Today the evidence caught up with the hype. Stanford published a review arguing the strongest results in AI tutoring come from tools that help human tutors — not ones that replace them. Common Sense Media finished its K-12 curriculum and found 70% of American teens already use AI for schoolwork while only 27% have ever had a teacher explain what AI actually is. And NASA quietly banned generative AI from a national student engineering challenge. Here is what mattered in AI and education over the last 48 hours.
1. Stanford: the strongest AI tutoring evidence is for tools that back up human tutors
Source: EdTech Innovation Hub, reporting on Stanford SCALE Initiative · 27 August 2026
Stanford's National Student Support Accelerator and AI Hub for Education published a review titled AI Tutoring is Not a Monolith: What We Actually Know. It separates AI tutoring into models based on how much human involvement survives, and its central finding is blunt: live, human-led tutoring still has the strongest evidence base, and fully automated tutoring has not built anything comparable.
The engagement numbers are the part worth sitting with. Across two randomised controlled trials in two school districts, between 40% and 47% of students left to use an AI tutoring platform independently never used it at all. Those who did engage used it for only four to five weeks of a multi-month intervention, averaging roughly two to five minutes a week. A separate study of 181,000 students on supplemental maths software found only 5% reached the recommended 30 minutes of weekly use while 41% never logged in. By contrast, when AI coached the human tutor with real-time recommendations, students were four percentage points more likely to master lesson topics — rising to nine points for students taught by lower-rated tutors.
The quiet lesson: dosage beats availability. A tool nobody opens has an effect size of zero.
2. Common Sense Media finishes its K-12 curriculum as 70% of teens report using AI for schoolwork
Source: EdTech Innovation Hub · 27 August 2026
Common Sense Media completed the K-12 rollout of its free Digital Literacy and Well-Being Curriculum, adding a 29-lesson high school programme across six pathways covering AI literacy, media literacy, online harms, health, financial literacy and a Freshman Foundations sequence. Alongside it, its Youth AI Safety Institute published survey research conducted by NORC at the University of Chicago with 1,017 US teenagers aged 13 to 17, fielded between 30 April and 14 May 2026.
Seven in ten teens said they use AI for schoolwork. But only 27% said a teacher had ever discussed what AI is and how it works, 30% had discussed safe use, and 26% had covered how to judge whether AI output is accurate. The usage picture is more nuanced than a cheating panic suggests: 45% use AI to generate ideas or work out how to start, 43% to check a completed answer, 40% to get feedback on their own writing — though 25% do use the answer as-is. Notably, 39% said they feel they are missing out on learning when AI completes assignments, and 44% of AI users said a tool had been blocked at school, of whom 59% simply switched to a personal device.
Common Sense's Ilana Lowery put it plainly on LinkedIn: AI is already in schools; AI literacy is not.
3. A 6,997-student trial finds an AI maths tutor helps after mistakes — and slows everything down
Source: EdTech Innovation Hub, reporting on an NBER working paper · 24 August 2026
A randomised trial across 20 schools in Hamilton County Schools, Tennessee, put nearly 7,000 middle schoolers through a single 50-minute maths lesson in late March 2026, half of them with a guard-railed AI tutor built into the practice platform. The paper, Making AI Tutoring Productive: Evidence from a Mastery-Based Math Practice Experiment, then tested 6,327 of them again a week later.
Among students on a mastery workflow who got a question wrong, the AI raised next-attempt accuracy by 8.5 percentage points and cut the extra attempts needed by roughly one — but added 2.88 minutes to getting there. On the delayed test, combining AI with mastery lifted correct answers on the practised question from 37.0% to 40.2%, a 3.2 point gap the authors themselves call suggestive rather than definitive at p=0.065. On the unpractised question there was essentially no difference at all. Co-author Alp Sungu summarised it neatly: the potential of AI tutors may depend on how the technology is operationalised in schools, not the technology alone.
Worth flagging: one assignment, a four-question follow-up test, and a working paper that has not been peer reviewed.
