Data Anak untuk AI Harus Dipandang Berbeda dari Data Pengguna Dewasa

Anak umur 11 tahun bikin akun.

Klik “setuju.”

Upload foto.

Ngobrol dengan chatbot.

Main game.

Voice chat.

Mungkin dia ngerti aplikasi itu seru.

Belum tentu dia ngerti satu keputusan kecil hari ini bisa membuat jejak data yang diproses bertahun-tahun.

Di sinilah kita perlu berhenti memakai logika: “semua user kan sudah consent.”

Anak bukan adult versi mini.

Indonesia secara hukum sudah mengakui perbedaan itu.

UU Pelindungan Data Pribadi menempatkan data anak sebagai data pribadi spesifik. Pasal 25 menyatakan pemrosesan data pribadi anak diselenggarakan secara khusus dan wajib mendapat persetujuan orang tua atau wali sesuai ketentuan peraturan perundang-undangan.

PP TUNAS dan Permen Komdigi Nomor 9 Tahun 2026 kemudian memperkuat pendekatan child-specific untuk produk digital, termasuk rentang usia, verifikasi pengguna anak, dan desain layanan sesuai tahap tumbuh kembang.

Jadi ketika AI mulai masuk ke game, sekolah, media sosial, dan companion app, baseline-nya harus jelas:

data anak perlu standar lebih tinggi.

Bukan karena setiap data anak otomatis berbahaya.

Karena anak punya waktu hidup yang lebih panjang di depan dataset itu, kapasitas memahami konsekuensi yang berbeda, dan lebih sedikit bargaining power terhadap platform.

Satu foto anak punya umur data yang panjang

Foto umur delapan tahun mungkin masih ada ketika orang itu umur 28.

Voice sample.

Game behavior.

Learning history.

Location.

Friend graph.

AI membuat dataset lama semakin “berguna” karena model masa depan bisa mengekstrak signal yang hari ini belum terpikir.

Itu alasan retention anak harus konservatif.

Kalau data dibutuhkan untuk satu semester, kenapa simpan sepuluh tahun?

Kalau image hanya untuk age verification, kenapa raw photo disimpan?

Child data minimization bukan teori.

Itu cara mengurangi future unknown.

Anak juga lebih gampang memberi data karena desain

Game bilang: “pilih avatar.”

Camera opens.

AI app bilang: “ceritain tentang diri kamu biar aku lebih kenal.”

Anak cerita.

Sekolah.

Teman.

Nama guru.

Alamat.

Hobi.

Masalah di rumah.

AI conversation terasa seperti teman.

Bukan form legal.

Justru itu yang bikin sensitivity naik.

Product untuk anak perlu membatasi pertanyaan.

Jangan meminta detail yang tidak perlu.

Kalau AI companion bisa bekerja tanpa nama sekolah, jangan tanya.

Kalau personalization bisa cukup dengan topik favorit, jangan bangun full profile.

Design should protect user from oversharing.

Parental consent tidak menyelesaikan semuanya

Orang tua klik setuju.

Done?

Belum.

Orang tua juga sering tidak membaca terms.

Consent bisa formal tapi tidak meaningful.

Selain persetujuan, perusahaan tetap perlu purpose limitation, security, minimization, transparency, dan age-appropriate design.

UU PDP memang tidak membangun seluruh rezim data hanya dari consent. Ada berbagai dasar pemrosesan lain sesuai konteks, dan prinsip pengendalian tetap berlaku.

Untuk data anak, pendekatannya harus lebih protektif.

Consent bukan tiket unlimited.

“Orang tua sudah klik” tidak boleh menjadi alasan untuk menyimpan semua.

Anak juga perlu penjelasan yang bisa dia pahami

Privacy policy: “pemrosesan untuk peningkatan layanan berdasarkan kepentingan sah…”

Anak 10 tahun ngerti?

Probably not.

Age-appropriate notice perlu.

“Kalau kamu kirim foto, foto dipakai untuk membuat avatar. Kami tidak akan memakainya untuk iklan.”

