence and context to understand what the system is doing, enough time and competence to challenge it, and practical power to override or remove it. A district that mandates an AI tutor while giving teachers no meaningful ability to inspect, limit, or stop it has recreated the fictional human in the loop, only this time at a desk instead of a hospital.

The student still has responsibilities. Students should learn to question sources, compare answers, expose uncertainty, and explain reasoning. But those habits are learning goals, not excuses for deploying unreliable systems and telling adolescents to catch the mistakes.

The Teacher Becomes the Architect

The rise of an artificial tutor changes where the teacher’s scarce attention can create the most value. Routine explanation, extra practice, and immediate feedback are precisely the activities that can sometimes be delegated. Choosing the learning objective, sequencing difficulty, reading the room, building trust, evaluating deeper understanding, and deciding when a student needs less help rather than more remain harder to automate responsibly.

There is also a second model of AI tutoring that deserves more attention: AI assisting the human tutor rather than replacing the tutor. Tutor CoPilot, developed by researchers at Stanford and partner institutions, provided real-time pedagogical suggestions to tutors working with students from underserved communities. In a randomized trial, students whose tutors had access to the system were more likely to master math topics, with the largest gains among students assigned to lower-rated tutors. The AI-supported tutors also used more probing questions and less generic praise.

That result matters because it shows a different path to scale. Instead of asking AI to become the teacher, an institution can use AI to distribute better teaching practices across a human workforce. The machine carries some expert guidance to the tutor, and the tutor remains inside the relationship with the student.

Neither model should become doctrine. A bounded AI tutor may outperform a human in one setting, while a human-AI team may be better in another. The governing question is which architecture produces actual learning, preserves the ability to intervene, and leaves someone answerable for the instructional choice.

Training Debt Can Begin in School

Chapter Six described training debt inside organizations. A company can remove the beginner work that once trained future experts, enjoy the productivity gain, and discover later that it no longer has enough people capable of checking difficult work. Education can create the same debt earlier.

Consider writing. If a student routinely delegates the first argument, the first outline, the first draft, the revision, and the fact-checking, the submitted essay may improve while the student’s ability to reason on a blank page develops more slowly. The same pattern can appear in mathematics, coding, foreign language, laboratory work, and research. Output quality can rise faster than underlying competence.

That is why the argument over AI in school cannot be reduced to cheating. Cheating matters, but the deeper risk is developmental. A student can follow every rule and still use a permitted tool in a way that eliminates too much of the practice needed to become competent.

The opposite risk also matters. Banning useful tutoring can withhold individualized help from students who cannot afford a human tutor or who need explanations a crowded classroom cannot provide. The World Bank result in Nigeria and the Tutor CoPilot result both warn against pretending that the safe policy is simply to keep AI out. A ban can preserve old inequalities as efficiently as bad deployment can create new ones.

The Nigeria trial also complicates the equity story. Students benefited across the baseline ability distribution, but the largest gains went to girls and to students who were already performing better. The same tool that widens access can therefore widen the gap between students who are better positioned to convert assistance into learning and those who are not. Chapter Nine’s conversion gap can appear inside a classroom as easily as inside a company.

The objective is to avoid training debt without preserving scarcity for its own sake. AI should let more students receive good explanations while ensuring that the student still performs the mental work that the curriculum is trying to build.

Assessment Has to Catch Up

Once students have powerful tutors outside the classroom, an old assumption begins to fail: the finished homework assignment is reliable evidence of the student’s independent capability. It may still show effort, curiosity, collaboration, or effective use of tools. It does not automatically show what the student can do alone.

Homework still matters, but the finished assignment can no longer be assumed to measure independent capability. Schools will need a richer mix of evidence: in-class writing, oral explanation, live problem solving, drafts that show development, projects with documented AI use, and assessments in which the tool is deliberately present or deliberately absent depending on the skill being measured.

The same AI use can be responsible in one assignment and destructive in another. If the goal is to teach research strategy, an AI system might help brainstorm search terms while the student verifies and synthesizes sources. If the goal is to measure whether the student can construct an argument independently, having the system draft the argument defeats the measurement. If the goal is professional tool use, forbidding AI may make the assignment less realistic.

AI policy is therefore delegation policy. It should tell students what may be delegated, what must remain theirs, what must be disclosed, and why. The rule should follow the learning objective rather than the mere presence of a chatbot.

