Ketika AI Menyimpulkan Informasi yang Tidak Pernah Kita Berikan Secara Langsung

Lo nggak pernah bilang ke aplikasi: “gue mungkin mau resign.”

Tidak pernah.

Tapi selama tiga minggu lo:

baca lowongan, upload CV, minta AI memperbaiki cover letter, cek kalender buat interview, cari tahu notice period.

Kalau semua signal itu masuk satu sistem, kesimpulannya tidak susah ditebak.

Ini salah satu hal yang membuat AI terasa cerdas.

Ia tidak cuma menerima fakta.

Ia menghubungkan.

Infer.

Membuat pola.

Masalahnya, inferensi bisa berubah dari helpful menjadi creepy dalam satu langkah.

Assistant menyimpulkan: “kayaknya lo sedang job hunting.”

Mungkin membantu.

HR system menyimpulkan hal yang sama lalu menurunkan peluang promosi?

Beda cerita.

Informasi yang tidak pernah kita berikan secara langsung bisa tetap punya consequence nyata.

Itulah kenapa privacy di era AI makin sulit dijelaskan dengan pertanyaan: “data apa yang kamu kumpulkan?”

Pertanyaan baru: “apa yang kamu simpulkan dari data itu?”

Satu data biasa bisa menjadi sensitif setelah digabung

User beli vitamin.

Tidak terlalu sensitif.

Search dokter.

Maybe.

Lokasi ke rumah sakit.

Maybe.

Chat tentang hasil tes.

Sekarang pattern.

AI bisa menyimpulkan kondisi kesehatan.

Belum tentu benar.

Tapi jika profile dipakai untuk advertising atau pricing, effect real.

UU Pelindungan Data Pribadi mendefinisikan data pribadi sebagai data tentang orang yang teridentifikasi atau dapat diidentifikasi, termasuk jika dikombinasikan dengan informasi lain.

Definisi ini penting.

Identifiability and combination already recognized.

AI makes combination cheap.

Dulu membutuhkan analyst.

Sekarang model bisa menghubungkan dalam detik.

Inference at scale.

Yang bikin rumit: inferensi bukan fakta

AI bisa melihat pola.

Tetap probabilistic.

User beli susu bayi?

Maybe punya bayi.

Maybe buat hadiah.

Sering ke klinik?

Maybe pasien.

Maybe antar keluarga.

Baca artikel utang?

Maybe punya masalah finansial.

Maybe sedang menulis artikel.

Context matters.

Kalau perusahaan menyimpan inference seperti: “financially vulnerable = true”

seolah fakta, problem.

Derived data needs confidence and expiry.

Not permanent label.

Especially high-impact.

Human beings change.

Model guess can be wrong.

Dan label yang salah bisa memengaruhi recommendation berikutnya hingga self-reinforcing.

Sistem melihat lo sebagai X, lalu terus menunjukkan X, lalu makin yakin lo X.

Feedback loop.

Ini bukan sekadar privacy.

Ini identity shaping.

Personalization punya sisi nyaman

Jangan bikin inferensi seolah semuanya buruk.

Netflix menebak genre.

Spotify menebak musik.

E-commerce menebak ukuran.

AI assistant menebak gaya tulisan.

Helpful.

Kita menikmati sistem karena tidak perlu mengatur semuanya manual.

Problem starts when inference becomes sensitive, consequential, hidden, or hard to correct.

A good rule: the more an inference can materially change how a person is treated, the stronger governance should be.

“Likes action movies.”

Low.

“Likely to default on payment.”

High.

“Prefers concise answer.”

Low.

“Likely pregnant.”

High sensitivity.

Context.

Not all personalization equal.

Profiling sudah masuk bahasa UU PDP

UU PDP memberi hak kepada subjek data untuk mengajukan keberatan atas keputusan yang hanya didasarkan pada pemrosesan otomatis, termasuk pemrofilan, jika menimbulkan akibat hukum atau dampak signifikan.

Ini bukan aturan yang melarang semua profiling.

It establishes a right in high-impact automated decision context.