Simple.

Untuk parent, detail legal lebih lengkap.

Dua layer.

Anak mendapatkan explanation.

Orang tua mendapatkan control.

Transparency harus bertemu kemampuan pembaca.

Tidak cukup satu document untuk semua usia.

AI di sekolah punya context berbeda

Sekolah mulai menggunakan AI.

Tutor.

Grading assistant.

Attendance.

Learning analytics.

Potential huge.

Personalized learning can help.

Tapi data pendidikan anak sensitif secara practical.

Nilai.

Kesulitan belajar.

Behavior.

Mistake.

Teacher note.

Jika AI membuat profile: “anak lemah matematika.”

Apakah label itu ikut sampai SMA?

Mungkin model salah.

Anak berkembang.

Data perlu expiry.

Educational inference shouldn’t become destiny.

AI should help teacher, not freeze child identity.

Correction and human review vital.

No permanent labels from temporary performance.

AI companion punya intimacy risk

Anak bisa cerita ke chatbot: “gue berantem sama teman.”

“gue takut pulang.”

“orang tua gue marah.”

This can be deeply personal.

Product may detect risk.

Maybe need safety intervention.

But collecting and retaining intimate conversation creates responsibility.

Who can access?

Parent?

Platform moderator?

Law enforcement?

When?

Need clear policy.

Total secrecy impossible if safety concern.

Total parental surveillance also problematic.

Child rights and safety tension.

Design must be thoughtful.

Not solved by one checkbox.

Advertising should be separated hard

If AI learns a child loves football, don’t automatically use that to target purchases.

Personalization for service and personalization for ads are different purposes.

Kids are easier to influence commercially.

PP TUNAS already focuses on child-risk product design.

Data governance should align.

Child safety data shouldn’t become growth data.

Age signal shouldn’t become ad segment.

School interest shouldn’t become shopping profile.

Purpose separation.

Very important.

Model training creates long-term uncertainty

Company collects child chat to improve model.

What happens later?

Model version changes.

Dataset copied.

Third-party training.

Can data be removed?

Hard.

This is why training use needs higher bar.

Maybe de-identification.

Aggregation.

Synthetic data.

Opt-in parent consent where appropriate.

Avoid raw child conversation by default.

Especially intimate or biometric.

“Improve AI” is broad.

Parents need concrete explanation.

What is improved?

What data?

How long?

Can withdraw?

If impossible to explain simply, scope may be too broad.

Biometric child data deserves maximum care

Face age estimation.

Voice authentication.

Gesture.

School attendance.

If biometric used to identify uniquely, UU PDP treats biometric as specific data.

And child data itself is specific.

Combined sensitivity.

Need DPIA when processing has high-risk characteristics under UU PDP.

Security.

Alternative method.

No unnecessary raw retention.

Imagine a database of millions of children’s faces.

High-value target.

Do we need it?

Sometimes maybe.

Often no.

Privacy-preserving age token better.

Children grow, models must forget some things

AI loves memory.

Kids change fast.

Favorite thing age 9.

Identity age 13.

Beliefs age 16.

Permanent personalization can become creepy.

Child AI should perhaps have shorter memory by default.

Reset.

Age transition.

Ask again.

Don’t bring embarrassing childhood chat into older teen experience.

Design for growth.

This is rarely discussed.

Data retention is developmental.

A child isn’t static account.

School graduation should trigger cleanup.

Age 18 maybe new terms.

Guardian link changes.

Lifecycle.

Parents also overshare

Sharenting.

Photo sejak bayi.

School uniform.

Location.

Birthday.

AI era adds new consequence.

Public photos can be scraped.

Face can be cloned.

Parents should think: does this need to be public?

No need panic.

But maybe don’t post: full name + school + schedule + live location.

Combination identifies.

UU PDP generally exempts personal or household activity from its scope, so don’t treat ordinary parent posting exactly like corporate data processing.

But safety logic still applies.

Privacy habit starts at home.

Child never chose their digital archive.

