AI governance and regulation (EU AI Act): Worked examples

6.4.13 · D3 · AI-ML › AI Safety & Alignment › AI governance and regulation (EU AI Act)

Shuru karne se pehle, teen plain-word definitions jo hum poore page pe use karenge. Agar parent note ne ye words casually use kiye, yahan unhe nail down kiya gaya hai. Ek aur shorthand jo legal text mein milegi:

Definition Teen words jisse hum sort karte hain (plus ek legal shorthand)

Risk tier ::: kaun sa ek chaar boxes mein se (Unacceptable / High / Limited / Minimal) system land karta hai. Ye decide hota hai system logon ke saath kya karta hai se, model kitna clever hai usse nahi.
Obligation load ::: paperwork aur safeguards ka wo dhera jo tier tumpe force karta hai. Unacceptable = "tum ise ship hi nahi kar sakte". High = poora saat-part checklist. Limited = "bas users ko batao". Minimal = "jo chaaho karo".
Trigger ::: scenario ka wo single fact jo ise tier upar push karta hai. Trigger dhundhna hi poori skill hai.
Article (short mein "Art.") ::: law ka ek numbered paragraph, jaise ek rulebook mein numbered rule. Jab hum Art. 5 likhte hain matlab "EU AI Act ka rule number 5", jo banned systems ki list hai. Ye sirf ek pointer hai text mein kahan rule hai — kuch mathematical nahi.

Scenario matrix

Is topic ka har scenario in cells mein se kisi ek mein aata hai. Har row ek "case class" hai; neeche ke worked examples cell ke saath tagged hain.

Cell	Case class	Kya alag banata hai	Worked example
C1	Clear Unacceptable	Ek hard red line (social scoring, subliminal manipulation)	Ex. 1
C2	Clear High-risk	Employment / medical / law-enforcement use	Ex. 2
C3	Clear Limited	Chatbot / deepfake — transparency only	Ex. 3
C4	Clear Minimal	Game AI, spam filter — koi obligation nahi	Ex. 4
C5	Boundary / "banned jaisa lagta hai lekin hai nahi"	Same tech, alag context tier flip kar deta hai	Ex. 5
C6	Degenerate input	Ek rule-based system jiska rights pe koi impact nahi	Ex. 6
C7	Multi-trigger (worst-tier-wins)	Ek system jisme kai risky functions hain	Ex. 7
C8	Real-world word problem	Ek messy startup pitch jise classify + cost karna hai	Ex. 8
C9	Exam twist	Jurisdiction / "EU ke bahar develop hua" trap	Ex. 9

Figure — AI governance and regulation (EU AI Act)

Is figure ko dhyan se padho, kyunki ye har example drive karta hai. Ye char stacked boxes ki ek decision ladder hai:

Bottom rung — Minimal (chalk blue): sabse sasta floor, 0 obligations. Games, spam filters yahan rehte hain.
Second rung — Limited (chalk blue): sirf transparency. Chatbots, deepfakes.
Third rung — High (pale yellow): poora 7/7 checklist. Hiring, medical, education.
Top rung — Unacceptable (chalk pink): banned. Social scoring, live mass biometric ID.

Rungs ke beech upar wale white arrows trigger points hain — tum tabhi chadhte ho jab koi specific fact fire karta hai:

Minimal → Limited "confusion trigger" pe fire karta hai (user ko nahi pata ki woh machine se deal kar raha hai).
Limited → High "rights or safety trigger" pe fire karta hai (system kisi fundamental right ya physical safety ko affect karta hai).
High → Unacceptable "red-line trigger" pe fire karta hai (system ek banned pattern se match karta hai).

Top se "fall off the top" label wala pink arrow key mental image hai: ek baar red-line trigger fire ho jaye, tum zyada paperwork wale upar ke rung pe nahi jaate — tum ladder se bilkul bahar ho jaate ho aur system ship nahi ho sakta. Tum hamesha bottom se start karte ho (sabse sasta) aur utna hi chadhte ho jitna strongest trigger force kare. Neeche ka har example is ladder pe ek walk hai — ya usse bahar.

