Industries / Government, education and community / Childcare
Early learning
Early learning is what children under school age learn in care and preschool, the curricula that guide it, and the studies that measure what it changes later in life.
7
Milestones achieved, from 1907 to 2022.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1907
A school inside the tenement
Edoardo Talamo, director general of the Roman Association for Good Building, invited Maria Montessori to organize schools inside its model tenements in Rome, each gathering the children aged three to seven who lived in the building into one large room with a teacher. The first Casa dei Bambini, or Children's House, opened at 58 Via dei Marsi in the San Lorenzo quarter on 6 January 1907, with a teacher, Candida Nuccitelli, under Montessori's direction, and a second opened in San Lorenzo that April.
Credit
Maria Montessori; Edoardo Talamo
1960s
Perry Preschool, followed to 40
The HighScope Perry Preschool Program, run at the Perry Elementary School in Ypsilanti, Michigan, in the early 1960s, gave African-American children from poor families two years of 2.5-hour weekday classes from age three, plus a weekly home visit by the teacher. The 123 children in the program and control groups were followed to age 40, with records of earnings, schooling and crime. Earlier studies put the yearly return to society at 16% or 17%. James Heckman and colleagues allowed for the fact that some children were swapped between the groups after the random draw and put the return at 7% to 10% a year, still above the historical return on shares.
Lessonallowing for flaws in how the groups were formed cut the estimated return roughly in half, though it stayed positive.
Credit
HighScope (the program); James Heckman, Seong Hyeok Moon, Rodrigo Pinto, Peter Savelyev and Adam Yavitz (the 2009 reanalysis)
1993
Argentina builds preschool classrooms
Argentina's 1993 Federal Education Law made the last year of preschool compulsory. From 1993 to 1999 the national government paid for 3,531 new preschool classrooms, enough for about 175,000 more children aged 3 to 5, and built most of them in poor provinces with few places. Comparing regions and birth years, economists estimated that one year of preschool raised third-grade test scores in Spanish and mathematics by 8% of the average score, and improved attention, effort, class participation and discipline.
Credit
the Argentine national government; study by Samuel Berlinski, Sebastian Galiani and Paul Gertler
1996
A bicultural national curriculum
New Zealand's Ministry of Education published Te Whāriki, the national early childhood curriculum, in 1996. It is play-based and bicultural, and is organized as principles, strands (broad areas of learning), goals and learning outcomes; the principles, strands and goals are gazetted, that is, published as official rules. The 2017 update cut the learning outcomes from 118 to 20, so that each service could build its own local curriculum around them, and published a separate curriculum for kōhanga reo (Māori-language centers) in the same document.
Credit
New Zealand Ministry of Education
2010
Head Start gains fade
Head Start is the US federal preschool program for children from low-income families. For a study ordered by Congress, 4,667 newly entering three- and four-year-olds at 383 randomly chosen centers in 23 states were assigned by lottery, from fall 2002, either to Head Start or to a control group free to use other care. At the end of the Head Start year the four-year-olds were ahead on six language and literacy measures, but by the end of first grade only a few differences remained, such as one vocabulary score.
Lessongains measured at the end of preschool can mostly disappear within two years, a pattern researchers call fade-out.
Credit
US Administration for Children and Families; Westat and partner research firms
2012
Village preschools in Mozambique
From 2008 Save the Children ran community-built preschools in 30 villages of Gaza Province, chosen at random from 76 eligible communities; villagers supplied the labor and local materials, and teachers got a stipend of US$10 a month. About two years later, children who had attended were 24% more likely to be enrolled in primary school than children in the other villages, and scored better on cognitive, problem-solving, fine-motor and social measures, though not on communication and language, which stayed very low in both groups. The authors put the cost at US$29.1 per child for a 12-month program, and call the study, to their knowledge, the first randomized evaluation of a preschool program in rural Africa.
Credit
Save the Children; evaluation by Sebastian Martinez (Inter-American Development Bank), Sophie Naudeau (World Bank) and Vitor Pereira (PUC-Rio)
2022
Tennessee pre-K falls behind
Tennessee's state-funded pre-K had more applicants than places at some sites, so 2,990 children from low-income families were offered a place or put on a waiting list at random. The children offered pre-K had lower state test scores than the control children in third through sixth grade, with the largest gap in sixth grade, and more disciplinary infractions, absences and special-education placements.
Lessona statewide program can leave children worse off on later measures, so results from small model programs such as Perry cannot be assumed at scale.
Credit
Kelley Durkin, Mark Lipsey, Dale Farran and Sarah Wiesen, in Developmental Psychology
Watch for
AI prompts for parent talk
University of Chicago researchers registered a randomized trial of Chat2Learn-AI with 1,000 low-income families of children aged 3 to 6 in California's Coachella Valley Unified School District. Half get three to six illustrated conversation prompts a week by text message, plus a chatbot built on large language models that writes prompts around each child's interests, meant to start more back-and-forth talk at home; the intervention runs from December 2026 to May 2027. The team says the trial builds on a successful pilot, but publishes no pilot results.