The U.S. Department of Education’s August 2026 guidance on classroom technology points in a compatible direction. “Responsible design is the floor,” the Department says. The letter asks schools to focus on instructional value, independent evidence, educator judgment, transparency, and student outcomes. It also builds on a July 2025 federal AI-in-education position organized around five principles: educator-led, ethical, accessible, transparent, and protective of student data. Its five practical questions are unusually concrete: what learning problem does the tool solve, when should it be used, for whom, for how long, and what evidence shows that it improves learning? Those are not specifically a Delegation Doctrine test, but they are recognizably questions about boundary, evidence, and ownership.

Current Evidence: What Remains When the Tutor Leaves

Reviewed September 11, 2026. Independent-learning evidence added October 10, 2026.

Pew Research Center reported in February 2026 that 54% of U.S. teens had used chatbots for schoolwork. One in ten said chatbots helped with all or most of their schoolwork, while 21% said some and 23% said a little. About half of all teens described chatbots as at least somewhat helpful for schoolwork, and 59% said AI cheating was a regular occurrence at their school. The figures establish that AI tutoring and assistance are already part of ordinary student behavior, not a hypothetical future.

The PNAS high-school mathematics experiment provides the clearest warning about the completion trap. Standard GPT-4 assistance improved assisted practice performance but produced a 17% decline on the later unassisted exam relative to the control group. A teacher-designed tutor that gave hints rather than simply supplying answers largely removed the damage. The study covered one school and mathematics review sessions, but its design directly separates performance with the tool from learning after the tool is gone.

An October 9, 2026 Berkeley report highlighted a peer-reviewed study presented at COLM that recruited 1,222 participants for three randomized experiments. AI assistance improved performance while available, but participants performed worse on subsequent unaided arithmetic or reading tasks. Persistence findings were mixed: in the larger arithmetic experiment, the overall increase in skipped problems was not statistically significant. The tests followed brief assistance and measured immediate effects, not permanent loss of ability.

A school therefore needs to examine what a student can do after assistance ends, including whether the student keeps working when an answer is difficult. Better assisted results alone cannot establish learning. The teacher should preserve opportunities to attempt, explain and revise the work independently.

The Harvard physics experiment offers an encouraging example. Students using the custom AI tutor made larger gains on the study’s tests, typically in less time than the classroom lesson. The system broke problems into manageable steps, gave feedback when students needed it, and let them work at their own pace. Those teaching choices matter. The experiment covered two lessons in one course; it does not establish that an ordinary chatbot will teach every subject or student better.

Tutor CoPilot shows that AI can also improve the human side of tutoring. The 2025 randomized trial found a 4-percentage-point increase in topic mastery overall and a 9-point increase for students of lower-rated tutors. The system encouraged pedagogical strategies such as guiding questions. That is evidence for delegation to the teacher’s tool rather than delegation of the teacher’s role.

The World Bank trial in Nigeria shows why access matters. Six weeks of teacher-supported AI tutoring produced meaningful gains in English and related digital skills in a low-resource setting. The intervention used a publicly available large language model but did not simply hand students a chatbot. Teachers helped students formulate prompts, supervised sessions, and connected the AI work to the curriculum. The architecture around the model was part of the intervention.

A 2026 randomized classroom study of 371 students in Grades 7 through 9 adds another useful caution. Researchers tested custom generative-AI interventions aimed at motivation and learning strategies against standard ChatGPT. They found some benefit for perceived utility value but no statistically significant advantage for effort, domain knowledge, or elaboration strategy use. Customization by itself did not create broad learning gains. The educational mechanism still matters.

On October 16, 2026, the University of Maryland highlighted a working paper on a trial involving 2,379 undergraduates and 30 instructors. Comparisons between sections of the same course found lower final grades and less course-platform participation when an AI assistant was offered. The grade estimate across the full sample was inconclusive. Only about 15% of students offered the tool used it, and most instructors retained a mode that could supply answers directly. These findings concern that rollout; grades do not independently establish learning.

A school should therefore evaluate the whole learning environment when it introduces an assistant. That includes the students who do not use it, the practice it may replace, and the help available from teachers and peers. Before expanding access, the institution should show what students can do independently and explain who will change or suspend the tool when learning suffers.

Beginning November 1, 2026, Microsoft says US school districts can request its school AI safety and privacy protections in new or existing agreements without waiting for renewal. The published standard requires student features that take consequential actions outside the provider's systems to be off by default unless an authorized school administrator enables them, with limited safety and security exceptions. It covers defined educational AI products, not every general-purpose tool a school might use. These are contractual commitments, not a national law or proof that every district has adopted them.

The school must still establish which products its agreement covers, who may enable consequential actions, and whether the controls work in practice. An available safeguard is not yet an implemented boundary. Contract review and testing should establish that boundary before students receive authority to act through the system. Learning outcomes require a separate test.

