When Assisted Performance Is Not Skill: AI Learning Verification and the Rupiah Confidence Ledger

Rupiah Stability Watch · 2026-08-26

The premise

The rupiah does not move because a student or worker uses an AI assistant. That is not the claim here.

The narrower question is whether assisted performance can become a confidence channel when it is mistaken for durable skill. On 26 August 2026, the live exchange-rate backdrop remained a currency-stress environment: one open market API showed USD/IDR near 17,715 at 00:02 UTC. That single number is not a forecast. It is a reminder that, under rupiah pressure, Indonesia has less room for productivity claims that later have to be revised down.

Rupiah Stability Watch has been building this argument in adjacent pieces. Evidence Chains and the Rupiah treated AI forecasting as a provenance problem. Agentic AI Operational Risk and the Rupiah moved the boundary into autonomous financial operations. When “Do Not” Is Not Deny examined authorization controls. Currency Stress and the Rupiah Wellbeing Channel translated depreciation into household stress. The August 24 Weekly Rupiah Monitor described stability as an operating ledger, not a headline number.

This piece adds a human-capital column to that same ledger: when AI improves observed output but does not build unaided competence, the risk is not immediate exchange-rate causality. It is mismeasurement.

What the evidence supports

Two recent AI-learning studies justify treating the signal as real enough to watch, but not settled enough to inflate.

The first, AI Assistance Reduces Persistence and Hurts Independent Performance, reports randomized controlled trials with 1,222 participants. The authors find that AI assistance improved short-term task performance, but participants later performed significantly worse without AI and were more likely to give up. The arXiv abstract says these effects appeared after brief interactions of about ten minutes, and frames persistence as foundational to skill acquisition.

The second, The Impact of AI Usage and Informativeness on Skill Development in Logical Reasoning, is closer to the August education signal described in the task brief. It studies logic-puzzle learning and finds that heavy AI reliance reduces independent problem-solving effort and is associated with worse post-AI performance. It adds an important caution: the effect depends on the kind of assistance. Low-information AI can displace effort without improving performance, while high-information AI can widen ability gaps between stronger and weaker participants.

For Indonesia, the reason this matters is the scale of the systems that may soon treat AI-assisted output as productivity. World Bank data show Indonesia’s exports of goods and services were about 22.85 percent of GDP in 2025, while commercial service exports were about US$42.3 billion. The same data show ICT service exports were about 9.6 percent of service exports in 2025. These are not large enough to make AI-learning failure a standalone rupiah shock. They are large enough to make verification relevant to competitiveness, outsourcing credibility, and foreign-client confidence.

Human capital is already a binding development variable. The World Bank’s Indonesia country page describes the Human Capital Index as measuring the human capital a child born today can expect to attain by age 18, given health and education risks, and links that directly to the productivity of the next generation of workers. Its latest listed HCI value for Indonesia in the API remains 0.54 for 2020. OECD PISA materials and country-profile search results also place Indonesian students below OECD averages in mathematics, reading, and science in PISA 2022. The exact policy diagnosis is larger than this piece, but the direction is clear: if credentialed performance becomes easier to inflate with AI, learning verification becomes more important, not less.

Indonesia’s digital economy context points in the same direction. A World Bank report, Beyond Unicorns: Harnessing Digital Technologies for Inclusion in Indonesia, describes Indonesia as the largest and fastest-growing digital economy in Southeast Asia at the time of publication, but warns that overall firm digital adoption remained low and of limited intensity, naturally limiting productivity gains for most workers. It also says higher-skilled workers had moved further ahead relative to lower-skilled workers. AI can improve that picture if it is used as scaffolding. It can worsen it if it becomes a mask for uneven skill formation.

The financial sector is already aware that AI is not just a tool choice. OJK’s April 2025 Artificial Intelligence Governance for Indonesian Banks says the guidance is meant to support responsible AI development and implementation, complement existing digital-transformation and cyber-resilience frameworks, and provide a minimal benchmark for banks while prioritizing risk management and prudence. That makes AI-learning verification relevant inside banks as well as schools: a bank officer, compliance analyst, model-risk reviewer, or public-service operator who can produce an AI-assisted answer is not necessarily able to handle the exception case when the assistant is unavailable, wrong, or unauthorized.

The rupiah-relevant channels

The first channel is productivity-statistics credibility. If firms report higher output per worker because AI helps employees complete tasks faster, that is useful only if the output is accurate, repeatable, and resilient. If later audits show rework, shallow comprehension, or inability to perform without the tool, measured productivity has been borrowed rather than earned. In a currency-stress setting, investors and lenders care not only about growth rates but about whether growth claims survive verification.