Selain itu, UU PDP mewajibkan penilaian dampak pelindungan data untuk pemrosesan berpotensi risiko tinggi, termasuk automated decision berdampak signifikan, evaluasi atau penskoran sistematis, pencocokan atau penggabungan kelompok data, serta penggunaan teknologi baru.

AI inference jelas bisa masuk wilayah yang membutuhkan perhatian lebih depending use.

Jangan tunggu “AI Law” untuk mulai govern.

PDP already relevant.

Inference dari percakapan lebih intimate

Search history fragment.

Conversation narrative.

User bilang: “gue nggak bisa tidur, kerjaan lagi kacau, pengeluaran bulan ini gede.”

AI can infer: stress.

Financial pressure.

Maybe health issue.

User mungkin cuma meminta tips productivity.

Apakah assistant perlu menyimpan inferensi: “financially distressed”?

Probably not.

This is where memory design matters.

Just because model can infer doesn’t mean product should persist.

Ephemeral inference can answer current question.

No need profile.

Data minimization should apply to derived data too.

Not only collected fields.

Inference dari wajah dan suara lebih kontroversial

Model bisa mencoba menebak: emotion, age, attention, identity.

Some useful.

Age assurance.

Liveness.

Accessibility.

But emotion inference is tricky.

Expression isn’t universal truth.

Nervous face in interview may just mean hot room.

Voice tired may be traffic.

Jakarta jam enam sore, siapa yang nggak kedengaran capek?

Turning ambiguous human signals into categorical profile can be unfair.

If stakes high, evidence standard must rise.

Hiring.

Education.

Insurance.

Public service.

Don’t let model theater become science.

“AI detected low confidence” sounds objective.

Maybe it’s just a guess with decimals.

Decimal points do not create truth.

Location reveals more than address

A single GPS point says: here.

Sequence says: home, office, religious place, clinic, political event, relationship pattern.

AI makes trajectory inference easy.

User may have consented location “for nearby service.”

Would they expect lifestyle profile?

Maybe not.

Purpose limitation matters.

Location data value comes from pattern.

Privacy notices should reflect likely secondary use.

If data is used only for service, don’t keep raw history unnecessarily.

Derived routine can be more revealing than each location.

This is a perfect example: inference sensitivity can exceed source sensitivity.

Company data has inference problem too

Employee AI sees: email tone, calendar, task, document activity.

Could infer: performance, burnout, intent to resign, team conflict.

Dangerous.

Company might think: “kami hanya analyse operational data.”

But derived profile affects people.

Workplace power imbalance.

Need strict boundary.

If productivity system starts making HR inference, governance changes.

Purpose shift.

Employee should know.

Human review.

No secret scoring.

Especially if promotion or termination involved.

Consumers need a “why does the system think this?” button

Recommendation app: “because you watched X.”

Good start.

Imagine broader: “We infer you’re interested in parenting because of these interactions.”

User: “No, gue beliin hadiah ke ponakan.”

Reset.

Correction.

This is not just nice UX.

It improves model.

And gives autonomy.

Inferred profile should be challengeable.

Maybe user doesn’t need see every 10,000 embedding dimension.

That’s useless.

Need human-readable categories.

Interest.

Age estimate.

Risk score where appropriate.

Source.

Confidence maybe.

Action.

Explainable profile.

AI could infer protected or sensitive traits accidentally

Even if company never collects certain attribute directly, proxy data may reveal.

Postal code.

Language.

Shopping.

Names.

Network.

This matters for fairness.

“we don’t collect sensitive attribute” isn’t always enough if model reconstructs it.

Testing needed.

Can model infer?

Does decision correlate?

Bias audit.

Privacy and discrimination overlap.

Remove direct field may not remove proxy.

This is classic ML problem, now more scalable with generative AI.

Regulators need technical capacity.

User also needs awareness: redaction isn’t always anonymity.

Remove name from document.

Content can still identify: job title, company, city, event.

AI combines.

Pseudonymization reduces risk, not guarantees.

Company should test re-identification where high sensitivity.

Not everything needs advanced privacy math.

Sometimes simply remove more context.

If analysis doesn’t need exact company, use category.

If doesn’t need age exact, use band.

Minimize.

Inference should expire

This is under-discussed.