On November 3, 2026, Google announced that access to Gemini, Gemini Notebook, Gemini in Classroom and Google Search would be off for K-12 student accounts until school administrators turned the products on. Google said it would give schools more details before the changes rolled out. The Associated Press reported that schools could choose among products, but that keeping Google Search would still mean accepting AI within Search.

An administrator can be responsible for what students use while the provider decides which choices are available. Before enabling a service, a school should check whether it can keep the research functions students need while limiting the kind of help an assignment permits. If those choices cannot be separated, the school should decide whether to use another research tool and make sure students can still find and evaluate sources.

Finally, the Department of Education’s August 2026 guidance says schools should use independent evidence, support educator training, review effectiveness regularly, and remove tools that repeatedly fail to improve learning. It states that “Responsible design is the floor,” builds on five federal AI principles announced in July 2025, and points to outcomes-based contracting in five states as a way to tie vendor payment to results. Whatever administration writes the guidance, that standard is hard to argue with. A school should not keep an AI system merely because it is modern, popular, or impressive in a demonstration.

Accountability Analysis

Principles 2, 4, and 8 do most of the work here, with Principle 3 defining the boundary around the assignment. Principle 2 asks whether the system’s capability has been mistaken for permission. If the student is meant to learn the skill, the AI’s ability to perform it is not enough reason to delegate the performance.

Principle 4 asks whether the people responsible for instruction can actually supervise the system. The four reviewer conditions from Chapter Ten apply: authority over the use, evidence and context, time and competence to disagree, and practical power to override. The teacher cannot be held responsible for an instructional system that the teacher is not allowed to govern.

Principle 8 reaches the student as well as the institution. Accountability decays when competence decays, and competence also fails to form when practice is repeatedly outsourced. If students become excellent at obtaining answers while remaining weak at producing or checking them, the school has created the educational version of training debt.

The accountability owner is therefore shared but not vague. Teachers and schools own the instructional boundary. Providers own the truthfulness, reliability, safeguards, and transparency of the system they sell. Students own the work the assignment explicitly reserves for them. Government and accrediting institutions own the rules they impose. Principle 6 requires consequences to reach the provider as well as the user. The Department of Education points to outcomes-based contracting in Louisiana, Arkansas, Indiana, Michigan, and Texas, where vendor payment can be tied to results. That is a concrete way to connect commercial reward to educational performance. The AI owns none of it.

Try it without the tutor

Choose one idea you have practiced with an AI tutor. Close the tool and explain the idea in your own words, then try a similar problem. Mark where you get stuck and ask your teacher or tutor for help with that step. The purpose is to discover what you can do independently and what still needs practice.

Accountability Verdict

Schools should use AI tutors where they demonstrably improve learning, especially where individualized support is scarce. A student who can receive a good explanation at midnight should not be denied it merely to preserve an older technology of scarcity.

Schools should also stop treating completion as the objective. The success measure is the student’s growing independent capability. That means guardrails that preserve practice, assessments that reveal what the student can do without the tool, and policies that define what may be delegated rather than pretending all AI use is the same.

The strongest educational AI may sometimes answer a question, sometimes ask one, sometimes offer a hint, and sometimes refuse to do the work. The teacher’s role is to define those boundaries and to revise them as evidence improves.

The test of a tutor is what remains when the tutor is gone.

Chapter Twelve moves from the classroom to the courtroom. There the person subject to an algorithmic judgment may not be learning a skill and may not have chosen the system at all. The question becomes harder: what happens when machine-assisted judgment can affect liberty, policing, or punishment?

Sources and Notes

1. Hamsa Bastani, Osbert Bastani, Alp Sungu, Haosen Ge, Özge Kabakcı, and Rei Mariman, “Generative AI without guardrails can harm learning: Evidence from high school mathematics,” Proceedings of the National Academy of Sciences, Volume 122, Issue 26, published June 25, 2025. https://www.pnas.org/doi/10.1073/pnas.2422633122. Accessed September 11, 2026. In a randomized field experiment with nearly 1,000 high-school mathematics students, GPT Base improved assisted practice grades by 48% and GPT Tutor by 127% relative to control; on the later unassisted exam, GPT Base students scored 17% below control while the guarded tutor largely eliminated the negative effect. The study used GPT-4. OpenAI systems are also used in this Living Edition’s production workflow; see the transparency statement. Source record: DA-CH11-S1. Archive reference: pending.