Platform should not punish privacy-conscious families

Some apps require face upload for fun feature.

If parent says no, child loses entire service.

Is that necessary?

Maybe offer non-AI avatar.

Alternative.

Data choice shouldn’t become coercive unless function truly requires it.

This is meaningful consent.

“Agree or no service” can be problematic when data is not necessary.

Product design can provide low-data mode.

Especially child product.

AI provider should conduct child-specific risk assessment

Not generic privacy checklist.

Ask:

Could model encourage oversharing?

Can strangers reach child?

Can AI infer sensitive profile?

What happens to memory?

Can parent control?

Can child delete?

Can data train model?

How is age verified?

What if age wrong?

These are child-specific.

Risk assessment should include child development expert.

Not only engineer.

Kids interact unpredictably.

Need testing.

Children should be involved carefully

Ask actual young users whether notice makes sense.

Can they find block?

Do they know bot is AI?

Do they understand public/private?

Ethical research.

Parental consent.

Safe design.

Adults designing for children often guess.

User research can reveal.

PP TUNAS is pushing product toward age-appropriate design.

Testing should follow.

Data Anak untuk AI Harus Dipandang Berbeda dari Data Pengguna Dewasa because time, vulnerability, and understanding are different.

An adult can make a bad privacy decision and deal with consequences.

A child may not even know a decision was made.

And the dataset can follow them longer than the childhood moment that created it.

The right goal isn’t to keep children away from AI.

They will use it.

For learning.

Creativity.

Games.

Communication.

The goal is to make AI adapt to children, instead of expecting children to understand an adult-scale data economy.

Collect less.

Keep shorter.

Explain simpler.

Separate advertising.

Protect biometrics.

Give parents tools without turning them into surveillance operators.

Ada satu hal yang sering dilupakan ketika perusahaan bicara “parental consent”: orang tua juga butuh desain yang masuk akal.

Kalau setiap aplikasi anak mengirim tiga halaman legalese, orang tua akan approve seperti kita menerima cookie banner.

Capek.

Consent fatigue.

Platform perlu merangkum hal yang benar-benar penting:

data apa, untuk fitur apa, apakah disimpan, apakah dipakai melatih AI, siapa yang bisa menghubungi anak, dan bagaimana mematikan fitur.

Lalu detail lengkap tetap tersedia.

Ini bukan menyederhanakan hukum.

Ini membuat informasi benar-benar terbaca.

Konteks Indonesia juga penting karena satu perangkat sering dipakai bersama.

Tablet keluarga.

HP orang tua dipinjam anak.

Laptop sekolah dipakai bergantian.

Sistem yang menganggap “satu device = satu usia” bisa salah.

Recommendation adult bisa muncul ke anak.

Atau child restriction mengganggu orang tua.

Account separation dan profile menjadi bagian child safety.

Bukan hanya verifikasi tanggal lahir.

Sekolah juga perlu hati-hati dengan vendor edtech yang tiba-tiba menambahkan AI.

Kontrak awal mungkin hanya untuk learning management system.

Update berikutnya menambah chatbot yang memproses jawaban murid.

Apakah data use berubah?

Apakah ada model provider baru?

Apakah percakapan disimpan?

Vendor update seharusnya memicu review, bukan dianggap feature biasa.

Karena orang tua menyetujui satu sistem bukan berarti semua AI feature masa depan otomatis masuk paket persetujuan yang sama.

Child data governance harus mengikuti perubahan produk.

Satu prinsip yang cukup sehat adalah “least surprising use.”

Kalau orang tua memberi data nilai anak untuk membuat rapor, mereka reasonably expect data dipakai untuk pendidikan.

Mereka mungkin tidak expect data yang sama dipakai membangun advertising profile atau melatih model komersial umum.

Semakin jauh penggunaan dari expectation awal, semakin besar kebutuhan transparency dan legal review.

AI membuat reuse sangat mudah.

Justru karena mudah, boundary harus lebih jelas.

And remember the most basic thing.

A child should be allowed to grow beyond the data profile a machine built about them yesterday.