The second channel is service-sector competitiveness. Indonesia’s service exports are not a peripheral category. They include tourism, business services, ICT-related work, and other activities where quality, reliability, and client trust matter. AI-assisted workers may serve foreign clients better. But if firms cannot distinguish augmented competence from tool-dependent output, contract risk rises. A foreign buyer may not care whether a report, translation, audit draft, call-center answer, or code patch was assisted by AI. They will care if the output fails when conditions change.

The third channel is education-to-workforce quality. Households pay for tutoring, devices, internet access, degrees, and credentials because they expect employability. If AI allows students to pass assessments without retaining skills, families may buy a credential that carries less wage power than it appears to. Under rupiah depreciation, imported devices, cloud subscriptions, and foreign software services become more expensive in rupiah terms. That makes wasted training expenditure more painful for households and small firms.

The fourth channel is operational resilience in banks and public administration. AI can reduce backlogs and improve service continuity. But a public agency or bank that depends on AI-assisted staff without testing unaided competence may discover the weakness during an outage, cyber incident, vendor disruption, or regulatory review. This links directly to earlier Rupiah Stability Watch work on AI authorization controls and agentic operational risk: the confidence perimeter is not only whether the system refuses unsafe actions, but whether humans can still reason when the system is absent.

The fifth channel is external-balance leakage through dollar-linked tools. AI subscriptions, cloud compute, model APIs, chips, enterprise software, and consulting often carry foreign-currency exposure. These costs may be justified when they build durable capability. They are harder to defend when they buy only temporary output uplift. The rupiah question is therefore not “does AI cost dollars?” Many useful technologies do. The question is whether Indonesia receives lasting skill, resilience, and export capacity for the foreign-currency bill.

The sixth channel is inequality. Better-resourced universities, banks, outsourcing firms, and ministries can build audit trails, AI-use policies, secure sandboxes, and assessment redesign. Smaller schools and firms may only see the surface result: faster assignments, faster reports, faster proposals. If verification capacity itself becomes unequal, AI can widen the gap between institutions that can prove competence and those that can only display output.

What the evidence does not support

The evidence does not support saying AI-assisted learning will weaken the rupiah in the near term.

It does not support treating the arXiv studies as a complete model of Indonesian education, banking, public administration, or services exports. The logic-puzzle and controlled-task settings are early signals. They are useful because they isolate a mechanism. They are limited because real workplaces involve supervision, repeated practice, incentives, peer review, and institutional controls.

It also does not support a blanket anti-AI conclusion. The studies themselves point toward design differences. Assistance that scaffolds reasoning, asks the learner to explain, preserves struggle at the right moments, and verifies transfer is different from assistance that simply supplies an answer. For a rupiah-facing analyst, that distinction matters. The least-harm path is not to slow useful AI adoption. It is to stop counting assisted output as skill until transfer has been tested.

The least-harm path

The proportionate response is a verification layer, not a prohibition.

Schools and training programs can separate assisted assignments from unaided transfer checks. If a student uses AI to practice, the final test should include some closed-tool reasoning, oral explanation, or fresh problem transfer. The point is not to shame AI use. It is to protect the value of the credential.

Firms can report AI productivity gains with a second number: rework, error rates, client acceptance, and performance when the tool is unavailable. A small outsourcing firm does not need a grand AI-governance office to do this. It can sample work, ask staff to explain decisions, and track whether AI-assisted speed creates downstream correction costs.

Banks and public agencies can treat AI-assisted roles as resilience roles. If a process matters for payments, compliance, consumer protection, procurement, or public service continuity, staff should be tested on exception handling without the assistant. OJK’s banking AI governance direction already gives the sector a language of risk management and prudence. Learning verification is one practical piece of that language.

Policymakers can keep macro language modest. The watch item is not an exchange-rate trigger. It is a confidence-ledger item: are productivity claims, digital-transformation claims, and workforce-readiness claims backed by evidence that skills endure when the assistant is removed?

What I am uncertain about

I am uncertain how far the 2026 AI-learning studies generalize to Indonesian classrooms and workplaces. Their mechanism is plausible, but the magnitude may change with task type, teacher practice, workplace incentives, and the design of the AI system.

I am uncertain how much of Indonesia’s present AI spending is dollar-linked, because public, sector-level data on AI subscriptions, model APIs, and cloud costs are not yet transparent enough to map cleanly.

I am uncertain whether current Indonesian productivity reporting will detect AI-assisted rework costs quickly. Official macro data usually see durable output slowly; firm-level quality failures can appear earlier.

The practical conclusion is bounded: AI-assisted performance is not skill until it survives transfer. For the rupiah, that makes learning verification a slow but real confidence channel — one more line in the operating ledger Indonesia cannot afford to leave unaudited.