Worked examples

Example 1 — Cell C1: Clear Unacceptable

Forecast: Ruko. Kaun sa box hai, aur kya koi paperwork ise rescue kar sakta hai?

Har function list karo. System logon ki overall trustworthiness ko unrelated behaviour se rate karta hai aur phir ek public service restrict karta hai. Ye step kyun? Trigger tab tak nahi milega jab tak tum pehle name nahi karte ki system logon ke saath kya karta hai.
Red lines se match karo. Ye textbook social scoring by a public authority hai — explicitly Art. 5 se prohibited (yaad karo: "Article 5", Act ka woh rule jo banned systems list karta hai). Ye step kyun? Unacceptable-risk ek closed list hai; list se match karna hi poori test hai — koi risk-weighing nahi.
Rescue dhundho. Kya koi transparency fix ya human-oversight fix hai? Nahi. Prohibited matlab prohibited; obligations ise scale down nahi karte. Ye step kyun? Beginners assume karte hain "oversight add karo = allowed". C1 ke liye ladder rung neeche nahi hai — tum top se gir jaate ho.

Answer: Tier = Unacceptable → bilkul banned. Obligation load = ∞ (yaad karo: koi finite pile of paperwork ise legal nahi banati — tum ship nahi kar sakte).

Verify: Parent note mein Charter derivation ke against tier cross-check karo: social scoring dignity (Art. 1) aur freedom of thought (Art. 10) violate karta hai. Ye check kyun matters karta hai: ye "banned" verdict ko alag source se independently confirm karta hai (human-rights law, risk-tier list se nahi) — agar do independent routes agree karein, toh humne list galat nahi padhi. Rights-violating system jisme koi proportionate law-enforcement carve-out nahi ⇒ prohibited. Consistent. ✓

Example 2 — Cell C2: Clear High-risk

Forecast: Tier AND obligation count guess karo aage padhne se pehle.

Function + kaun impact hota hai, name karo. Ye ek medical decision influence karta hai jo life-threatening ho sakta hai. Kyun? Medical devices ek named high-risk category hai — domain khud trigger hai.
Confirm karo ki ye banned nahi hai. Disease diagnose karna red-line list mein nahi hai. Kyun? C1 ko hamesha rule out karo C2 settle karne se pehle — ladder danger ke liye upar se climb ki jaati hai, lekin tum phir bhi confirm karte ho ki tum edge se bahar nahi ho.
Poora checklist load karo. Parent ke compliance formula ke saare 7 obligations apply hote hain: risk management, data governance, technical documentation, record-keeping, transparency, human oversight, accuracy/robustness/cybersecurity. Kyun? High-risk "all or nothing" hai — tum subset nahi choose kar sakte.
Numeric bars pin karo. Parent note example clinical thresholds deta hai: sensitivity $\geq 95%$ , specificity $\geq 90%$ . Kyun? Accuracy obligation sirf concrete pass/fail numbers se meaningful hai.

Answer: Tier = High-risk. Obligation load = 7 / 7, sensitivity $\geq 0.95$ aur specificity $\geq 0.90$ as hard bars ke saath.

Verify: Maano ek validation run mein 100 sick patients mein se 96 true positives aur 100 healthy patients mein se 92 true negatives aate hain. $sensitivity = \frac{96}{100} = 0.96 \geq 0.95 ✓ specificity = \frac{92}{100} = 0.92 \geq 0.90 ✓$ Ye check kyun matters karta hai: accuracy obligation ye kehne se satisfy nahi hoti ki "hum high-risk hain, humne try kiya" — ye tabhi satisfy hoti hai jab measured performance legal numbers clear kare. Ye plug-in ek concrete deployment dikhata hai jo actually audit pass karega, prove karta hai ki thresholds reachable hain aur ki humne jo bars bataye (0.95, 0.90) wahi test ho rahe hain. Donon bars cleared. Scanner types ke across robustness clause testing force kyon karta hai dekhne ke liye 6.4.4-Robustness-and-adversarial-examples dekho.