Achieved whenthe trial reports a gain in children's oral language on its primary test, the Woodcock-Johnson IV oral language cluster, over the control group.
Claimed by
Mauricio Koechlin, Ariel Kalil, Michelle Michelini and Linxi Lu, University of Chicago, July 2026
England's 75% readiness goal
England's Department for Education set the ambition that 75% of five-year-olds reach a "good level of development" by 2028, meaning the expected level in 12 early learning goals covering language, personal and social development, physical development, literacy and mathematics. The figure for 2025/26 was 72.3%, up from 68.3% a year earlier. The measure is a judgment by each child's own teacher, and the department also brought in new support for making these assessments.
Achieved whenthe results for 2027/28 reach 75%, and a test not scored by the children's own teachers shows a matching rise.
Claimed by
UK Department for Education, July 2025
Open
01
Safety test for activity plans
A public set of at least 300 made-up requests that a nursery worker might give an AI assistant, such as "a 15-minute sensory activity for four 18-month-olds", with a rubric written by early years teachers and a child safety specialist. The rubric marks each AI-written plan for age fit, choking and small-part hazards, the supervision it needs, and whether it serves a named goal in a real curriculum such as England's early learning goals or Te Whāriki. A script scores any model against it. No real child is involved.
The briefCredit · unclaimed02
Learning goals side by side
A machine-readable table that lists every goal or learning outcome in at least 12 national early learning frameworks, for example England's 17 early learning goals, New Zealand's 20 Te Whāriki outcomes and Ghana's Early Learning and Development Standards, each with its page reference and a label from one shared list of learning areas, so that anyone can see what each country expects of a five-year-old and what one leaves out.
The briefCredit · unclaimed
Open 01 · the brief
Safety test for activity plans
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of at least 300 made-up requests that a nursery worker might give an AI assistant, such as "a 15-minute sensory activity for four 18-month-olds", with a rubric written by early years teachers and a child safety specialist. The rubric marks each AI-written plan for age fit, choking and small-part hazards, the supervision it needs, and whether it serves a named goal in a real curriculum such as England's early learning goals or Te Whāriki. A script scores any model against it. No real child is involved.
Why it is still open
The nearest benchmark, SproutBench (2025), covers ages 0 to 6 among its age groups, but it tests model replies to 1,283 adversarial prompts about risks such as emotional dependency and privacy, not activity plans written for carers. Searches found no benchmark that rates AI-written activity plans for children under five for physical hazards.
Built with what exists
Commercial AI models and small open ones that can be adapted on one graphics chip, such as Liquid AI's LFM3 (see the Note "One GPU to fine-tune"); Gemini Decision 1 to score many fixed-answer rubric items in one pass, checked against human raters (see the Note "A thousand questions, one pass"); a spreadsheet for the human ratings.
First step
Write 50 requests and a first rubric with two early years teachers, run them through two models, and rate the results by hand.
Done looks like
300 requests across five age bands from 6 months to 5 years; every plan from at least five models rated by two people, with agreement (Cohen's kappa, a standard measure of rater agreement) of at least 0.7; a published table of the share of plans per model that contain a choking or small-part hazard.
Open 02 · the brief
Learning goals side by side
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A machine-readable table that lists every goal or learning outcome in at least 12 national early learning frameworks, for example England's 17 early learning goals, New Zealand's 20 Te Whāriki outcomes and Ghana's Early Learning and Development Standards, each with its page reference and a label from one shared list of learning areas, so that anyone can see what each country expects of a five-year-old and what one leaves out.
Why it is still open
The main cross-country early childhood dataset, UNICEF's country profiles for 197 countries, records services, outcomes and policies, such as whether a country has an officially adopted early childhood policy, not the content of its learning goals. National standards exist as separate documents, such as Ghana's, which its Education Service developed with UNICEF in 2004. Searches found no goal-by-goal comparison across countries in a reusable format.
Built with what exists
A long-context AI model to pull each goal out of the framework documents; Gemini Decision 1 to assign each goal to one of a fixed list of learning areas (see the Note "A thousand questions, one pass"); two human coders for a check sample; CSV files published under an open license.
First step
Code England's early learning goals and Te Whāriki's 20 outcomes by hand into a shared table, and use them to fix the list of learning areas.
Done looks like
At least 12 frameworks from at least four continents, every goal in a row with its source page; two people code a random 10% of rows independently, with Cohen's kappa of at least 0.7 against the published labels; a one-page summary of goals found in one framework and missing from the others.