It's Sunday afternoon, the math homework has stalled on fractions again and, while you look for help on your phone, the perfect ad pops up: an artificial intelligence tutor that adapts to your child's pace, talks to them, asks them questions, celebrates their right answers. Five more follow, almost identical. "Personalized." "Adaptive." "Like a private tutor, but available 24 hours a day."
The problem isn't that those promises are false. It's that, from the outside, they all sound the same, and the real difference between them lies somewhere the ads never show: the evidence.
The Evidence Ladder
Before looking at specific apps, it helps to have a map. Not all "proof" is worth the same, and in educational technology there are basically three rungs:
Testimonials: happy families, five-star ratings, success stories. They're sincere, but they can't tell whether the child improved because of the app, because of the family's extra effort, or because they were going to improve anyway.
Company data: surveys of their own users, before-and-after grades. Better than nothing, but without a comparison group we don't know what would have happened without the app.
Randomized trials: students are randomly split between those who use the tool and those who don't, and the two groups are compared. It's the only method that lets you attribute the improvement to the tool.
One piece of context that helps read the numbers: in rigorous education trials, the average effect of an intervention is around 0.10 standard deviations. Any result above that is already notable.
With that map in hand, let's get to the apps.
The Flashiest Ones: Lots of Promise, Little Independent Proof
Synthesis Tutor is probably the most eye-catching one right now, and it's already available in Spanish too. It's a math tutor for kids aged 5 to 11 that works by voice: the child reasons out loud and the AI responds, detects misunderstandings and adjusts the difficulty. It has a headline-worthy origin: it grew out of the experimental school its founder, Josh Dahn, set up in 2014 inside SpaceX at Elon Musk's request. Its Socratic approach and the "micro-assessments" built into each lesson are pedagogically sensible. But, as of today, its website doesn't cite any independent efficacy study: what it offers are family testimonials and the fact that tens of thousands of families use it. Independent reviews agree that it fits self-motivated or advanced kids very well, and the rest less so.
Smartick is the best-known Spanish option: 15-minute daily sessions with exercises the AI adapts to each child's level. The company publishes striking figures (that 94% of its students improve in arithmetic and problem solving, or that their average grade rises by almost two points), but they come from its own surveys and data, not from a trial with a comparison group.
None of this means they're bad apps. It means that, for now, we don't rigorously know how much they help, or whom.
The Tutor Almost Nobody Used
The most instructive case isn't an app without studies, but one that does have one. Khanmigo, Khan Academy's AI tutor, is designed exactly the way experts recommend: it doesn't give the answer, it guides with questions. A two-year trial in 18 secondary schools in Tennessee, led by Philip Oreopoulos (University of Toronto), put it to the test with low-achieving students.
The result has two sides. Access was almost universal: 96% of students tried it at least once. But actual use was scarce: the typical student only messaged it on a third of the days they practiced, and only in 17% of the sessions where they made mistakes, which is precisely when a tutor is useful. Many messages were one-word replies, clicks on suggested prompts or attempts to get the answer straight out of it. Students did improve, but the researchers attribute the improvement mainly to Khan Academy practice, not to the chatbot.
The authors' conclusion is one worth writing down: "human attention is the key ingredient." Or, as Oreopoulos puts it: if a motivated student wants to use the technology as a good tutor, it's ready; the problem is that most don't use it that way.
The Ones That Do Have Trials (and Aren't the Flashiest)
Eedi, a British math platform, was designed around something very specific: diagnosing the exact mistake each student makes. Its AI tutor is deliberately limited and only steps in when a student has gotten a specific question wrong. In a randomized trial with 2,901 students aged 11-12 in 20 UK schools (2023-2024 school year), those who used Eedi gained the equivalent of about two extra months of progress (an effect of 0.17). Those who used it regularly, just 10-15 minutes a week, doubled that effect. Two honest caveats: the trial was commissioned by Eedi itself, although it was carried out by an external evaluation team; and the comparison group wasn't "doing nothing" but mostly using other digital platforms, which makes the result more demanding.
ASSISTments is, of all of them, the one with the strongest evidence, and perhaps the least glamorous. It isn't a chatbot: it gives immediate feedback while the student does math homework and prepares reports so the teacher knows what to review the next day. It has two large randomized trials with students aged 12-13: one in Maine (46 schools, 3,050 students), with an improvement equivalent to 60% of what is normally gained in a school year, and a replication in North Carolina (63 schools, 9,073 students) whose effects were still there a year later. That earns it the top rating ("Top Tier") from Evidence-Based Programs. And an important detail: the effect was larger for students who started out further behind. Its limitation for a family in Spain is practical: it's designed to be used through the school and its content follows the US curriculum.
What the ones that work have in common. If you line up the tools with the best evidence, the pattern is clear, and it has nothing to do with how sophisticated the AI is:
They act on the specific mistake, not on the topic in general. They detect exactly what went wrong and respond to that.
They give hints or feedback, not the answer. The effort stays with the child, which is what we already covered in La IA que hace mejor a tus hijos en clase (y la que los hace peor) es la misma IA (in Spanish).
They ask for little time, but consistently. In Eedi, 10-15 minutes a week of regular use was enough to double the effect.
They keep an adult in the loop. In ASSISTments, the teacher gets the reports; in Khanmigo, according to its own researchers, what was missing was precisely human attention.
Five Questions Before You Pay for the Subscription
We already covered on Papitek what to ask before trusting any AI educational app. For math apps, these five are the ones that filter the most:
A. Is there any study with a comparison group, or just testimonials? Look for a research section on the app's website. If there are only families' opinions, treat it as a bet, not a guarantee.
B. Who did the study? A trial commissioned by the company isn't invalid, but it's worth more if an external team runs it and publishes its methods.
C. Does it give the answer or give hints? Test it: make a mistake on purpose and see what it does. If it corrects you and explains why, good. If it just hands you the result, bad.
D. How much time does it ask for, and how often? Be wary of the ones that need hours. The ones that work tend to ask for a little time, consistently.
E. What role does it leave for you? The best tools tell you where your child gets stuck so you can support them. If the app promises you won't have to do anything anymore, that's exactly the opposite of what the evidence says.
A Good Coach Doesn't Run for You
A good track coach doesn't carry you on their shoulders to the finish line. They watch you run, spot the movement that's off, correct that movement and let you try again. Math apps work the same way: the ones that work aren't the ones that seem smartest, but the ones that most resemble that coach. They detect the mistake, give the right hint and leave the effort where it belongs, without leaving out the adult alongside.
The next time an ad promises you a tutor that does it all, maybe the question isn't "how much does it cost?" but this: what proof does it have that my child is going to learn, and not just finish the homework?