The London-based education company Eedi, together with Google DeepMind, has published the full results of its second randomized study on AI tutoring in math. 1,179 students at ten public schools in the UK practiced on Eedi's platform for twelve weeks, ten of them with instruction.
Each student was permanently assigned to one of four groups and always received the same type of help after every wrong answer: the platform's usual explanation, a human tutor in a text chat, the AI tutor alone, or the AI tutor supplemented by a summary of the student's own earlier mistakes. The AI tutor is based on a version of Gemini adapted for classroom use. A human reviewed every AI message before it reached the teenagers.
What did AI tutoring achieve in the practice session?
Right after the mistake, the AI helped as well as a human did. After a conversation with a tutor, learners solved a similar task correctly in 56 to 58 percent of cases, after the usual explanation only in 43 percent. The additional information about earlier mistakes did not noticeably make the AI tutor better.
The Eedi study in numbers
1,179 learners at ten schools in the UK, results from October 29, 2026
- 43 %correctly solved follow-up tasks after the usual explanation
- 56–58 %correctly solved follow-up tasks after a conversation with a tutor, AI or human
- 94.7 %of AI messages were approved unchanged by the reviewers
- 7AI messages with an incorrect math statement reached learners
In the standardized math test Star Maths, the group with the AI tutor was not ahead after ten weeks.
Sources: Eedi, Google Blog
Google emphasizes in its report that Gemini guided learners toward the solution path in almost 99 percent of sessions instead of giving away the answer. In more than 85 percent of chats, the learners did the calculating themselves.
Why does the math test show no lead?
In the standardized Star Maths test, the 916 learners who took part both before and after performed similarly with the AI tutor as with the usual explanation. The group with human tutors was slightly lower, possibly by chance. Tutoring through conversation is more thorough, but covers less material. Google puts the difference at about 14 tasks per session compared with 19 with the usual explanation.
Eedi considers ten weeks too short. An earlier trial spanning two years suggests that clear effects only show up after a year or later. Eedi regards only as preliminary the finding that girls with a tutor performed somewhat better than without a tutor, while boys performed worse.
How often was the AI wrong?
The reviewers approved 94.7 percent of AI messages unchanged. Eleven AI messages contained an incorrect math statement. The reviewers caught four; seven reached the teenagers despite the review. The human tutors also made seven genuine calculation errors. For the one-year follow-up study with more than 15,000 learners at 26 schools, started in September 2026, Eedi has changed the procedure: humans no longer review every message, but supervise the conversations and intervene in a targeted way. Eedi calls on all providers of AI for the classroom to disclose their error rates.
What this means for you
Eedi does not mention an offering in German for its tutor. Gemini for schools, however, is also available in Germany: Google provides Gemini for Education to schools and universities free of charge, provided they meet the requirements for educational institutions. With a school account, you can use Gemini as soon as your school has enabled the service. Students under 18 get their own version with additional safeguards.
If you use AI for learning, it's best to have a mistake explained step by step and then work through two or three more tasks without help. This is exactly the amount of practice the learners with a tutor lacked. Briefly double-check every calculation step by the AI, because seven incorrect statements got through even with human review.
Feel free to explicitly ask Gemini in the prompt to ask questions instead of stating the solution. How to phrase such prompts is shown in the Gemini tutorial on prompt engineering.
As a teacher, it's best to plan AI explanations in addition to practice time and not shorten any tasks for it. How widespread explaining via AI already is was shown by the student survey in Hesse: 64 percent there have AI explain classroom material to them.
Eedi published the results on October 29, 2026.
Sources