Example 3 — Cell C3: Clear Limited

Forecast: Definitions mein se kaun sa ek word poori obligation hai yahan?

Steps se pehle, transparency rule ko plain words mein unpack karo taaki neeche ki equation earned lage, sirf drop na ho.

Definition Transparency obligation, term by term

Limited-risk duty ek do alag cheezein ka sum hai jo user ko di jaani chahiye. Ise likhte hain: $User Awareness = Clear Disclosure + Right to Know$ " $=$ " ko "se achieve hota hai" padho aur " $+$ " ko "AND — donon parts zaroori hain, either/or nahi":

Clear Disclosure ::: ek active statement jo system tumpe push karta hai, jaise "You're chatting with an AI" line type karne se pehle dikhti hai. Provider ko khud volunteer karna padta hai.
Right to Know ::: tumhara standing entitlement puchne aur bataye jaane ka ki machine hai, koi insaan nahi, interaction ke peeche — chahe tumhe banner dikhaya na gaya ho.
" $+$ " ka matlab hai donon hold karne chahiye. Banner bina confirm karne ke, ya confirm-on-request system bina banner ke — har ek adha duty fail karta hai. Koi partial credit nahi — ye logical AND hai, arithmetic sum nahi.

Function 1: chatbot. Users ek machine se interact karte hain jo insaan samjha ja sakta hai. Ye step kyun? Limited-risk ka trigger confusion hai ki tum kisse/kya baat kar rahe ho.
Transparency formula apply karo. Upar ki definition use karte hue, donon terms satisfy karo: Clear Disclosure push karo ("You're chatting with an AI") aur Right to Know honour karo (user request pe confirm kar sakta hai). Ye step kyun? Log insanon aur machines ke saath alag behave karte hain — disclosure informed choice restore karti hai, aur sirf donon parts ka AND ise puri tarah restore karta hai.
Function 2: synthetic spokesperson = deepfake. AI-generated/synthetic label lagna chahiye. Ye step kyun? Video ko real endorsement samjhe jaane se rokta hai (misinformation control).

Answer: Tier = Limited-risk. Obligation load = sirf transparency — chatbot ke liye ek disclosure line, video pe ek "synthetic" label. Koi documentation, oversight, ya audits nahi.

Verify: Contrast test — human-confusion element hatao (jaise bot clearly "AI Assistant" kehta hai apne naam mein to koi reasonable user fooled nahi hota). Ye check kyun matters karta hai: test karta hai ki humne sahi trigger identify kiya — agar classification disclosure banner hatane ke baad bhi survive kare lekin kuch higher nahi aata, toh koi rights/safety trigger kabhi present nahi tha aur hum tier under-shoot nahi kar sakte. Disclosure trigger kamzor hota hai, lekin label sasta hai, isliye hum comply karte hain waise bhi. Koi higher trigger present nahi ⇒ High nahi ho sakta. Ladder se consistent (Minimal se ek rung upar). ✓

Example 4 — Cell C4: Clear Minimal

Forecast: _Kitni paperwork? (Ye ek trick hai — number guess karo.)**

Pucho: kya ye kisi fundamental right ya safety ko touch karta hai? Screen pe monsters attack karna: nahi. Spam filter karna: koi life-altering decision nahi kisi insaan ke baare mein. Ye step kyun? Agar koi bhi trigger fire nahi karta, tum by default ladder ke bottom pe ho.
Confirm karo koi confusion trigger nahi hai. Game monster ko koi real insaan nahi samjhta, aur spam filtering koi conversation nahi hai. Ye step kyun? Limited tier bhi rule out kar deta hai.

Answer: Tier = Minimal-risk. Obligation load = 0 (sirf voluntary codes of conduct).

