A discrete choice experiment evaluating how parents in Kenya evaluate artificial intelligence tools designed for early childhood education reveals that families prioritize rigorous safety controls and core academic subjects while strongly rejecting designs that demand ongoing parental co-use. The investigation, conducted by researchers Antony Mbithi and Loise Maina, analyzed 10,824 discrete decisions made by 451 parents of children aged zero to eight years to quantify trade-offs among pedagogical content, interface interaction styles, supervisory mechanisms, technical assistance, and subscription pricing. The experimental findings indicate that parents exhibit their highest willingness-to-pay for comprehensive parental control mechanisms, valued at 3,598 Kenyan shillings, or approximately 27.68 United States dollars per month, followed closely by mathematics and problem-solving curricula at 3,185 shillings, roughly 24.50 dollars per month. Conversely, early learning configurations requiring mixed interaction, where children engage digital tools under continuous parental supervision, suffered a substantial penalty of negative 1,776 shillings monthly.
The study addresses a critical empirical void in low- and middle-income countries, where the rapid expansion of educational technology intersects with persistent socioeconomic disparities and high-stakes schooling systems. While early childhood education across birth to age eight represents a formative developmental stage with documented long-term cognitive and socio-emotional returns, access to quality resources remains profoundly unequal. Digital applications powered by artificial intelligence promise personalized instruction, yet they simultaneously introduce unique hazards for young learners, including automated behavioral tracking, unverified algorithmic recommendations, and the displacement of essential interpersonal contact. Grounded in the Technology Acceptance Model, Parental Mediation Theory, and Lancaster’s Consumer Choice Theory, the inquiry examines how caregivers function as decisive gatekeepers. Parents must continually negotiate the perceived developmental utility of software against technical vulnerabilities, digital literacy limitations, and strict household budgetary boundaries, making rigorous preference modeling essential for guiding equitable product deployment across emerging digital education markets.
To capture these latent trade-offs without the confounding distortions of unconstrained survey ratings, the researchers structured a discrete choice experiment grounded in Random Utility Theory. An initial candidate pool of ten potential attributes derived from developmental psychology and technology adoption literature was refined through a structured pretest of thirty participants to balance statistical power with respondent cognitive load. Five focal attributes were selected, each defined across three distinct levels: child interaction mode, educational content focus, parental control mechanisms, support provision, and monthly subscription cost. The experimental design established fully interactive operation, reading and vocabulary content, low tool-managed oversight, and zero support as baseline reference categories. Subscription fees were specified continuously across zero, 2,500, and 6,000 Kenyan shillings per month. Using a fractional factorial design generated via Stata’s DCREATE procedure, the authors optimized orthogonality and D-efficiency, constructing twelve paired choice scenarios containing two unlabeled product profiles without an opt-out alternative.
The empirical sample comprised 451 caregivers recruited through pre-primary education centers, community organizations, parental associations, and digital platforms across Kenya, drawn from an initial pool of 506 recruited individuals to yield an effective response rate of approximately ninety percent. To prevent the systematic exclusion of digitally marginalized households, data collection combined the online Kobo Toolbox platform with in-person enumerators who read choice tasks aloud and recorded responses on digital tablets. The final participant cohort had an average age of thirty-three years and consisted of 51.7 percent female and 48.3 percent male respondents. Formal educational attainment varied widely, including 2.2 percent with primary schooling, 38.1 percent holding bachelor degrees, and 3.3 percent possessing postgraduate credentials. Household monthly incomes ranged from under 24,000 shillings for 23.1 percent of families to over 100,000 shillings for 18.4 percent, while geographic representation spanned 46.6 percent urban, 29.3 percent suburban, and 24.2 percent rural environments.
The statistical architecture relied on McFadden’s conditional logit framework alongside a panel mixed logit specification to evaluate choice probabilities and unobserved preference heterogeneity across the balanced panel of 10,824 decisions. While the standard conditional logit model established baseline parameters under the assumption of independent and identically distributed extreme value error terms, the panel mixed logit specification modeled non-price attributes as normally distributed random parameters while holding the cost coefficient fixed. This fixed cost parameter permitted consistent derivation of marginal rates of substitution and willingness-to-pay calculations using the Delta method for standard error estimation. Maximum likelihood estimation for the mixed logit utilized 500 Halton draws per respondent to accurately approximate multidimensional probability integrals. Estimated price coefficients emerged consistently negative and statistically significant at minus 0.000235 in the conditional logit model and minus 0.000265 in the mixed logit specification, confirming that escalating subscription costs consistently suppress tool adoption probabilities.
Estimation results demonstrate that pedagogical orientation and supervisory oversight constitute the primary determinants of parental utility. Average marginal effects indicate that the incorporation of mathematics and problem-solving material elevates the likelihood of product selection by 16.9 percentage points relative to creative and imaginative content, generating a regression coefficient of 0.749 in the conditional logit model. Reading and vocabulary development produced a smaller positive shift of 7.4 percentage points with a coefficient of 0.321 and a willingness-to-pay of 1,365 shillings. Demand for protective governance proved even more decisive: moderate parental control increased choice probability by 10.9 percentage points with a coefficient of 0.465 and a valuation of 1,977 shillings, whereas comprehensive control surged choice probability by 19.6 percentage points with a coefficient of 0.846. The mixed logit model revealed substantial unobserved variance around both mathematics content and comprehensive oversight, demonstrating that while average demand remains elevated, underlying parental valuations diverge sharply.
