Foundations — AI governance and regulation (EU AI Act)

6.4.13 · D1 · AI-ML › AI Safety & Alignment › AI governance and regulation (EU AI Act)

Pehle aapko EU AI Act wale parent note ko padhne se pehle, usme chupi hui words aur pictures ko samajhna hoga. Yeh page har ek cheez ko zero se banata hai, us order mein jo agli cheez ko samajhna aasaan bana de.

1. "AI system" hota kya hai?

Ek box ki picture karo. Left side pe, ek arrow andar ja raha hai (input — ek job application, ek X-ray, ek sentence jo aapne type kiya). Right side pe, ek arrow bahar aa raha hai (output — "hire / reject", "cancer / healthy", ek reply). Box ke andar ek rule hai jise kisi ne directly nahi likha; box ne hazaron past examples dekhke khud figure out kiya.

2. "Risk" — dar ko ek number-ish idea mein badalna

Poora Act ek word pe tikaa hai: risk. Istemal hone se pehle isey define karte hain.

Do knobs socho. Ek knob hai severity (kitna damage — ek spam email mild hai, ek wrongful arrest severe hai). Doosra knob hai probability (yeh kitni baar galat jaata hai). Koi bhi knob upar ghuma do aur danger light zyada chamkegi.

3. Char tiers — ek switch nahi, ek staircase

Jab aap risk measure kar sakte ho, tab systems ko levels mein group kar sakte ho. Act char use karta hai.

Side se dekha hua ek staircase socho. Sabse upar ka step (unacceptable) pe ek wall hai — wahan nahi ja sakte. Har step neeche jaane pe kam chains hain. Sabse neeche ka step (minimal) khula maidan hai.

4. "Compliance" — woh checklist jise system ko pass karna hota hai

Woh badi brace parent mein — $\text{Compliance}=\{\text{Risk Management}+\text{Data Governance}+\dots \}$ — real arithmetic nahi hai. $+$ signs ka matlab "AND" hai, aur brace ka matlab "yeh sab milake" hai. Isey ek checklist ki tarah padho jahan har box tick hona zaroori hai, add up karne wala total nahi.

5. Checklist ke andar ki vocabulary

Har requirement ek term hai jo parent freely use karta hai. Yeh rahi zero se.

6. "Bias" aur "fairness metrics" — data governance kyun exist karta hai

Parent do ideas pe lean karta hai bina define kiye: bias aur fairness metrics.

Training examples ki do piles socho: Group A ki ek tall pile, Group B ki ek tiny pile. Box ko A pe bahut practice milti hai aur B pe almost nahi — toh A ke liye accha guess karta hai aur B ke liye bura. Woh lopsidedness hi bias hai.

7. Rights: red lines ke peeche ka "kyun banned"

Parent EU Charter of Fundamental Rights se bans justify karta hai. Tumhe sirf plain idea chahiye:

Unacceptable-risk systems (social scoring, mass live facial recognition, manipulation) banned hain kyunki yeh in rights pe directly attack karte hain, sirf risky hone ki wajah se nahi. Isliye yeh walled-off top step pe hain: koi bhi paperwork ek rights-violation ko acceptable nahi bana sakta. Is concern ka door ka end — jahan systems insaniyat ko hi threaten karte hain — long-term existential risks mein cover hai, aur ise rokne ki research AI safety research mein hai.

Foundations topic ko kaise feed karti hain

Ise top-down padho: yeh jaanna ki AI box kya hai tumhe uske risk ki baat karne deta hai; risk se tiers mein sort ho jaate ho; high tier ko ek compliance checklist chahiye; top tier rights ki wajah se banned hai. Saari raahein Act ki taraf jaati hain.

Equipment checklist

Right side cover karo. Agar har ek ka jawab de sako, tab tum parent note ke liye ready ho.