Product & Research

The Indonesian government, through its National Strategy for Artificial Intelligence, is preparing to integrate AI into the primary and secondary school (K-12) curriculum starting from the 2025/2026 academic year. However, the readiness of this policy faces major challenges on the ground:

• Infrastructure Gaps: Approximately 65% of schools in Indonesia do not yet have stable internet, and 35% lack a reliable electricity supply.
• Training Disparity: Access to teacher competency development remains uneven and is heavily centralized in urban areas. Furthermore, teachers’ AI literacy levels have not yet been accurately mapped.

To look at the actual readiness beyond policy documents, the research team surveyed 132 K-12 teachers using open-ended questions to elicit honest understandings from the educators.

Main Findings: “Know How to Use, But Misunderstood”

The study found a unique mismatch. Teachers are highly enthusiastic about adopting AI, yet their understanding of how this technology actually works remains very low.

Misconceptions About How AI Works

• Surface-Level Definition: 55.4% of teachers define AI merely as “technology that mimics humans” without knowing its underlying mechanisms.
• ChatGPT Understanding Error: When asked to explain the process by which ChatGPT generates answers, the majority gave ambiguous, normative responses. In fact, 21.5% of teachers mistakenly thought ChatGPT works like Google, meaning it searches and retrieves data directly from the internet/data centers (knowledge retrieval misconception), rather than predicting words based on language probabilities.

AI Adoption Patterns in Schools

Despite minimal theoretical understanding, AI adoption in schools has apparently been happening organically for the sake of work efficiency:

Moral Dilemma: Between Efficiency and Character Risks

The study recorded massive support: 96.2% of teachers support AI use for teachers and 84.6% support it for students. Teachers find it incredibly helpful because AI significantly cuts down their administrative workload.

However, behind this support, teachers harbor deep concerns (opposition risks) if AI is used without strict supervision:

• Threats to Academic Integrity: The greatest risk feared by teachers is plagiarism and academic cheating (23.8%). Teachers worry that students use AI as a “shortcut” to complete assignments without actually learning.
• Dependency and Mental Laziness: Teachers are anxious that this technology will erode students’ critical thinking skills because they become accustomed to passively receiving instant answers.
• Loss of the Human Touch: There is concern that over-reliance on AI could reduce the vital social-emotional interactions between teachers and students in the classroom.

Recommendations for Curriculum Policy

To ensure that the implementation of the new curriculum does not become a mere formality, this study suggests three strategic steps for the government:

1. Conceptual Literacy-Based Training: Teacher capacity-building programs must go beyond just teaching how to use the tools. Instead, they must explain the limitations, biases, risks of information hallucination, and the foundational mechanics of AI so that teachers can make sound pedagogical decisions.

2. Differentiated Support: The government must map out training equitably, providing greater assistance to schools outside of Java and in rural areas that lack infrastructure, ensuring the digital divide does not widen further.

3. Ethical Guidelines and School Governance: Schools need clear regulations regarding the boundaries of when students are allowed to use AI (for instance, for brainstorming or as a personal tutor) and when AI is strictly forbidden to maintain academic integrity.

Research Team: Alham Fikri Aji, Afifa Amriani, Rendi Chevi, Ayu Purwarianti, & Derry Wijaya.

Climate change is no longer just a prediction of the future, but a reality that we feel every day. Starting from the increasingly scorching air temperatures, erratic flash floods, to shifts in the planting season that confuse our farmers. However, amid the urgent need for real action, there is another major challenge lurking in the digital world: hoaxes and climate change misinformation.

In Indonesia, conversations around environmental issues are often obscured by false information. Some say that global warming is just a conspiracy, while others spread false claims about the causes of natural disasters. The impact is fatal. The public became skeptical, hesitant to act, and environmental rescue policies became difficult to receive full support.

Unfortunately, hoax makers are very smart. They not only spread fake news in standard Indonesian but also entered through regional languages that were closer to the hearts of the people. This is where the big problem lies: we lack digital tools or linguistic resources to detect hoaxes in regional languages. As a result, conventional hoax filtering systems often “escape” when reading texts outside the official Indonesian language.

Getting to Know NusaClimate: A New Weapon Against Hoaxes

Seeing this critical gap, a group of researchers in Indonesia AI Institute (IAII) is doing research to develop an innovative solution called NusaClimate.

NusaClimate is the first multilingual giant data collection (corpus) that was deliberately created to detect people’s attitudes or stances on the issue of climate change. This dataset collects 50,613 text data covering four languages at once:

• English
• Minangkabau Language
• Balinese
• Bugis Language

The presence of three regional languages (Minangkabau, Balinese, and Bugis) is very important because all three are classified as languages with minimal digital resources (low-resource languages). With NusaClimate, artificial intelligence (AI) now has an adequate “dictionary” to understand the local context in depth.