Verify: Obligation count 0 floor hai; koi lower tier nahi hai girne ke liye, aur chadhne ke liye koi trigger nahi. Ye check kyun matters karta hai: confirm karta hai ki hum ladder figure ke bilkul bottom rung pe landed — agar kisi check ne negative obligation count produce kiya hota matlab scheme mein Minimal ke neeche hole hai, jo hona nahi chahiye. Count = 0 legal aur stable hai — parent se "most AI applications ko koi burden nahi" se match karta hai. ✓

Example 5 — Cell C5: Boundary — "looks banned but isn't"

Ye wahi trap hai jo parent ke [!mistake] callout ne warn kiya tha. Iske liye figure chahiye kyunki context hi poora answer hai.

Figure ek algorithm (facial recognition) ko teen columns mein fan out karte dikhata hai. Columns padhte hue:

Left column (chalk pink) — RED / banned column. Iske teen stacked conditions exactly live (real-time), public space, aur mass identification hain. Ye exactly isliye banned hai kyunki teeno conditions ek saath hold karti hain.
Middle column (pale yellow) — High-risk. Conditions: post-hoc (recorded footage) + targeted (ek suspect) + judge approval.
Right column (chalk blue) — Minimal. Conditions: private device + opted in + one user.

Phrase "live + public + mass" yaad rakho — woh trio hi poori red line hai.

Forecast: Same tech, teen answers. Kaun sa banned hai, kaun sa High, kaun sa Minimal/Limited?

(a) Live + public + mass. Teeno red-line conditions ek saath present hain (ye exactly wahi teen hain jo figure ke red / pink column mein stacked hain: live, public, mass). Ye step kyun? Public mein real-time mass biometric ID permanent chilling effect create karta hai — exactly banned pattern. ⇒ Unacceptable (banned).
(b) Post-hoc + targeted + judicial oversight. "Live" aur "mass" conditions gone hain; ek narrow law-enforcement exception warrant ke saath apply hoti hai. Ye step kyun? Real-time/mass elements hatane se prohibition ka trigger hata deta hai; lekin law-enforcement context mein logon ko identify karna phir bhi High-risk hai (oversight, logging chahiye). ⇒ High-risk (allowed with full checklist).
(c) Private device, ek consenting user. Koi public space nahi, koi mass surveillance nahi, tumne opt in kiya. Ye step kyun? Koi bhi biometric-surveillance trigger fire nahi karta; aur koi confusion trigger bhi nahi hai — apna phone unlock karta insaan jaanta hai ki machine usse authenticate kar rahi hai, isliye Limited-tier "kya tumhe pata tha ye AI hai?" duty ke paas bite karne ke liye kuch nahi. ⇒ Minimal — nahi Limited, kyunki jo cheez ise Limited tak lift kar sakti thi wo confusion trigger tha, aur woh koi nahi hai.

Answer: (a) Banned, (b) High-risk, (c) Minimal (Limited step 3 ke reason se ruled out hai: koi confusion trigger nahi). Same algorithm; context tier flip kar deta hai.

Verify: Parent ke [!mistake] fix ke saath cross-check karo: banned = real-time + public + mass; still-allowed = post-event analysis aur non-public device unlock. Ye check kyun matters karta hai: parent note is exact boundary ke liye hamaari ground truth hai — teen answers ko iske teen named rows pe map karna prove karta hai ki humne koi aisa boundary nahi ijaad kiya jo law nahi kheenchti. Hamare teen answers exactly un teen rows pe map karte hain. ✓

Example 6 — Cell C6: Degenerate input

Forecast: Kya sirf rules hone ka, learning nahi, ise law se bilkul bahar rakhta hai?