In sharp contrast to prevailing developmental recommendations advocating joint media engagement, configurations requiring mixed interaction elicited intense parental disutility. The requirement for collaborative adult-child participation reduced adoption probability by 9.8 percentage points, imposing a severe monetary discount penalty of negative 1,776 shillings, which equates to roughly seventy-one percent of the benchmark 2,500-shilling subscription tier. Conversely, passive video and audio delivery generated a statistically negligible effect of minus 0.8 percentage points and an insignificant willingness-to-pay of negative 154 shillings. The researchers attribute this mixed-mode penalty to severe household time constraints, irregular work schedules, and boundary ambiguity surrounding instructional responsibility. Meanwhile, technical assistance exhibited clear diminishing returns, termed the support plateau: basic operational tips increased choice probability by 7.2 percentage points with a willingness-to-pay of 1,287 shillings, while comprehensive expert guidance yielded an almost identical gain of 6.9 percentage points and a comparable valuation of 1,247 shillings, highlighting that elaborate customer support infrastructures fail to generate proportional consumer value.
Affordability emerged as the decisive systemic constraint governing potential technology diffusion, exposing a stark equity paradox across socioeconomic strata. On average, each 1,000-shilling monthly fee increment diminishes selection probability by 5.6 percentage points, but subgroup analyses uncovered profound structural cleavages across income and formal schooling brackets. For low-income households earning below 24,000 shillings monthly, a 1,000-shilling increase suppresses adoption probability by 32.0 percentage points, even though these same families demonstrate the highest willingness-to-pay for core capabilities, including 3,101 shillings for mathematics and 3,575 shillings for comprehensive oversight. In stark opposition, high-income households proved statistically insensitive to price changes, assigning negative or negligible incremental valuations to the same features, including minus 487 shillings for mathematics and minus 1,376 shillings for high control. Economically advantaged parents effectively regard robust safeguards and structured academics as non-negotiable baseline standards rather than premium enhancements warranting supplementary payment, creating deep divisions across the early childhood technology landscape.
Heterogeneity analyses across geography, parental gender, and child age further illuminated the fragmented architecture of early educational technology markets. Urban parents expressed the highest recorded valuation for full supervisory control at 6,440 shillings, reflecting heightened digital risk awareness and unsupervised device access, while non-urban caregivers prioritized mathematics at 3,442 shillings, valued comprehensive assistance at 2,736 shillings, and penalized mixed interaction at negative 3,882 shillings. Mothers exhibited greater demand for full control at 4,513 shillings and reading content at 1,768 shillings compared to fathers, who valued control at 2,637 shillings and reading at 946 shillings. Educational divides proved equally severe: caregivers with lower educational attainment faced a 47.3 percentage-point adoption decline per 1,000-shilling price increase while expressing strong attribute demand, contrasting with a minor 7.3 percentage-point decline among highly educated respondents. Furthermore, parents of older children aged five to eight placed higher valuations on mathematics and control than parents of infants and toddlers.
The authors emphasize critical methodological and contextual limitations, noting that hypothetical discrete choice experiments capture stated preferences rather than revealed market behavior, and the omission of an opt-out choice prevents estimation of unconditional market penetration rates. Furthermore, cross-sectional designs cannot observe preference evolution as parents gain direct operational familiarity, while assisted sampling still underrepresents completely disconnected households. Crucially, parental preference patterns favor exam-oriented academic drills over play-based pedagogies, creating tensions with developmental science guidelines that recommend holistic socio-emotional enrichment. Because private market incentives threaten to price out economically vulnerable households who express the strongest functional demand, commercial mechanisms alone risk exacerbating educational stratification. The researchers conclude that mitigating these risks requires intentional interventions, including targeted public subsidies, tiered subscription structures, certified open-access tools, enforceable child data privacy regulations, and software architectures that alleviate supervisory burdens rather than imposing impractical domestic co-use mandates.
Subject of Research: Parental preferences and willingness to pay for AI-powered early childhood education tools in Kenya
Article Title: Parental preferences for AI-powered early childhood education tools: a choice experiment
Article References: Mbithi, A., & Maina, L. (2026). Parental preferences for AI-powered early childhood education tools: a choice experiment. AI, Brain and Child, 2(1), Article 1. https://doi.org/10.1007/s44436-025-00025-z
Image Credits: AI Generated
DOI: 10.1007/s44436-025-00025-z
Keywords: early childhood education, artificial intelligence, discrete choice experiment, educational technology, parental mediation, willingness to pay, Kenya, digital divide, Parental, preferences, AI-powered, early
News Source: Courtney Benton. (November 7, 2026). Kenyan Parents Value AI Safety and Math Over Shared Screen Time. Scienmag.