2. Greg Kestin, Kelly Miller, Anna Klales, Timothy Milbourne, and Gregorio Ponti, “AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting,” Scientific Reports 15, 17458, June 3, 2025. https://www.nature.com/articles/s41598-025-97652-6. Rechecked September 20, 2026. The randomized crossover study took place in a Harvard undergraduate physics course. Students using a custom, pedagogically structured AI tutor showed substantially larger learning gains than students in an active-learning class covering the same material and had a median time on task of 49 minutes compared with a 60-minute classroom lesson. The authors caution against assuming that unstructured AI use will produce the same results. Source record: DA-CH11-S2. Archive reference: pending.

3. Rose E. Wang, Ana Trindade Ribeiro, Carly D. Robinson, Susanna Loeb, and Dorottya Demszky, “Tutor CoPilot: A Human-AI Approach for Scaling Real-Time Expertise,” EdWorkingPaper 24-1054, Annenberg Institute at Brown University, November 2025 version. https://edworkingpapers.com/ai24-1054. Accessed September 11, 2026. In a randomized trial involving more than 700 tutors and 1,000 students from underserved communities, students whose tutors had access to Tutor CoPilot were 4 percentage points more likely to master math topics, with gains of 9 points for students of lower-rated tutors. Analysis of more than 350,000 messages found greater use of probing questions and less generic praise. This is a working paper rather than a final peer-reviewed journal article. Source record: DA-CH11-S3. Archive reference: pending.

4. Martín De Simone, Federico Tiberti, Maria Barron Rodriguez, Federico Manolio, Wuraola Mosuro, and Eliot Jolomi Dikoru, “From Chalkboards to Chatbots: Evaluating the Impact of Generative AI on Learning Outcomes in Nigeria,” World Bank Policy Research Working Paper 11125, 2025. https://doi.org/10.1596/1813-9450-11125. Accessed September 11, 2026. The randomized trial evaluated a six-week after-school program using Microsoft Copilot, powered by GPT-4, with teacher support for first-year senior secondary students. The reported effect was 0.31 standard deviations on a combined assessment and 0.23 on English. The authors estimate the learning gain as comparable to 1.5 to 2 years of business-as-usual schooling in that context. Effects were positive across the baseline ability distribution, with the largest gains among female students and those with higher initial academic performance. OpenAI technology underlies the model used in the intervention; OpenAI systems are also used in this Living Edition’s production workflow. Source record: DA-CH11-S4. Archive reference: pending.

5. Colleen McClain, Monica Anderson, Olivia Sidoti, and William Bishop, “How Teens Use and View AI,” Pew Research Center, February 24, 2026. https://www.pewresearch.org/internet/2026/02/24/how-teens-use-and-view-ai/. Accessed September 11, 2026. Pew surveyed 1,458 U.S. teens and their parents from September 25 to October 9, 2025. Fifty-four percent of teens reported using chatbots for schoolwork; 10% said chatbots helped with all or most of their schoolwork, 21% some, and 23% a little. Fifty-nine percent said AI cheating was a regular occurrence at their school. Source record: DA-CH11-S5. Archive reference: pending.

6. U.S. Department of Education, “U.S. Department of Education Releases Guidance on Responsible Use of Education Technology in the Classroom,” August 20, 2026. https://www.ed.gov/about/news/press-release/us-department-of-education-releases-guidance-responsible-use-of-education-technology-classroom. Accessed September 11, 2026. The Department’s Dear Colleague Letter emphasizes instructional value, independent evidence, educator judgment, transparency, implementation support, and regular review of learning outcomes, stating that “Responsible design is the floor.” It builds on a July 2025 federal AI-in-education position organized around five principles: educator-led, ethical, accessible, transparent, and protective of student data. The August letter asks five practical questions about what learning problem a tool solves, when it should be used, for whom, for how long, and what evidence shows that it improves learning. It also points to outcomes-based contracting in Louisiana, Arkansas, Indiana, Michigan, and Texas as a mechanism for tying vendor payment to results. This is federal guidance, not a binding universal curriculum rule, and states and local communities retain substantial authority over classroom implementation. Source record: DA-CH11-S6. Archive reference: pending.

7. Tim Fütterer, Lisa Bardach, Jochen Kuhn, Stefan Daniel Keller, et al., “Enhancing School Students’ Self-Regulated Learning through Generative AI Support: A Randomized Controlled Trial,” Educational Psychology Review 38, Article 42, April 14, 2026. https://doi.org/10.1007/s10648-026-10133-8. Accessed September 11, 2026. The randomized classroom study assigned 371 students in Grades 7 through 9 to two targeted generative-AI interventions or a standard ChatGPT control across six 45-minute sessions. The targeted interventions showed some benefit for perceived utility value but no statistically significant advantage on effort, domain-specific knowledge, or elaboration-based strategy use. The study is used as a counterweight to claims that customization alone guarantees better learning. Source record: DA-CH11-S7. Archive reference: pending.