Test karo ki ye in scope hai — carefully. Ek common trap: "koi learning nahi ⇒ AI system nahi." Galat. EU AI Act ki AI system ki definition explicitly logic- aur knowledge-based (rule-based) approaches include karti hai, sirf machine learning nahi. Isliye ek deterministic rule engine phir bhi AI system count ho sakta hai. Ye step kyun? Ye hi degenerate cell ka poora point hai — naive escape hatch ("sirf rules hain") kaam nahi karta.
Toh function ke basis pe classify karo, baaki sab ki tarah. Ye logon ke saath kya karta hai? Tax deterministically compute karta hai — koi discretionary, rights-affecting decision nahi. Tax computation ek named high-risk domain nahi aur koi red line nahi. Ye step kyun? In scope hona sirf ye batata hai ki sorting machine apply hoti hai; tier phir bhi logon pe impact se decide hota hai.
Tier land karo. Koi rights/safety trigger aur koi confusion trigger fire nahi karta ⇒ ladder ka bottom. Ye step kyun? Same reasoning as Example 4 — har trigger ki absence matlab Minimal.

Answer: Tier = Minimal-risk (ye Act ki broad definition ke under likely ek AI system hai, lekin harmless). Obligation load = 0 binding obligations, sirf voluntary codes. "AI-powered" marketing na duties add karta hai, na remove — function decide karta hai.

Verify: Degenerate check — classification learning vs. rules distinction pe depend nahi karni chahiye (Act woh distinction refuse karta hai) aur na hi marketing label pe. Ye check kyun matters karta hai: ek robust classifier ko same answer dena chahiye chahe system kaise implement ya advertise kiya gaya ho; agar "ML" ko "rules" mein flip karne se tier change hota, hamaara rule galat feature pe react kar raha hota. Donon invariances hold karte hain: rule-based phir bhi in scope hai, aur tier function ke basis pe 0 obligations hai. Answer stable. ✓

Example 7 — Cell C7: Multi-trigger (worst-tier-wins)

Forecast: Teen functions, teen possible tiers. Poora product kaun sa tier inherit karta hai?

Har function ko alag tier karo.
- (i) CV screening ⇒ High (employment named high-risk domain hai).
- (ii) Chatbot ⇒ Limited.
- (iii) Layout recommender ⇒ Minimal. Ye step kyun? Combine karne se pehle har trigger dhundhna padta hai — ek miss karo aur under-comply ho jaoge.
Worst-trigger-wins apply karo. $max (High, Limited, Minimal) = High$ . Ye step kyun? Ek dangerous sub-function harmless ones ke peeche chhup nahi sakta — sabse strong trigger govern karta hai.

Answer: Poora platform = High-risk ⇒ poora 7/7 checklist, plus chatbot sub-part pe transparency label. Screening model mein reward-hacking se savdhan raho — 6.4.2-Reward-hacking dekho — jahan model proxy metric game kare jaise "keyword match" true fit ki jagah.

Verify: Tiers ko Minimal=0, Limited=1, High=2, Unacceptable=3 rank karte hue, platform ka tier $max (2, 1, 0) = 2 = High$ hai. Ye check kyun matters karta hai: pin karta hai ki kaun sa operator sub-tiers combine karta hai. Agar humne galti se average liya hota max ki jagah to $\frac{2 + 1 + 0}{3} = 1 =$ Limited milta — jo ek genuinely high-risk hiring tool ko almost koi safeguards ke bina ship karne deta. Check prove karta hai ki max (mean nahi) safe operator hai. ✓

Example 8 — Cell C8: Real-world word problem

Forecast: Tier guess karo, phir euro cost €10k ke andar guess karo.

Classify karo. AI education access decide kar raha hai ⇒ named High-risk domain. Ye step kyun? Education explicitly listed hai; essay score admission gate karta hai, ek life-altering outcome — ye trigger hai.
Confirm karo ki record-keeping obligation real hai. High-risk har prediction ke liye ek log mandate karta hai, isliye €2-per-essay term ek legally required cost hai, optional tooling nahi. Ye step kyun? Cost term tabhi include kar sakte ho jab koi obligation ise force kare; yahan record-keeping karta hai.
Cost set up karo. First-year cost = fixed compliance + per-essay logging: $C = 40000 + 2 \times 10000$ Ye step kyun? One-off fixed cost ko per-unit variable cost se alag karna standard "fixed + variable" cost model hai.
Compute karo. $C = 40000 + 20000 = 60000 euros.$ Ye step kyun? Projected volume (10,000) ko variable term mein plug karna concrete first-year figure deta hai.