4. NASA bans generative AI from its 2027 Student Launch challenge
Source: EdTech Innovation Hub · 26 August 2026
NASA opened applications for the 2027 Student Launch challenge with a new rule barring students from using generative AI or other AI tools for any challenge-related work. It applies to both the college division and the middle and high school division, and it is broad: browser-based generative AI, AI-assisted coding tools and AI-generated imagery are all out, and challenge content must not be uploaded to or processed by AI systems at all.
The rule sits inside NASA's existing requirement that students complete 100% of the project themselves, with safety-critical work on motors and ejection charges still handled by adult mentors. The engineering brief itself is a serious one — university teams must build a rocket that reaches 4,000 to 6,000 feet, then autonomously locate a passive visual target after landing, photograph it and calculate the distance to it without GPS and without moving, all within 15 minutes and before any human interaction. Applications close on 14 September at 8:00am Central, teams are announced 1 October, and Launch Week runs 8-10 April 2027 in Huntsville, Alabama.
Expect more competitions to copy this wording. It is the cleanest AI ban anyone has written this year.
5. Common Sense Media is putting ChatGPT for Teens through an independent risk assessment
Source: EdTech Innovation Hub · 26 August 2026
The Youth AI Safety Institute at Common Sense Media is preparing an independent assessment of OpenAI's ChatGPT for Teens, with results expected in the coming weeks. It follows the organisation's earlier evaluation of ChatGPT, which it rated High Risk overall and Unacceptable Risk for mental health support.
The testing focus is specific. Researchers want to know whether students can bypass the Responsible Homework Reminder within the same conversation, by switching mode, or simply by using ChatGPT logged out. They are also checking the accuracy of learning visualisations, and how reliably teen users are identified at all — including logged-out users and those who enter a false age — plus how fast safety alerts reach a linked parent account and how often accounts are actually linked. Developmental behavioural paediatrician Jenny Radesky flagged the customisable voice modes as her biggest concern, questioning what Professional, Friendly, Candid, Quirky and Cynical personalities do to teenagers' confidence, motivation and emotional attachment.
Robbie Torney and Radesky: the features are well intentioned, but external testing is needed to confirm they perform as intended.
6. US Education Department tells schools to judge EdTech by outcomes, not screen time
Source: EdTech Innovation Hub · 25 August 2026
In a Dear Colleague Letter dated 20 August, the US Department of Education urged states and districts to evaluate education technology by whether it improves learning rather than treating screen time as the primary test. Districts are encouraged to seek products with demonstrated positive outcomes, including through randomised controlled trials, to review whether existing tools still serve instructional goals, and to remove technology that repeated findings show is not working.
Assistant Secretary Kirsten Baesler drew a sharp line between distraction and instruction, arguing that phones and algorithmic social media undermine focused classrooms while instructional technology should be judged by its impact on learning, not simply by whether it happens on a screen. The letter puts pressure on vendors too: publish rigorous independent evaluations where feasible, give classroom-grounded implementation guidance, and be transparent about limitations as well as capabilities. Responsible design, it says, is the floor. It names Louisiana, Arkansas, Indiana, Michigan and Texas as states exploring contracting models built around shared performance measures.
Two stories, one theme: 2026 is the year AI in education started being asked for receipts.
Stanford Just Killed the AI Tutor Hype — Here's What Actually Works
It is the rare research story with a number that stops people scrolling. Nearly half of students handed an AI tutor never opened it once, while the version that actually improved mastery was the one nobody markets — AI coaching the human teacher in the background. That contrast is the whole video.
Stanford just reviewed the evidence on AI tutors — and in two trials, between 40% and 47% of students never opened the thing once. The tools that actually moved the needle were not tutoring students at all.
What to do with this
If you are a student: the Common Sense data suggests the highest-value uses are the ones you already half know — getting unstuck at the start, checking a finished answer, asking for feedback on work you wrote yourself. The 25% who paste the answer straight in are the group reporting they feel they are missing out on learning.
If you are a teacher or a parent: the Stanford review and the Education Department letter point the same way. Ask what a tool is supposed to improve, ask for the evidence, and watch whether students actually use it — because a 41% never-logged-in rate is the real failure mode, not cheating.
Want this digest in your inbox every morning?
One short email a day on AI and education — what changed, and what it means for students and teachers.