8. Grace Liu, Brian Christian, Tsvetomira Dumbalska, Michiel A. Bakker, and Rachit Dubey, “AI Assistance Reduces Persistence and Hurts Independent Performance,” COLM2026, arXiv:2604.04721v5, October3,2026. https://arxiv.org/pdf/2604.04721 Jason Pohl, Berkeley News, “Using AI for just 10 minutes erodes your ability to persist at hard things,” October9,2026. https://news.berkeley.edu/2026/10/09/using-ai-for-just-10-minutes-erodes-your-ability-to-persist-at-hard-things/ Accessed October10,2026. Three online randomized experiments;1,222 recruited, with exclusions reported separately. The larger arithmetic experiment’s aggregate skip-rate difference was not significant. Immediate unaided testing cannot establish durability or all-task effects. The educational recommendation is author analysis. OpenAI systems are used in this Living Edition’s production workflow; see the transparency statement. Source record: DA-CH11-S8. Update 082. Archive reference: pending.

9. Jing Liu, Tracy Sweet, Marty Haoyuan Chen, Jonathan Engelberg, Megan C. Masters, Martyn Clark, Axel Persaud, Alia Lancaster, Jeffrey K. Hollingsworth, and Jennifer King Rice, “The Effects of Course-Integrated AI Tutoring on Student Performance and Engagement: A Randomized University Trial,” EdWorkingPaper 26-1598, October 2026; paper dated September 30, 2026. https://doi.org/10.26300/y3f8-vh05. Paper and PDF: https://edworkingpapers.com/ai26-1598 and https://edworkingpapers.com/sites/default/files/ai26-1598.pdf. University of Maryland College of Education, “Study Suggests More Guidance Needed in Use of AI Course Assistants,” October 16, 2026. https://education.umd.edu/news/10-16-26-study-suggests-more-guidance-needed-use-ai-course-assistants. Accessed October 17, 2026. The new event is the university's October 16 discussion of an earlier paper, not an experiment conducted that day. Fall 2025 trial at one university; instructor-level random assignment within course blocks; the estimates concern offering access, not randomly assigned individual use. The same-course subsample contained 1,353 students. Its adjusted final-grade result was statistically significant; the full-sample grade result was not. About 14.63% of students offered access ever used it. Other AI tools remained available. Course grades and platform activity do not independently establish learning, and these results do not establish that all AI tutoring harms students. Used for the specified rollout and its limits. The institutional recommendation is the author's analysis. The tested assistant used GPT-4o; OpenAI tools are also used in this book's production workflow. Source record: DA-CH11-S9. Update 102. Archive reference: pending.

10. Justin Spelhaug, Microsoft On the Issues, “Expanding AI literacy with the Public Library Association and community-based programs,” October 22, 2026, https://blogs.microsoft.com/on-the-issues/2026/10/22/expanding-ai-literacy-with-the-american-library-association-and-community-based-programs/; National Academy for AI Instruction, “AI Safety & Privacy Standard for Schools: Memorandum of Agreement,” September 2026, Version1.0, pp.1-3 and8-9, https://www.aft.org/sites/default/files/media/documents/2026/NAfAI-School_AI_Privacy_Standards.pdf. Accessed November 1, 2026. Microsoft's announcement establishes the November1 request-availability milestone, not verified adoption by every district. The published standard defines covered educational products, excludes general-purpose products merely used by educational institutions, provides an incorporation process, and specifies student action controls with safety/security exceptions. Actual district agreements and implementation must be checked separately. The text does not claim a nationwide statute, universal coverage or independent effectiveness. The second paragraph is author analysis. Source record: DA-CH11-S10. Update 126. Archive reference: pending.

11. Chris Phillips, Google, “Designing responsible AI for learning,” November 3, 2026, https://blog.google/products-and-platforms/products/education/ai-learning-era-updates-2026/; Barbara Ortutay, Associated Press, “Google will cut off search access for students at schools that reject AI,” November 3, 2026, https://apnews.com/article/ai-schools-classroom-students-google-7ba4219e505db1c11d672f6eefe7155a. Accessed November 4, 2026. Google establishes the prospective K-12 account controls; AP describes separate product switches and AI integration within Search. Neither establishes current rollout or learning outcomes. The second paragraph is author analysis. Source record: DA-CH11-S11. Update 130. Archive reference: pending.