How El Salvador Is Using Grok to Tutor Students, and How Parents Can Do the Same at Home

If you homeschool, you have probably seen the headline: El Salvador is using Grok to tutor students, and the scores look like Germany's. The partnership is real. The test is real. The idea that students in general now match Germany, or France, is not. Here is what was announced, who was tested, and how a parent can use Grok at home without handing over the teaching.

What was announced in December 2025

In December 2025, xAI and the government of El Salvador announced a plan to put Grok in public schools. EFE's September 2026 report and a write-up of the pilot describe it as xAI's "world's first nationwide AI-powered education program": Grok in more than 5,000 public schools over two years, personalized tutoring for more than a million students, and teachers kept in the work as partners. President Nayib Bukele reposted the announcement. The cost was not disclosed.

That is a rollout plan. It is not a score.

The pilot was a few hundred schools, not the whole country

It began as a pilot of about 200 first-phase schools, and 171 of them were assessed. In September 2026, Bukele wrote on X: "These are the results of our pilot program, which includes AI tutors and has been running in 171 public schools for just over a year." He said the work had expanded to more than 1,000 schools and would reach every public school within 18 months.

On that same account he named the partners: "Our AI tutors and learning platforms are being developed with @grok, @Kira_Learning, and support from I Hate Flying Bugs." Grok is in the stack. It is not the only name on it. Kira Learning's AI tutor is sold to schools, with coursework the school sets. Kira's public pages do not say that product is built on Grok. This article does not call Kira a Grok tutor.

Grok is one piece of a wider reform. The World Bank's September 2026 account describes curriculum, materials, teacher support, school leadership, technology, and assessment moving together. A tutor can matter and still not be the only reason a score moves.

The PISA for Schools numbers

In mid-2026, 1,198 volunteer students from those 171 pilot schools took PISA for Schools. The World Bank says the sitting was in July. EFE and Bukele have said June. Either way, this was not El Salvador's national PISA assessment. It was a test of a group of schools in a program.

The World Bank published the pilot scores next to the national average:

The World Bank says the OECD called these results "similar to the averages for Sweden and Germany" on PISA 2022. France is not named in that comparison.

Similar to Germany is a loose phrase

Germany's own PISA 2022 averages, as published in the German federal statistics table of those results, were 475 in mathematics, 480 in reading, and 492 in science. Set the pilot beside those figures and the shape of "similar" shows up.

• Mathematics: 417 against 475. That is about 58 points lower.
• Science: 462 against 492. That is about 30 points lower.
• Reading: 492 against 480. This is the close one.

Hear the math gap first. Fifty-eight points is not a rounding error. Reading is where the pilot sits nearest Germany's 2022 average, and science falls between the two. "Similar to Sweden and Germany" is the World Bank's wording, quoting the OECD. It holds loosely, and mostly for reading. That comparison does not name France, so this article does not say the pilot matches France.

They were volunteers, and the gain is not pinned on Grok

Bukele said it himself: "We never claimed these results represent the national average." The 171 schools were not a representative sample. The World Bank adds that the volunteers "are likely among the higher performers in their schools." A high score can describe who showed up.

The cause is unproven too. Bukele wrote: "These results also do NOT prove that AI tutors, by themselves, caused the improvement, although we believe they are a MAJOR part." The World Bank points the same way. The scores are a sign from a package of changes, not proof that one tutor produced them.

The national test, a few months earlier, was quieter

The national PISA 2025 results came out in September 2026. The data were collected in November 2025, at the very start of the reform. El Mundo reported that math and reading were about flat compared with 2022, and that science improved. National scores remain far below France and Germany. That picture and the pilot picture are not the same group of children.

What an outside critique adds

An OSF preprint asked how much of this selection alone could explain. If the volunteers were the top 5.5% of students in math, the top 11.7% in science, or the top 17.2% in reading, they would match the German averages with no learning gain at all. The pilot tested 7.0 students per school, against 25.4 in the national sample. Taken at face value, the jump implies about 6.5 years of schooling in one year. That is a reason to slow down, not a verdict on every assumption in the paper. It does show why "volunteers from 171 schools" cannot be retold as "El Salvador's students."

