AI reading assessment

Chapter 1 · The problem

Testing how a child reads aloud needs an adult to sit with each child, one at a time

A Grade 3 child sits across a desk from an adult holding a tablet. She reads a short passage aloud while a timer runs, and the adult taps each word she gets wrong. At the end, the tablet works out how many words she read correctly in a minute.

That number, correct words per minute, is how most reading programs measure fluency. A child who reads slowly and stumbles on common words cannot yet read to understand. The test is simple to describe and expensive to run, because it needs one adult listening to one child.

I knew the problem from my earlier job, before Wadhwani AI

In 2022, at Central Square Foundation, I helped a state government run a sample baseline test of Grade 3 children, and I built its digital oral reading test on Tangerine, an open-source assessment app. That is how I knew, before joining Wadhwani AI, how much this test costs in people. CSF's 2022 article on the baseline assessment describes the field work (Exhibit 1).

The digital tool cut printing costs to almost nothing. The next cost to cut was people: a sample of one grade, once, still needed over a thousand trained adults, and a teacher who wanted to test her own class regularly would have to do all the listening herself.

Speech AI can do the listening

If a model can hear which words a child read correctly, the adult no longer has to. The teacher records the child reading in an app, and the score comes back by itself. That is what the oral reading fluency tool at Wadhwani AI does. In one state, the program reported 1.68 million assessments in four weeks with no outside investigators.

Sources: CSF's 2022 article on the baseline assessment; Wadhwani AI program deck on the solution, 2025.

Chapter 2 · The insight

Teachers trusted the scores and needed to know what to do next

When I joined Wadhwani AI in September 2025, the reading tool was already in use across two whole states. It returned correct words per minute, accuracy, and a mark of right or wrong on every word. In one state, most teachers found it easy and believed it. The next question was what they would do with the result.

Teachers said what they needed from the results

In field visits, teachers explained what held them back from acting on the scores. The program's field notes record their reason:

A number for each child told a teacher who was behind. The next step was telling her what to do with those children on Monday morning.

A perception study in one state later asked teachers directly (Exhibit 2). Most trusted the tool and used it without help. The study pointed to the next step: using results to plan lessons and targeted practice.

834 teachers Teachers' own answers in a survey.

A field check traced wrong scores to how recordings were made

In November 2025, senior state officials in one state said scores did not match the reading they saw in classrooms. A field check of 30 recordings in 4 schools found that 12 did not match the child, all from 2 of the schools. The cause was teachers recording in batches, sometimes on a child's behalf, not the model.

The check recommended verifying the child in real time and detecting mismatches automatically. Flagging an adult's voice, and two children's voices in one recording, went onto the product plan. Designed A wrong recording puts a wrong score on a real child's record, so this matters as much as model accuracy.

Sources: Program field notes; perception study of 834 teachers and 213 school heads; field validation report, November 2025; product plan, 2026.

Chapter 3 · The loop

Assess, diagnose, remediate: scoring and four reading groups, then practice for each group

The plan was a loop. A child reads, the AI scores, the child joins a reading group, and the teacher gets practice for that group. Then the child is tested again. By mid-2026 the first three steps worked, and the fourth was in design.

Groups came from simple rules that each state approves

Four groups run from children still learning letters to fluent readers. The plan had been a model that would find natural groups in the scores. It ran late, so the team placed children by ranges of fluency and accuracy instead. Experts set the ranges by labeling sample recordings, and each state vetted them. A rule a teacher and an official can read is easier to trust than a cluster nobody can explain.

Practice for each group is the next product feature

The teaching half of the loop was in design. The remediation lesson plan was at a concept stage, and a module that drafts practice with a language model was planned for 2027. The team compared three providers for that module on how closely each followed the intended pedagogy, with a pedagogy expert reviewing the output. Designed

In one state, the state runs its own 100-day remediation program for the three lower groups, using lesson plans in its teacher app.

Sources: Program product plan, deployment plan and concept note, 2026; state meeting minutes, October 2025; one state's evaluation protocol, October 2025.

Chapter 4 · State rollout

Approval to reach 15M+ children in 8 states, with five new agreements in my time

Two states use it at census scale. During my tenure the program signed agreements with five more states and a letter of intent with another, and ran its first pilot in one of them. I led that expansion, winning approval from state education secretaries to reach 15M+ children in 8 states.

What I did in the expansion

I planned the school sample for one new state's voice recordings: 300 schools across Grades 1 to 5. I worked on the final round of that state's agreement before it was signed in November 2025. In July 2026 I presented the program's reach and plans at a sector consultation. State program officers ran each state day to day.

Each new language needed children's voices before it could launch

A passage that is easy in one script can be hard in another. The team built a reading framework for each language, grading words by letter count, matras and conjuncts, and from it banks of leveled paragraphs. Then it collected children reading aloud, sampled by district to cover each language's accents: from 250 to more than 400 hours in each of three new states.

The partial character set came with a rule for passages. In one new language the training data reliably covered 108 of more than 400 letters. Passages must then avoid the letters the model cannot yet score, so a child is not marked wrong for the model's gap.

For planning, the team sized the market from what a live state had actually achieved. About 95 million children study in Grades 1 to 8 in government schools. One live state had reached about 57 to 64% of its eligible children, so the obtainable market was set at 55 to 60 million. Designed

Sources: Program agreements sheet, portfolio overview and expansion sheet, 2025 to 2026; one state's data-collection plan and school-sample calculator, October 2025; consultation deck, July 2026; voice-data updates, June 2026; market-size note, 2026.

Chapter 5 · Evidence

The first two states have assessed more than 9 million children since 2023

The two census-scale states have assessed more than 9 million children since 2023. Approvals in 8 states cover 15M+ children. This chapter says what each figure counts and how much fell in my time.

9M+

unique children assessed so far in the first two states, 2023 to June 2026

~35M

assessments behind them, same period

~21M

assessments from October 2025 to June 2026, in the same two states

What the reach counts

The program's grant sets targets for reading gains. Chapter 6 sets out how I would measure them.

Sources: Program deck, July 2026, with figures to 25 June 2026; program assessment-numbers sheets; program reporting to its funder, July 2026; one state's evaluation protocol, approved October 2025.

Chapter 6 · What comes next

Next: measure reading movement by group, and catch bad recordings

Two kinds of work come next. One is measurement: showing how children's reading moves, group by group. The other is the product: catching the recordings that produce wrong scores, and finishing the teaching half of the loop.

The measure I would publish first

I propose one table per state: for each reading group, the share of children who moved up a group between two tests, with the number of children, the dates and the source. Next to it, the share of teachers who opened the group materials before the second test, once those exist. I would add a comparison group to the next study, so that movement can be set against schools without the tool.

Sources: Program product plan, 2026; my proposal.

Reflections

The scoring scaled to two whole states, and the teaching half comes next