Data Minimization Kembali Penting di Era Semua Hal Ingin Dikumpulkan

Ada penyakit klasik di produk digital:

kalau bisa dikumpulkan, simpan dulu.

“Siapa tahu kepakai.”

Email? Ambil.

Tanggal lahir? Ambil.

Lokasi? Ambil.

Kontak? Kalau bisa.

Voice? Menarik.

Riwayat klik? Tentu.

Chat? Lebih bagus lagi.

AI membuat godaan ini naik level karena data memang memberi kemampuan.

Lebih banyak konteks bisa menghasilkan jawaban lebih personal.

Lebih banyak histori bisa meningkatkan prediction.

Lebih banyak dokumen membuat enterprise search terasa pintar.

Jadi argumen “data mungkin berguna nanti” sekarang terdengar lebih meyakinkan daripada sebelumnya.

Justru karena itu data minimization kembali penting.

Bukan sebagai slogan privacy yang membosankan.

Sebagai strategi agar perusahaan tidak tenggelam oleh aset yang sekaligus menjadi liability.

UU Pelindungan Data Pribadi Indonesia punya prinsip yang cukup langsung: pengumpulan data pribadi dilakukan secara terbatas dan spesifik, sah secara hukum, transparan, serta pemrosesan dilakukan sesuai tujuan.

Data minimization bukan berarti tidak boleh punya data.

Artinya jangan mengambil lebih banyak daripada yang memang dibutuhkan.

Bedanya kelihatan kecil.

Efeknya besar.

Form registrasi adalah tempat paling gampang melihat

Aplikasi cuma butuh email untuk login.

Kenapa minta tanggal lahir?

“Buat personalization.”

Apakah personalization benar-benar dipakai?

Nggak.

Kenapa minta gender?

“Buat analytics.”

Analytics apa?

Belum tahu.

Kenapa minta nomor HP?

“Backup.”

Padahal recovery email cukup.

Satu field ekstra terasa murah.

Tapi kalau user 10 juta, satu field menjadi 10 juta record yang harus diamankan.

Data minimization memaksa product manager menjawab: kalau field ini hilang, fungsi inti rusak nggak?

Kalau nggak, pertimbangkan buang.

Buat AI, pertanyaannya makin penting: apakah model butuh identitas atau cuma context?

User ingin AI membantu membuat budget.

Model butuh: income, expense, goal.

Apakah butuh nama lengkap?

NIK?

Alamat rumah?

Tidak.

AI sering bekerja sama bagusnya dengan data pseudonymous.

Gunakan.

More context tidak selalu berarti better context

Ini counterintuitive.

Orang suka menganggap: kasih semuanya biar AI makin pintar.

Padahal terlalu banyak data bisa membuat noise.

RAG mengambil dokumen tidak relevan.

Model bingung policy lama.

Personal assistant membawa preference yang sudah outdated.

Knowledge base penuh duplicate.

Data minimization juga information quality.

Curate.

Bukan hoard.

Untuk perusahaan, less data can mean better answer.

Satu folder policy authoritative sering lebih useful daripada 15 tahun arsip campur.

Konteks yang kecil tetapi bersih mengalahkan tumpukan sampah digital.

Storage murah menipu kita

Dulu orang hapus file karena hard disk penuh.

Sekarang cloud murah relatif.

Tidak ada pressure.

Folder 2016 masih ada.

Chat.

Recording.

Log.

Backup.

“Ngapain hapus?”

Karena setiap byte yang tetap hidup membawa obligation.

Access control.

Security.

Discovery.

Deletion request.

Incident response.

Kalau breach, data lama ikut.

Storage cost mungkin receh.

Risk cost belum tentu.

Retention bukan keputusan infra.

It is risk decision.

AI bahkan meningkatkan value data lama untuk attacker

Dataset tua dulu tidak terlalu menarik.

Sekarang bisa: train, clone, profile, infer.

Old voice recordings.

Face image.

Customer conversation.

Support ticket.

AI can extract structure from messy archive.

Useful to company.

Also useful to attacker.

Data that was harmless-ish when collected may become more revealing as models improve.

Future capability risk.

One more reason not to keep everything forever.

Meeting recording adalah contoh nyata

Perusahaan hanya butuh action items.

Tapi simpan full audio.

Kenapa?

“Storage included.”

Setahun kemudian ada ribuan jam percakapan internal.

HR issue.

Customer detail.

Strategy.

Casual comment.

Searchable by AI.

What did we gain?

Maybe nothing.

Better pipeline: record temporarily, make transcript, approve summary, delete raw audio according to policy unless needed.

Less data.

Same business value.

AI product can be smarter by forgetting

Ini terdengar aneh karena industri berlomba membangun memory.

Tapi good assistant should know what not to keep.

Preference: “gue suka jawaban ringkas.”

Useful.

Temporary detail: “besok meeting jam 2.”

After tomorrow, maybe irrelevant.

Sensitive one-off: “ini hasil medical check.”

Maybe don’t persist unless user wants.

Memory needs decay.

Priority.

User control.

OpenAI’s memory system, for example, now lets some users review and manage stored memory, while Temporary Chat can be used when information shouldn’t become future personalization context.

Again, specific product.

Broader principle: memory should be selective.

Not vacuum cleaner.

Data minimization juga bikin compliance murah

Imagine company receives deletion request.

System A punya data di 4 places.

System B punya data di 27 places.

