Dytective online dyslexia screener reaches 80.6% recall

Rello et al.'s random-forest screener, tested on 3,644 children from 113 organizations in five countries, 79.8% accuracy; the tool has been used over 200,000 times. A 2026 paper reports confounding.

ⓘ

AI system

System

Dytective online gamified dyslexia screening test with a random-forest classifier

Model

A random-forest classifier with 200 trees and no maximum height.

Release

Implemented with Weka 3.8.3. No separate release identifies the fitted classifier.

Application developer

Luz Rello and co-authors (IE University, Northeastern, University of Washington, Carnegie Mellon, University of Barcelona), per the paper's author list

Role

The classifier estimated dyslexia risk from the screening assessment.

Configuration

The Weka 3.8.3 implementation used 200 trees without a maximum height.

Inputs and outputs

Not assessed

Autonomy

Not assessed

Workflow

Thirty-two gamified linguistic exercises administered in a browser. A labeled sample of 3,644 children, 392 with professional diagnoses, gathered through 113 organizations. A random-forest classifier trained on performance measures.

Human oversight

Not assessed

Relation to AI

The random-forest classifier is the screening decision. The 2026 re-analysis shows that demographic variables alone reach nearly the same AUC in this dataset and that false-positive rates differ twelve-fold between language groups. So the measured accuracy may partly reflect sample composition.

AI contribution

The random-forest classifier is the screening decision. The 2026 re-analysis shows that demographic variables alone reach nearly the same AUC in this dataset and that false-positive rates differ twelve-fold between language groups. So the measured accuracy may partly reflect sample composition.

Benefits

Outcome identifiers

Not assessed

Benefit

79.8% accuracy with 80.6% recall and 79.3% precision

Category

Improved Task Accuracy

Status

Not assessed

Beneficiary groups

People. Organizations.

Beneficiaries

Not assessed

Reach

Not assessed

Distribution and access

Not assessed

Duration

Not assessed

Expected?

Yes

Costs and harms

Not assessed

Measurement

Studies

Not assessed

Study design

Not assessed

Measurement

Not assessed

Setting and sample

Not assessed

Measures

79.8% accuracy with 80.6% recall and 79.3% precision

Comparison

Not assessed

Results

79.8% accuracy with 80.6% recall and 79.3% precision

Additional value from AI

Not assessed

Limitations

Not assessed

Denominators

Not assessed

Materials for replication

Not assessed

Money

Kind of value

Not assessed

Amount

Not assessed

Currency and prices

Not assessed

Period and unit

Not assessed

Realization

Not assessed

Recipients and payers

Not assessed

Gross or net

Not assessed

Included costs

Not assessed

Comparison

Not assessed

Method

Not assessed

Return and payback

Not assessed

People and places

Participants

Developer: Luz Rello and co-authors (IE University, Northeastern, University of Washington, Carnegie Mellon, University of Barcelona), per the paper's author list. Operator: Schools, dyslexia centers and non-profits in Argentina, Chile, Colombia, Spain and the USA (study participants); public online use in Spanish-speaking countries.

Jurisdictions

Not assessed

Ownership and control

Not assessed

Dependencies

Not assessed

Cross-border connections

Not assessed

Geopolitical implications

Not assessed

Follow-up

Conditions for replication

Samples balanced on native language and other demographics, which the 2026 re-analysis finds this dataset is not. Professional diagnosis as the reference standard. Spanish-language content for Spanish-speaking populations.

Replication attempts

1 follow-up entries. Benefit did not recur: 1.

Later findings

23 September 2026. An independent re-analysis of the same dataset reported that four demographic columns alone nearly match the 192 task metrics' discrimination and that including them yields a twelve-fold false-positive disparity between language groups, undercutting the screener's reported performance.

Availability

Not assessed

Institutional decision

Not assessed

Grounds

Not assessed

Action and result

Not assessed

Success criterion

Not assessed

Evidence

Source numbers

AIBD-0221-source-0001–0004

Citations

Numbered references below.

Language

Not assessed

Source relationships

Developer, Independent

Original evidence

Not assessed

Access and versions

Not assessed

Supported findings

Not assessed

Evidence strength

Not assessed

Missing evidence

Not assessed

Grades

Assessed benefits

People. Organizations.

Depth

Not assessed

Reasons

People: No individually attributable outcome: the sources report classifier accuracy (80.6% recall) and aggregate use (over 200,000 people), not what followed a screening result. Organizations: No outcome documented for the 60 participating schools or the other organizations beyond supplying participants.

Observed or intended

Not assessed

Rubric

Not assessed

Assessor and date

Not assessed

Alternative assessments

Not assessed

Fill

Independently studied

Entry details

Kind

Not assessed

Field

Education / learning-disability screening. School-age dyslexia screening.

Purpose

Dytective was designed to screen children for dyslexia risk cheaply online. The accuracy figures and usage are that purpose.

Dates

Date basis unassessed: 2020 (precision unassessed). Updated: 21 October 2026.

Stage

Not assessed

Version

SHA-256: e723d6e44874…

References

Other identifiers

Not assessed

Category mappings

Not assessed

Brief

Not assessed

Revisions

4 revisions. Latest: 21 October 2026.

Sources

  1. Peer-reviewed study Developer

    Rello L, Baeza-Yates R, Ali A, Bigham JP, Serra M (2020). Predicting risk of dyslexia with an online gamified test. PLoS ONE 15(12): e0241687 (PMC7710040).

  2. Peer-reviewed study Independent author

    Sudha M, Murali S (2026). GRAPHOS re-examined: graph-theoretic features do not improve dyslexia screening, and a demographic confound in the Dytective dataset. Frontiers in Digital Health 8; doi 10.3389/fdgth.2026.1927725 (PMC13656794).

  3. Peer-reviewed study Developer

    Rello L, Baeza-Yates R, Ali A, Bigham JP, Serra M (2020). Predicting risk of dyslexia with an online gamified test. PLoS ONE 15(12): e0241687 (publisher page).

  4. Peer-reviewed study Independent author

    Sudha M, Murali S (2026). GRAPHOS re-examined: graph-theoretic features do not improve dyslexia screening, and a demographic confound in the Dytective dataset. Frontiers in Digital Health 8; doi 10.3389/fdgth.2026.1927725 (publisher page).

Collections: Scrutiny and findings.