How Technology & Experiment Works

How is that much data processed into a hoax extermination system? The answer lies in a technology called the Encoder-based Language Model.

Think of this system like a highly sensitive language detective. When there is a new claim circulating on social media, the AI will perform a semantic comparison (word meaning). The system will match these claims to the premise (scientific facts or valid data in the NusaClimate dataset), even though they are written in different regional languages (cross-lingual).

Through this framework, the IAII researchers are in the progress of building a real-time climate misinformation checker tool that can be used by the wider community to directly filter which news is valid and which is a hoax right away.

Why Should AI Be “Trained” Again?

To ensure this AI works intelligently, the IAII researchers performed a process called Fine-Tuning. Why is this important?

The AI models are basically good at reading language in general, but they need to be “trained” specifically in order to understand environmental scientific terms and local slang related to climate. In this experiment, the researchers tested three popular giant language models:

1. IndoBERT (from IndoNLU) – Very good at understanding the structure of the formal Indonesian language.
2. IndoBERT-Nusa – An improved version to understand the language variations in the archipelago.
3. XLM-RoBERTa Large – A powerful international multilingual model in bridging different interlingual meanings.

The experiment was conducted through the supervised finetuning method, in which the AI is trained to use optimal hyperparameters for two main tasks: detecting the attitudes of the text to climate misinformation and grouping topics and subtopics around the environmental issue.

Aspect-Based Sentiment Analysis (ABSA) research is present as a scientific breakthrough to dissect public opinion in super detail (fine-grained) directly on the specific aspects of a review. By targeting the world’s best standards (SOTA), this research combines the richness of regional languages in Indonesia with the sophistication of the latest large language models (LLMs).

Have you ever read internet reviews that have mixed content?

“The hotel is very clean and the mattress is soft, but unfortunately the restaurant food is bland and the reception service is sluggish.”

For humans, we know these consumers love the room amenities but are disappointed with the food and service. However, traditional AI (Artificial Intelligence) will be confusing. The old AI could only read one whole sentence and then guess one label: Positive or Negative. Because the content is the opposite, the old AI usually gives up and labels it Neutral. As a result, hotel owners lose valuable information about which parts need to be repaired.

To bridge this gap, a cutting-edge research is developing by Indonesia AI Institute (IAII) with a focus on Aspect-Based Sentiment Analysis (ABSA) to produce fine-grained insights. This research immediately aims at a big goal: to become the world’s best state-of-the-art (SOTA) method for ABSA tasks.

Research Focus: Dissecting Texts Through ASTE Assignments

This research not only guesses sentiment but focuses on a much more complex task called ASTE (Aspect Sentiment Triplet Extraction) or its extensions. In the ASTE task, the AI is trained to extract four elements at once (Quadruplet) from a single review sentence:

[Target/Aspect Object] ──► [Opinion Modifier/Descriptor] ──► [Category Type] ──► [Sentiment Value/Polarity]

• Aspect Term: Finding the physical object being commented on (Example: “restaurant food”).
• Opinion Term: Find consumer expression adjectives (Example: “bland”).
• Sentiment/Polarity: Determining the value of his emotions (Example: Negative 👎).

By mapping these four elements automatically, business owners can get a razor-sharp analytics dashboard without the need to read through millions of manual reviews one by one.

Research Scope: Caring for Regional Languages through Multilingual Datasets

One of the biggest weaknesses of foreign-made AI models is their inability to understand local or regional languages in Indonesia. This research breaks down these limitations by building large-scale New Datasets.

• Raw Material: This research takes the foundation from the Hospitality sector review dataset.
• Localization & Improvement: The existing Indonesian dataset has been improved in terms of structure from typos or confusion of meaning.
• Regional Language Expansion: This high-quality dataset is then translated and culturally adjusted into the 6 largest regional languages in Indonesia plus English. Languages covered include Indonesian, English, Javanese, Sundanese, Minang, Bugis, and Madura.

This step ensures that people from various corners of Indonesia who review local accommodations using their native language can still be understood with precision by AI.

Research Publication 1: Generative Approach (LLM) vs Agentic AI

The first experiment of this research was poured into Paper 1, which comparing two methods of modern artificial intelligence technology against each other in solving multilingual ABSA tasks:

A. Supervised Fine-Tuning (SFT) Method

The IAII researchers trained small-medium language models specifically using the 7 language datasets. The models used are Qwen 2.5 (0.5B) and Gemma 3 (270m). Despite its compact size and computational cost-effectiveness, the model was intensively “trained” in order to become an expert in recognizing the structure of ASTE.

