I’m aware of the irony in writing a whole post about letter grades for AI when I write about killing off grades for students. I’ve been paying some attention lately to the work happening with KORA.
KORA is a nonprofit benchmark that tests how AI behaves with kids. In July, it tested 12 popular AI chat apps, running 781 simulated conversations for each version it tested, covering 26 child-safety risks across three age ranges (7 to 9, 10 to 12, and 13 to 17). Seven of the apps were tested twice: once on an account set up for a 14-year-old, and once on a regular adult account, because plenty of kids are on those too. KORA also audited each app’s store listing, policies, settings, and interface. The results are on its apps leaderboard.
Of the 19 versions tested, one earned a B: MagicSchool on a student account, at 74%. Character.AI and PolyBuzz got Ds at 27%. Everything else got a C, including tools districts put in front of students: SchoolAI (57%), Khanmigo (56%), and the teen versions of ChatGPT (59%), Copilot (53%), and Gemini (45%).
Who made this and why
In Before You Share That AI Headline, I went back to the five key questions I used for years teaching Media Literacy. A leaderboard is a message too, so I ran this one through them.
Who created it? KORA describes itself as independent and says it has been funded entirely by its founder, Mathilde Collin, so far. It credits more than 30 child-safety experts, psychologists, and researchers with shaping its list of risks, and its code is public. Why is it being sent? To publish how these systems behave with kids. KORA is also careful about what its scores mean. Its FAQ says a strong score “does not imply approval or recommendation,” and that high scores don’t eliminate risk.
How was it made? Another AI plays the child, and AI judges grade the conversations. That’s the only practical way to run 781 conversations per app, and it’s also a reason to read the scores as a floor. I’d want to know how often a human reviewer would agree with the judges, and I couldn’t find that number published.
The scores
KORA reports four categories for each app. The one closest to what we care about in school is learning and development. MagicSchool was the only app above 65% there. Most of the general assistants (ChatGPT, Copilot, Gemini, Meta AI, Snapchat, TikTok) landed somewhere between 35% and 43%.
KORA also scores behaviors I’ve worried about for a long time. Does the tool agree with a kid just to keep them happy? Does it act like their friend? Does it point them toward a trusted adult when it should? I’ve called LLMs bias engines (LLM: Synonym for Bias), and I think agreement is more dangerous than hallucination because agreement feels like quality. A ten-year-old has even less reason than I do to question a tool that tells them they’re right.
What the scores can’t tell a district
KORA tested each app the way a kid would find it: default settings, no special setup. It’s also not what students in your district get. Districts set these tools up. A teacher builds a classroom space, someone turns features on or off, someone writes the instructions a custom bot follows. Two districts using the same product can be handing students very different tools. I’ve said for years that the default version of everything sucks, and districts change those defaults all the time. KORA didn’t test a district’s version of any of these tools.
Some tools districts buy aren’t on the list at all (Brisk and Gemini for Education, for example). The tests ran in July, and these products change every few weeks. And a safety score doesn’t settle the rest of the decision: whether a tool meets your student data privacy obligations, whether it works with a screen reader, whether it fits the AI polic. A lot of districts are writing those policies right now because their states require it.
What I’d do with the leaderboard
- Look up every student-facing AI tool you’ve approved, and write down which ones KORA hasn’t tested.
- Read the four category scores under each letter grade. An app can do fine overall and poorly on the risks that matter most for your youngest students.
- Test your own setup. Have a few adults talk to the tool as a student would, using your district’s actual configuration. KORA publishes example scenarios that make a good starting list.
- Share what you find with families. Most parents were never asked about these tools.
- Put a date on the calendar to check again.
I’d much rather a district that allows these tools make the call with evidence like this in hand than with a vendor’s claims.
Scores cited are from the KORA Apps leaderboard (apps evaluated July 1 to 26, 2026). Inclusive AI Strategies has no affiliation with KORA.