User “likely moving house” based on behavior this month.

Six months later, stale.

“Likely job seeker.”

Stale.

“Interested in wedding venue.”

Very stale after wedding.

Ad-tech is notorious for following old intent.

AI memory could do same.

Inference needs TTL.

Time-to-live.

Some preference persistent.

Some seasonal.

Some momentary.

Good personalization knows decay.

Bad personalization turns every phase of life into permanent identity.

Humans deserve to move on.

High-risk inference should have owner

If company uses score: fraud risk, credit, employee performance, health triage,

someone must own.

What data?

Model?

Validation?

Appeal?

Bias?

Update?

Not anonymous “algorithm.”

Inference becomes business decision input.

Owner required.

Logs.

Review.

If model changes, revalidate.

This is model risk management.

AI doesn’t make responsibility disappear.

It makes it easier to hide.

Don’t.

There is also positive use: inference can protect

Fraud detection infers unusual transaction.

Cybersecurity infers account takeover.

Child safety infers user may be underage.

Medical system infers deterioration.

Useful.

Sometimes critical.

So the answer isn’t ban inference.

It’s govern based on consequence.

A fraud score may need action quickly.

But user needs path if wrong.

A child-age signal can trigger verification, not permanent label.

A health alert should assist clinician, not silently diagnose.

Use inference to focus attention.

Not replace reality.

Ketika AI Menyimpulkan Informasi yang Tidak Pernah Kita Berikan Secara Langsung, privacy becomes less about secrecy and more about interpretive power.

Who gets to interpret our traces?

For what purpose?

How confident?

How long?

Can we correct?

Does the inference change our treatment?

Those questions are going to matter more than another 40-page privacy policy.

Because the strange thing about AI is this:

we can keep a secret perfectly and still reveal enough pieces around it for a machine to guess.

Sometimes the guess will help.

Ada satu jebakan lain: model bisa membuat inference terasa lebih pasti daripada evidence-nya.

Dashboard menampilkan: “attrition risk 82%.”

Angka seperti itu bikin orang percaya.

Padahal 82 persen bukan fakta bahwa employee akan resign. Itu output model berdasarkan dataset, feature, dan definisi tertentu.

Kalau manager melihat angka tanpa konteks, keputusan manusia bisa menjadi otomatis secara psikologis.

“AI bilang 82.”

Selesai.

Ini automation bias.

Maka high-impact inference sebaiknya selalu datang dengan context yang cukup: apa yang diukur, untuk tujuan apa, berapa limitation, dan tindakan apa yang boleh dilakukan.

Jangan memberi decimal kalau tidak membantu.

Sometimes category lebih honest: “perlu review.”

Lalu manusia melihat evidence.

AI juga bisa menghasilkan inference dari data yang sebenarnya tidak intended untuk purpose tersebut.

Customer support chat dikumpulkan untuk menyelesaikan complaint.

Beberapa tahun kemudian data dipakai memprediksi “customer profitability.”

Apakah purpose ini compatible?

Need legal review.

AI membuat secondary use tempting karena dataset sudah ada.

“Sayang nggak dipakai.”

Itu bukan justification.

Justru inference governance perlu mulai dari purpose.

Apa yang ingin kita ketahui dan kenapa?

Bukan: “data kita bisa menebak apa aja?”

Pertanyaan kedua terdengar inovatif, tetapi bisa mengubah perusahaan menjadi mesin profiling tanpa batas.

Ada juga practical habit untuk user: ketika AI memberi kesimpulan tentang diri lo, treat it as hypothesis.

“Dari percakapan kita, kamu sepertinya…”

Maybe useful reflection.

Not identity certificate.

AI melihat potongan.

Bukan seluruh hidup.

Sistem tidak tahu percakapan yang tidak pernah terjadi, keputusan yang tidak direkam, atau konteks yang kita sengaja tidak berikan.

Inferensi adalah compressed interpretation.

Dan semua compression membuang sesuatu.

Itu reason bagus untuk menjaga humility, baik dari sisi user maupun perusahaan yang memakai model untuk menilai manusia.

Sometimes it will be wrong.

The dangerous moment is when nobody remembers it was ever a guess.