Who AI reading tools actually help, and who they slow down

Two children get the same comprehension passage. One puts on headphones and lets the tablet read it aloud while she follows the words. The child beside her tries the same thing, because it looks easier. Twenty minutes later the first has understood more of the passage than she usually would. The second has worked harder for nothing.

That is not a story about effort or attitude. It is close to what a study published this year in the journal Reading and Writing found when researchers in Taiwan sat 47 junior high students in front of an eye tracker and gave them the same texts twice, once read silently and once with text-to-speech. Sixteen of the students had dyslexia, fourteen had ADHD, and seventeen were typically developing. The three groups did not respond the same way at all. For the students with ADHD, comprehension under text-to-speech was significantly better than under silent reading. For the typically developing students, the audio cut across their normal reading pattern and raised their cognitive load, with no comprehension benefit to show for it.

This is the part of the AI story that gets lost in the noise about chatbots and homework. The clearest good that AI does for children is not a general boost for everyone. It is a specific accommodation for a specific difficulty, and it works because it is targeted.

The child the tool was built for

In Singapore the group is not small. MOE told Parliament in January 2021 that, based on data from 2016 to 2019, about 3.5% of Primary 3 students were reported as having dyslexia, with roughly another 1% identified at Secondary 1, which puts Singapore inside the international range of 3% to 10%. In a class of thirty, that is one child, sometimes two.

Some of the help is already sitting in the school platform. In an October 2024 parliamentary reply on special needs education, MOE said the Student Learning Space carries “AI and non-AI enabled features that support accessibility, such as video and audio transcriptions and text-to-speech feature”, and that AI-enabled feedback assistants and adaptive learning systems have been progressively rolled out to customise learning for students with diverse needs. The same reply noted that MOE funds assistive technology devices for students with special educational needs in both mainstream and SPED schools, and that specialised SEN officers in mainstream schools grew from around 450 in 2017 to about 680 as of July 2024.

Worth noticing what that means: the read-aloud button is not a purchase or a download. It is sitting inside the platform your child already logs into for homework.

The gains are real, and smaller than the marketing

Be careful with the size of the claim. The best available synthesis is a 2018 meta-analysis in the Journal of Learning Disabilities by Wood, Moxley, Tighe and Wagner, which pooled studies of text-to-speech and read-aloud tools for students with reading disabilities and found an average effect of 0.35 (95% CI 0.14 to 0.56). That is a modest, genuine effect, and the authors themselves put it carefully: these tools “may assist students with reading comprehension”, with more research needed on what makes them work.

The Taiwan study adds a wrinkle worth understanding. For its dyslexia group, text-to-speech did not significantly raise comprehension scores. What it did was cut the visual work, reducing the processing demand on key terms while comprehension held steady with silent reading. In other words, the child understood the same amount for noticeably less effort.

For a ten-year-old who has to get through six subjects before dinner, that is not a consolation prize. Effort saved on decoding is effort available for the actual thinking, and for the sheer act of keeping going. It just will not show up on the next spelling test, which is exactly why families abandon a tool that was working.

What we would actually do

Match the tool to the difficulty, then check the effect yourself. If your child struggles to hold attention on a page, the evidence for read-aloud support is encouraging, though it rests on small studies and should be treated as a lead rather than a finding. If the difficulty is decoding, expect a stamina gain rather than a marks gain, and judge it on that. If your child reads comfortably already, audio is more likely to get in the way, so leave it off.

Three practical moves:

• Ask the form or SEN teacher which accessibility features are turned on in your child’s SLS account, and ask them to show you where.
• Trial it in one subject for a fortnight, and watch the last fifteen minutes of homework, not the marks. Is your child still with it, or fried?
• If a paid device or software would help, look up the Assistive Technology Fund before you buy. It subsidises up to 90% of approved devices with a lifetime cap of $40,000, and from January 2026 the per capita household income ceiling rose from $2,600 to $4,800, which brings a lot of ordinary families inside the gate for the first time.

None of this is solved. A 2025 UNESCO IITE review of inclusive-technology practice named the standing obstacles plainly: accessibility, affordability and digital literacy. A tool nobody was taught to use is a tool nobody uses.

Still, when people ask what AI is genuinely good for with children, this is the honest answer, and it is quieter than the headlines. Not a machine that makes every child cleverer. A machine that removes one specific barrier for one specific child, so the effort goes where the learning is. Worth asking the school about at the next parent-teacher meeting.

(Disclosure: Mentus AI builds AI mentors for children, so we have a commercial interest in this area. The evidence above is other people’s, and we have linked all of it.)

Related reading: when AI helps kids learn, and when it quietly doesn’t and why Singapore’s school AI listens instead of answering.