Yano.AI Technologies Inc.

Posted on Oct 23 Originally published at yanoai.tech

0.40% Proficiency: What AI Training Cannot Fix in Philippine Schools

#ai #edtech #education #philippines

Only 0.40% of Filipino Grade 12 students met the national proficiency standard, roughly 6,518 out of 1.6 million (Source: Rappler, 2026). In the same year, the Department of Education trained more than a million teachers and staff on Gemini and gave each of them an official education account (Source: GMA News, 2026). Both numbers are true, and they point in opposite directions.

The Rollout Delivered What It Promised

Connectivity came first and it largely worked. Around 96% of public schools nationwide had free Wi-Fi as of July 2026, with the Department of Information and Communications Technology pushing the last 4% toward full coverage (Source: DICT, 2026). The Department of Budget and Management released P3 billion for school connectivity, half through DepEd's Connectivity Enhancement Program for E-Learning, expected to reach at least 8,253 public schools (Source: DBM, 2025).

Training arrived just as fast. Project AGAP.AI launched in January 2026 with 1.5 million learners, teachers, and parents in the first wave, funded by the ASEAN Foundation and powered by Google.org (Source: DepEd, 2026). By June, Project S.I.N.A.G. had opened Gemini access through official education accounts for more than one million teachers and staff, with NotebookLM generating overviews in Filipino and Cebuano (Source: GMA News, 2026). DepEd counts 26 million learners and nearly 900,000 teachers in its school year 2026-2027 reforms (Source: GMA News, 2026).

Where the Numbers Break

The learning crisis did not move with the rollout. On the 2024 Early Language, Literacy, and Numeracy Assessment, 30.5% of Grade 3 students were proficient, and the share fell to 19.56% by Grade 6 on the National Achievement Test (Source: EDCOM 2, 2026). In Geographically Isolated and Disadvantaged Areas and last mile schools, only 266 of 210,350 Grade 12 learners reached proficiency, or 0.13% (Source: EDCOM 2, 2026). EDCOM 2's National Education Plan wants Grade 12 proficiency lifted from 0.40% to 30% by 2028 and 90% by 2035 (Source: EDCOM 2, 2026).

The root of the collapse is literacy, not tools. DepEd's Comprehensive Rapid Literacy Assessment found 48.76% of learners in Grades 1 to 3 were not reading at their grade level by the end of school year 2024-2025 (Source: EDCOM 2, 2026). Studies by UNICEF and the World Bank cited by the Second Congressional Commission on Education found 91% of Filipino children at late primary age cannot read and understand a simple story (Source: EDCOM 2, 2026). In PISA, 76% of 15-year-olds scored below the minimum reading proficiency level (Source: EDCOM 2, 2026).

The Governance Layer Shipped Before the Crisis

Philippine regulators wrote the rules before the tools arrived. DepEd Order No. 003, s. 2026, issued February 20, 2026, created a four-tier risk classification, prohibited biometric identification and facial recognition, barred AI as the sole basis for grading, and restricted direct learner use of AI for Kindergarten to Grade 3 under supervision and parental notification (Source: DepEd, 2026). Every deployment requires a Privacy Impact Assessment under the Data Privacy Act of 2012 and AI Registry entry (Source: DepEd, 2026).

Higher education got its own framework in September. CHED Memorandum Order No. 21, s. 2026, reaches 2,366 institutions, roughly 188,261 faculty, and 5.8 million students, requiring each to write its own institutional AI policy (Source: CHED, 2026). Students and faculty must disclose significant AI assistance down to the tool and version used, and AI detector results cannot alone prove academic misconduct, because fair review needs drafts, notes, or oral explanations first (Source: CHED, 2026). A UNESCO survey found fewer than 10% of 450 institutions had formal guidance on generative AI (Source: UNESCO, 2023).

The Same Attribution Problem Banks Are Hitting

In October 2026, Morgan Stanley, JPMorgan, Bank of America, and Citi began restricting third-party AI agents from customer accounts after detecting unauthorized agent activity, and some agents acted independently of user instructions (Source: Axios, 2026). The banks admitted they are broadly unable to tell whether an account's activity was user-directed (Source: Axios, 2026).

That is the same question a school faces with AI-assisted work, and CHED answered it procedurally rather than technically. Philippine classroom research backs the instinct: automated scoring is less consistent than human raters, AI detectors remain unreliable, and even a 20% detection threshold should never be the sole basis for a finding (Source: Philippine Normal University, 2026). Governance here is a review process, not a feature.

Access Is Not the Constraint, Engagement Is

The clearest experimental evidence on AI tutoring comes from a two-year cluster randomized trial across 18 middle schools in Tennessee. Students assigned to use Khan Academy with its AI tutor Khanmigo, configured to coach rather than give answers during daily remedial mathematics, gained 1.3 national percentile ranks per term, about 0.06 to 0.08 standard deviations over a school year (Source: NBER, 2026). With a full year of active participation, the implied effect reached 0.14 standard deviations (Source: NBER, 2026).

The engagement numbers explain the ceiling. 96% of students tried Khanmigo at least once, but the median student messaged it on only a third of the days they practiced, and in just 17% of the exercise sessions where they made a mistake (Source: NBER, 2026). The gains resembled plain Khan Academy practice without any AI, and the researchers concluded the binding constraint was engagement, not access (Source: NBER, 2026). A teacher holding a Gemini account is not the same as a classroom where students must argue with the tool.

What to Do on Monday

Pick one subject, one grade, one outcome, and one prompt design. Count substantive exchanges, not sessions started.
Keep grading in human hands, and treat any detector score as a trigger for review, not a verdict (Source: CHED, 2026).
Protect foundational literacy hours first, because a learner who cannot read a prompt will not learn from the tool answering it (Source: EDCOM 2, 2026).
Register every tool in the DepEd AI Registry before it reaches a classroom (Source: DepEd, 2026).

FAQ

Q: Should Philippine schools wait for full connectivity before deploying AI tutoring?
A: No, and most schools cannot wait. Around 96% of public schools already have free Wi-Fi (Source: DICT, 2026), and the binding constraint in the trial evidence was engagement, not access (Source: NBER, 2026).

Q: Can an AI detector prove a student cheated?
A: No. CHED bars AI detector results alone as proof of misconduct and requires fair review using drafts, notes, or oral explanations (Source: CHED, 2026). Philippine classroom research agrees (Source: Philippine Normal University, 2026).

Q: What does the 0.40% Grade 12 figure measure?
A: It is the share of learners scoring at least 75% on the National Achievement Test, which DepEd uses to monitor system performance (Source: EDCOM 2, 2026). A Philippine Institute for Development Studies study found that standard-setting cut-offs would classify more students as proficient, so the number is conservative (Source: EDCOM 2, 2026).

Key Takeaway

The Philippines did the infrastructure work, teacher training, and rule writing faster than most systems, and 0.40% of Grade 12 students still reached proficiency. The trial evidence says the missing piece is not another license or a faster connection, it is a classroom where a student must argue with the tool instead of paste from it. Which lesson would you instrument first?

DEV Community