AI evidence extraction

Policy, Risk and Innovation: A Mixed-Methods Framework for Using AI to Foster Inclusion in Marginalized Communities in Bangladesh

Ritesh Karmaker, Vladimir M. Cvetković · Preprints.org · 2026

AI-generated evidence extraction, verified across multiple analytical personas. Not a substitute for the peer-reviewed original.

5/10

Relevance

0/4

Quality (LMQS)

Evidence

Citations

This is an AI analysis. Read the peer-reviewed original at the publisher: https://doi.org/10.20944/preprints202603.0981.v1

Methodology & findings

Study design

Convergent mixed-methods design integrating quantitative student survey (N = 213 across seven institutions) with qualitative data from 37 stakeholders (teachers and policymakers) collected through semi-structured interviews and focus group discussions, conducted March–September 2024..

Sample

N = 250, 6 groups

Primary method

Multivariable regression analysis to identify predictors of composite educational outcome score. Disparity analyses for urban-rural inequities. Mixed-methods integration of quantitative survey data with qualitative interview and focus group data. Specific software packages not stated in abstract.

Main result

The study found that "AI tool adoption was the strongest predictor of a composite educational outcome score (β = 0.38, p < 0.001), followed by institutional support (β = 0.25, p = 0.01)". Additionally, "the policy implementation gap—defined as the mismatch between policy intent and on-the-ground delivery—was negatively associated with outcomes (β = −0.12, p = 0.04)", and "the model demonstrated strong explanatory power (R² = 0.67; F(4, 208) = 42.3; p < 0.001)".

Reports effect sizes and confidence intervals.

Research paradigm

mixed_methods (pragmatism)

Author conclusions

The authors conclude that "the paper proposes a policy–innovation framework centered on localized AI toolkits, sustained teacher upskilling, device-access interventions, and enforceable fairness and transparency safeguards to advance equitable learning opportunities". The framework is grounded in findings showing that "three binding constraints" emerge from qualitative analysis: "limited teacher AI preparedness, affordability barriers, and trust concerns related to privacy and algorithmic bias".

Risk of bias

Potential selection bias: convenience sampling of seven institutions in Sherpur Sadar Upazilla may not be representative; Temporal bias: data collection over 6-month period (March–September 2024) may not capture seasonal variation; Measurement bias: composite educational outcome score construction not detailed in abstract; Confounding: unmeasured variables affecting both AI adoption and outcomes; Selection bias: Study limited to Sherpur Sadar Upazilla, which may not be representative of all marginalized communities in Bangladesh; Potential social desirability bias in qualitative interviews with teachers and policymakers; Self-selection bias in student survey participation; Cross-sectional design limits causal inference despite regression analysis; Selection bias: institutions and participants sampled from single upazilla (Sherpur Sadar) in Bangladesh; Cross-sectional design limits causal inference; Self-report bias in student survey responses

Open questions raised

The authors identify the need for interventions addressing: (1) teacher AI preparedness and upskilling; (2) affordability barriers to device and internet access; (3) trust and transparency concerns around privacy and algorithmic bias; (4) mechanisms to bridge the policy implementation gap between policy intent and ground-level delivery.
Implementation of localized AI toolkits tailored to marginalized communities
Sustained teacher upskilling programs for AI adoption
Device-access interventions to address digital divide
Development of enforceable fairness and transparency safeguards
Longitudinal tracking of policy implementation fidelity

Data: not_statedCode: not_statedExtracted from: pdfAgreement 58%

Explore related topics

Responsible AI Use AI Governance Mixed Methods