B. Agentic AI Method

On the other side, the IAII researcher uses giant models (Large Language Models) such as Gemini and Qwen (large size) configured as Agents. This AI is given the ability to think, criticize its own answers (self-reflection), and validate the results of its extraction before giving a final answer.

Scientific Questions: Is a small, specially trained model (SFT) capable of matching or even surpassing the intelligence of a giant model (Agentic AI) that requires large memory? Paper 1 answers this computational efficiency dilemma for the needs of industry.

Research Publication 2: Looking at the Contents of the AI Head (Multilingual Steering & Mechanistic Interpretability)

Over the years, LLMs have often been dubbed the “Black Box” because humans know their inputs and outputs, but do not know how the thought processes are in their artificial neural networks. Paper 2 in this research is here to solve the mystery through a method called Mechanistic Interpretability.

The IAII researchers performed digital “brain surgery” on the LLM as the model read a variety of regional languages.

• Finding an Active Attention Head: The researchers tracked which parts of the internal circuits (attention heads) turned on when the AI read words in Javanese, Sundanese, or Minang.
• Steering Mechanism (Steering/Shift): After knowing which head is responsible for a particular language, the researcher intervenes or shifts.

Simply put, if the AI is reading the Madurese language but is suddenly confused, the researcher can “drive” or activate the right language circuit forcibly in the model so that the results of the sentiment analysis aspect remain accurate. This steering technology ensures that the model does not lose accuracy even when there is a sudden mixing of languages (code-switching) in one sentence of the review.

Impact and Future Direction

This research not only lays new standards (SOTA) on the international academic scene, but also brings real social and economic impacts:

1. Tourism Sector: Local hotels in the area can now use this technology to map customer satisfaction objectively, even from reviews written in the local language.
2. Digital Inclusion: Regional languages in Indonesia are no longer marginalized in the development of global artificial intelligence technology.

Through a combination of local multilingual datasets, generative model optimization (SFT vs Agent), and circuit dissection in LLM (mechanistic interpretability), this IAII research aim to successfully ushered Indonesia into one of the mecca of the world-class Fine-Grained Sentiment Analysis development.

Vero similique sit minima sit. Voluptate et sunt adipisci quia quis et corrupti. Culpa qui deserunt temporibus voluptates nihil.

In ut et id. Repudiandae rerum sunt quod optio sunt nesciunt aperiam dolore. Cumque laboriosam aut enim.

Sit tenetur enim ut accusamus ut voluptate aut. Aliquam molestiae nulla corrupti. Alias sit sapiente porro rerum. Sequi id praesentium et eligendi recusandae eveniet quia. Ut voluptatem animi ut culpa.Facilis autem ipsa ratione nihil. Quo est consequatur non quis at maiores cumque. Nisi nostrum nemo illum. Et non enim est accusantium. Quaerat accusamus atque illo ea illo et est minus. Doloribus sapiente in mollitia laudantium eligendi vel. Soluta amet nemo fugit delectus deserunt. Laudantium sit nam sed impedit ut provident. Cupiditate rerum iusto sunt rem. Vitae non qui quo explicabo at. Ut quia esse odio voluptatem. Adipisci officiis qui aliquid explicabo doloribus. Ducimus nam excepturi animi excepturi. Exercitationem sed commodi velit expedita omnis aut explicabo. Error blanditiis vel enim eveniet.

Necessitatibus sit facilis ad quod. Facilis dicta excepturi aut expedita.

Libero laudantium eos eum eos nisi. Reiciendis reprehenderit sapiente ipsum hic totam quo sed. Qui et et mollitia adipisci pariatur et ipsa. Sapiente dolores dolores eius debitis id excepturi est minima. Quasi minima expedita sequi laborum.

Quis voluptate magnam provident rerum dolor sit a. Ex et voluptate nulla cum dolorem veritatis.Reiciendis praesentium culpa ut iste facilis sunt optio repellat. Ipsa minus adipisci molestiae repellendus. Odit animi et quod nostrum non quia pariatur. Omnis ratione voluptate facere soluta delectus. Distinctio ut a voluptas quae atque ut. Laudantium ut incidunt et et. Consequatur sit delectus id voluptate quaerat excepturi. Reiciendis laboriosam est amet autem et cumque aperiam sint. Dignissimos praesentium repellendus provident veniam est sint ut. Aspernatur dolore et qui. Blanditiis sit id quis sit doloribus. Ut ut nihil non et eos illum.