The sentence that survives

Here is the claim the sources can carry. In a non-representative group of 1,198 volunteer students from 171 pilot schools, inside a reform that includes Grok-based AI tutors, PISA for Schools scores were well above El Salvador's national average. The World Bank, citing the OECD, called them broadly comparable to Germany's and Sweden's 2022 averages, closest in reading. Who would have scored that high without the program, and how much of the change belongs to Grok, is not shown.

What the research says about AI tutoring

None of these studies used Grok. El Salvador is the only Grok example with outcome data this check found, and that data is the volunteer PISA for Schools sitting above, not a randomized trial of the model. No randomized trial of Grok turned up.

The World Bank's October 2026 feature describes the tutor as guiding each student with questions and step-by-step hints so they reach the answer on their own. It says the reform began with 10 schools in 2024, reached about 200 by mid-2025, and now reaches nearly 1,000, with tools built by an in-house team and private partners. That page does not name Grok. A write-up of the OSF preprint dates the xAI agreement to December 11, 2025, and notes that the tutor was therefore present for at most about half of the period the June 2026 assessment covered. The conventional pieces of the reform had been there longer.

Nigeria, 0.31 standard deviations

A World Bank randomized trial in Edo State used Microsoft Copilot (GPT-4), with a teacher in the room, for about six weeks in June and July 2024. First-year senior secondary students, about age 15, in nine schools in Benin City, had twice-weekly after-school English sessions. Policy Research Working Paper 11125 (De Simone and coauthors, May 20, 2025) reports 0.31 standard deviations on a composite test, 0.23 on English, and 0.21 on the end-of-term English exam. The World Bank's January 2025 blog describes the same six weeks and calls the gain about 0.3 standard deviations. The program was short, on one site, in schools chosen because they had computer labs. The composite includes AI and digital-skills items. Gains were larger for girls and for students who started ahead.

Harvard physics, 0.63 standard deviations

Kestin and coauthors, in Scientific Reports in June 2025, randomly compared a GPT-4 tutor, steered by carefully written prompts, with an in-class active-learning lesson. The students were 194 undergraduates in introductory physics at Harvard, in a crossover design. Linear regression put the gain at 0.63 standard deviations. Quantile regression, used because many students hit the top of the test, put it between 0.73 and 1.3. These were short post-tests on single lessons, and the tutor was built for those lessons. It was not an off-the-shelf chatbot.

Rori in Ghana, 0.36 standard deviations

Henkel and coauthors studied Rori, a WhatsApp math tutor for grades 3 to 9 in 11 schools. Students had two 30-minute sessions a week for eight months, at about $5 a student. Schools were assigned at random. The year-1 paper on arXiv (February 2024) reports 0.36 to 0.37 standard deviations of math growth. It was preliminary, the number of schools was small, and it had not been peer-reviewed when that preprint was posted.

Stanford Tutor CoPilot

Wang and coauthors at Stanford (working paper, October 2024) gave the AI to the tutors, not to the students. More than 700 tutors and about 1,000 underserved K-12 students were in the trial. Topic mastery rose from 62% to 66%, four percentage points. Students of lower-rated tutors gained about nine percentage points. The authors put the cost at about $20 per tutor per year. Mastery was scored inside the session. Their estimate of the end-of-year MAP effect is about 0.02 standard deviations.

Khanmigo in Tennessee, a small effect

The longest school trial so far is NBER Working Paper 35620 (Oreopoulos and Low, August 2026). Eighteen Tennessee middle schools used Khanmigo for two years, set up to coach rather than hand over the answer, during remedial math. The pre-registered cluster trial found about 0.06 to 0.08 standard deviations a year, and about 0.14 if a student was fully active. That is in the same range as Khan Academy practice without the AI tutor. Use was low. The median student messaged the tutor on only about a third of practice days, and in 17% of the sessions where they made a mistake.