Answer: Tier = High-risk; first-year compliance cost = €60,000. Data governance ensure karna chahiye ki training essays diverse student backgrounds se hon, warna model disparate impact risk karta hai — same failure mode medical example ki tarah.

Verify: $40000 + 2 \cdot 10000 = 60000$ . Ye check kyun matters karta hai: pehle, units euros mein aane chahiye — $euros + (euros/essay \times essays) = euros$ — confirm karta hai ki humne rate ko total ke saath mix nahi kiya. Doosra, variable share check karo: logging €20,000 hai, jo $\frac{20000}{60000} = 33.3%$ total ka hai — ek-tehai bill, isliye ye material hai aur model mein correctly belong karta hai round off nahi hona chahiye. ✓

Example 9 — Cell C9: Exam twist (jurisdiction trap)

Forecast: "Wo American hain, isliye EU unhe touch nahi kar sakti" — sach ya jhooth?

Jurisdiction trigger dhundho. Act kisi bhi AI system pe apply hota hai jiska output EU market mein use hota hai, regardless ki ye kahan develop ya host hua. Ye step kyun? Parent ki definition exactly yahi kehti hai — EU mein placement/use, developer ka address nahi, legal hook hai.
Yahan apply karo. French applicants outputs use karke screen ho rahe hain ⇒ output EU mein use ho raha hai ⇒ in scope. Ye step kyun? Affected log aur decision donon EU ke andar hain, isliye "abroad develop hua" fact irrelevant hai.
Ab function se tier karo. Employment screening ⇒ High-risk ⇒ poora 7/7 checklist. US developer ko ek EU authorised representative bhi appoint karna padega. Ye step kyun? Ek baar in scope mein, classification same function-based test hai jaise kisi bhi example mein — nationality kabhi enter nahi karti.

Answer: Haan, Act apply hota hai (extraterritorial reach); tier = High-risk, 7/7 obligations. "Abroad banaya" koi escape hatch nahi hai.

Verify: Ex. 2 aur Ex. 7 ke saath consistency check: employment + EU mein logon ko affect karna ⇒ har baar High, developer location irrelevant. Ye check kyun matters karta hai: twist sirf ek variable change karta hai — developer ki nationality — jo Act ki scope definition ke according factor nahi hai. Ek correct rule isliye same tier return karni chahiye jaise plain employment cases; agar hamaara answer flip karta, hum galti se jurisdiction ko tier decision mein leak karne dete. Tier unchanged hai ⇒ rule correctly behave karta hai. ✓

Recall Self-test (answer karne ke baad reveal karo)

Police dwara public crowd ka live facial recognition kaun sa tier hai? ::: Unacceptable — prohibited (live + public + mass teeno fire karte hain). US mein bana hiring AI lekin EU applicants screen kar raha hai — in scope? ::: Haan; output EU mein use ho raha hai ⇒ High-risk. Ek product jisme High sub-function aur Minimal sub-function ho, kaun sa tier inherit karta hai? ::: High (worst-trigger-wins / max rule). EduRank ka first-year cost (€40k fixed + €2 × 10,000 logs)? ::: €60,000. Ek purely rule-based tax program "AI" ki tarah marketed — in scope, aur obligation load? ::: In scope (Act rule-based systems cover karta hai), lekin Minimal-risk ⇒ 0 binding obligations. Phone Face ID Minimal kyun hai, Limited nahi? ::: Koi confusion trigger nahi — tum jaante ho apna phone ek machine hai jo tumhe authenticate kar rahi hai, isliye transparency duty ke paas bite karne ke liye kuch nahi.

"Hum ye zahmaat kyun lete hain" ki deeper foundations ke liye, 6.4.5-AI-safety-research aur 6.4.12-Long-term-existential-risks dekho; kyun transparency ko decision explain karna padta hai iske liye 6.4.3-Interpretability-and-explainability dekho. Parent pe wapas: EU AI Act overview.