Mana lebih mudah?

Less collection reduces: legal complexity, security surface, backup, audit, vendor dependency.

Privacy and operational efficiency align.

CISO often wants less.

Marketing wants more.

Product wants more.

AI team wants more.

Someone needs authority to ask: “do we really need this?”

Data governance cannot be advisory only.

There needs to be friction for new collection.

Not huge committee.

Simple justification.

Purpose.

Retention.

Owner.

If no owner, don’t collect.

“Semua data bisa dipakai AI nanti” bukan purpose

Ini kalimat yang harus ditolak.

Terlalu broad.

What AI?

For what?

Which decision?

Who benefits?

How long?

Data lake built on vague future value becomes data swamp.

And when AI connects, swamp becomes searchable.

Then everyone panics.

Better to define use case.

Collect sufficient data.

Evaluate.

If more needed, expand.

Iterative.

Same way we build product.

No need hoard upfront.

Data minimization bukan anti-personalization

Personalization can use less.

Age band instead of birthdate.

City instead of exact GPS.

Preference tag instead of entire chat history.

Aggregated purchase pattern instead of raw item forever.

Device-level processing instead of cloud upload.

Derived feature with expiry.

There are design choices.

Privacy-friendly doesn’t mean generic product.

It means information architecture intentional.

Sometimes coarse data is enough.

This is especially relevant for children

Platform needs know: under 16?

It doesn’t necessarily need full KTP.

Age token.

Guardian status.

Less data.

Child safety and privacy can reinforce.

If every platform asks identity document, protection creates a new risk.

Data minimization keeps assurance scoped.

Same principle national digital identity.

Service may need eligibility, not full identity profile.

Share attribute.

Not database dump.

AI engineers need minimization mindset too

Not just lawyers.

When building dataset: drop unnecessary columns.

Tokenize identifiers.

Separate secrets.

Use synthetic data for testing where appropriate.

Sample.

Mask.

If debugging retrieval, you don’t need real payroll.

Use canary.

Development environments are leak risk.

Real production data often copied because easiest.

Don’t.

Data minimization must be practical in engineering workflow.

Tooling helps.

Automatic PII detection.

Redaction.

Schema contracts.

Access.

Company culture matters.

People collect because no downside visible

Product KPI rewards more data.

“Profile completion.”

“Data enrichment.”

Rarely rewards: “we deleted 30 percent useless data.”

Maybe it should.

Data debt is real.

Could track: unowned datasets, expired data, duplicate PII, retention violations.

Cleanup can be OKR.

Not sexy.

But AI era makes it strategic.

The cleanest AI deployment often starts with less input, not more model.

Data minimization also reduces hallucination blast radius

If model only sees allowed context, it can’t leak context it never received.

If RAG corpus excludes restricted docs, one whole class of incident disappears.

If agent has no access to customer identity, prompt injection can’t exfiltrate it.

Security through least privilege.

Same principle.

Minimize data and minimize permission.

Two sides.

OWASP’s GenAI Data Security initiative emphasizes that generative systems aggregate user inputs, retrieved documents, embeddings, tool outputs, prompts, and agent memory across multiple trust domains.

Every extra stream expands attack surface.

The model may be sophisticated.

The risk equation stays primitive: what isn’t there can’t leak.

Not every deletion is permanent knowledge loss

Organizations fear: “kalau dihapus, nanti butuh.”

Archive can exist under controlled retention if lawful and justified.

Data minimization doesn’t demand corporate amnesia.

It demands tiering.

Active knowledge base: clean.

Archive: restricted.

Legal hold: separate.

Backup: controlled.

Not all data equally accessible.

The real problem is keeping everything in hot storage forever.

Especially searchable by AI.

Data Minimization Kembali Penting di Era Semua Hal Ingin Dikumpulkan because AI changed the value proposition of data.

More data can produce more capability.

Exactly.

But more data also creates more places to leak, more things to secure, more information to delete, more stale context, and more ways for a system to know something it never needed to know.

The mature AI company will not be the one with the biggest pile.

Ada satu cara sederhana membuat minimization tidak terasa abstrak: beri harga internal pada data.

Bukan rupiah literal.

Risk score.

Kalau tim ingin menambah field baru, mereka harus “membayar” dengan alasan, owner, retention, dan control.

Sensitive data lebih mahal.

Biometric paling mahal.

Public preference murah.

Dengan begitu product team berhenti menganggap semua field gratis.

Data punya carrying cost.

Perusahaan juga bisa menjalankan “data diet” tiap kuartal.

Cari: kolom tidak pernah dipakai, log terlalu lama, file orphan, account mati, embedding dari source deleted, data test yang ternyata production copy.

Delete.

Seperti bersihin kulkas sebelum belanja lagi.

Tidak glamorous.

Tapi setelah beberapa bulan, system jauh lebih sehat.

Dan ketika AI baru ingin diluncurkan, tim punya corpus yang lebih bersih.

Banyak organisasi berpikir AI readiness berarti mengumpulkan lebih banyak.

Sering justru sebaliknya.

AI readiness adalah tahu data mana yang memang layak dibawa ke masa depan.

Kadang keputusan paling cerdas untuk AI bukan menambah dataset baru, tetapi berani bilang: data ini ternyata tidak perlu disimpan.

It will be the one that knows why every important dataset exists.

Maybe the most future-proof database field in 2026 is the one the product team decided not to collect.