Read together, structured AI tutoring has produced moderate to large gains in some randomized trials, including Nigeria, Ghana, and the Harvard lessons. The largest and longest school trial, Khanmigo, found small effects, limited mainly by how little students used the tutor. None of that evidence is a trial of Grok. El Salvador remains the only Grok example, and it is not a randomized one.

How a parent can use Grok at home

You do not need a ministry to borrow the useful part. You do need to stay the teacher. These uses sit beside a curriculum you already trust.

A grade-level explanation

Name the age. "Explain fractions to a 10-year-old, in short steps." If the answer is too cute or too much like a lecture, say so and ask again. You are steering a tutor, not accepting the first speech.

Guiding questions, one at a time

Say, "Don't give me the answer. Ask guiding questions one at a time." Your child works the problem out loud. You listen, and you step in when a turn is wrong. The finished number should not be the first thing on the screen.

A quiz, or a mistake left unfixed

Ask for a short quiz or a few flashcards on the lesson you just taught. You can also paste a worked answer and say, "Point out the mistake. Do not fix it." Then you decide whether the note is fair. A confident sentence can still be wrong.

Voice practice

The Grok apps' voice mode can carry spoken practice: reading aloud, a language phrase, or an oral quiz. Sit where you can hear both sides.

A photo, and a hint instead of the answer

Photograph the worksheet and ask for the next small step, not the solution. xAI's terms say you may provide images as input, and that you are responsible for what you send. If the page shows a child's full name, cover it first.

You are a Socratic tutor. I am a parent sitting with my child. The child is [age] years old. The subject is [subject]. Today's lesson is [topic].

Rules:
- Do not give the final answer unless I type "answer".
- Ask one guiding question at a time, then wait.
- If we are stuck, give the next small hint, not the solution.
- Use language a child of this age can follow.
- If you are not sure of a fact, say so.
- Do not invent a grade, a score, or a source.

Age, the account, and you in the room

The practical ideas above assume an adult is running the session. The age rules should come from xAI, not from a blog. On October 17, 2026, the consumer terms (last updated October 12, 2026) said: "You must be at least 13 years old or the minimum age required in your country to use the Service." A teenager of 13 to 17 "must have your parent or legal guardian's permission to use the Service, and they must agree to our Terms of Service." If you are the parent of a minor who uses the service, you and the minor are both bound by the terms, and you are responsible for all of that use.

The same terms say you may not share your account or make it available to anyone else. A child under 13 is not a user under those rules, and borrowing your login is not a workaround the terms allow. The consumer FAQ says the same thing in shorter form: "Grok is not directed to children or minors under the age of 13." You can still use Grok yourself while you prepare a lesson or sit beside an older child. The terms say a parent who uses the service to help with conversations with their children, including educational ones, must use the data controls in the product and pick the features that fit. They also ask parents to stay involved and to watch how teenagers use it. Grok inside X follows X's terms, the FAQ says, not only these consumer terms.

Kids Mode is not described in those terms or in that FAQ. A parent guide from Internet Matters describes a control in the app: open the menu (two horizontal lines at the top left), choose Kids Mode, turn on the top toggle, then turn on the lower toggle and set a PIN so Kids Mode cannot be switched off without it. Confirm those steps in the app you installed. No official xAI help page for Kids Mode turned up in this check, and app settings change.

The terms also warn that the model can be wrong: a step, a date, or a fact that sounds smooth and is not. Check the arithmetic. If a teacher assigned the page, say that an AI hint was used. You decide whether the answer stands.

What this does not choose for you

Grok can explain, quiz, and hint. It does not file a state notice or pick the books. The CurriculumQuiz matcher ranks programs from official pages and does not ask for a child's name. Quiz answers are never stored or sent. The site uses Google Analytics, which sets cookies, for anonymous usage stats. Use the tutor for the stuck problem. Use a real plan for the year.