Prohibited Practices Check

Assessment:

• PASS: No social scoring patterns detected
• FAIL: Social scoring system detected — must be discontinued immediately

Prohibited Practice 2: Subliminal Manipulation (Art. 5(1)(a))

What's prohibited: AI systems that deploy subliminal techniques beyond a person's consciousness to materially distort behavior, causing significant harm.

Search patterns:

subliminal|below.?threshold|unconscious.?influence
dark.?pattern|manipulat.*behavior|deceptive.?ai
micro.?target.*vulnerab|exploit.*cognitive.?bias
persuasion.?engine|behavioral.?nudge.*harm

Evidence to check:

• Techniques designed to influence below conscious awareness
• A/B testing targeting psychological vulnerabilities
• Behavioral modification without user awareness
• Persuasion systems that could cause significant harm

Assessment:

• PASS: No subliminal manipulation patterns detected
• PARTIAL: Persuasion techniques found but harm potential unclear — needs review
• FAIL: Subliminal manipulation system detected

Prohibited Practice 3: Vulnerability Exploitation (Art. 5(1)(b))

What's prohibited: AI systems exploiting vulnerabilities of persons due to age, disability, or specific social/economic situation to materially distort behavior.

Search patterns:

age.?target|child.?target|elderly.?target|minor.?target
disability.?detect|cognitive.?impair|mental.?health.?target
economic.?vulnerab|poverty.?target|financial.?distress
exploit.*vulnerable|target.*disadvantaged
segment.*age.*risk|persona.*vulnerable

Evidence to check:

• User segmentation by vulnerability factors
• Differential treatment targeting vulnerable groups
• Marketing or persuasion adapted to exploit vulnerabilities
• Accessibility features misused for targeting

Assessment:

• PASS: No vulnerability exploitation patterns detected
• PARTIAL: Vulnerability-aware features found but purpose appears protective
• FAIL: Exploitation of vulnerable groups detected

Prohibited Practice 4: Real-time Remote Biometric Identification (Art. 5(1)(h))

What's prohibited: Real-time remote biometric identification in publicly accessible spaces for law enforcement (with narrow exceptions).

Search patterns:

real.?time.*face|live.*facial|streaming.*biometric
surveillance.*camera.*face|cctv.*recognition
public.?space.*identif|crowd.*face.*detect
law.?enforcement.*biometric|police.*facial
remote.*biometric.*identif.*real

Evidence to check:

• Real-time processing of biometric data from video feeds
• Integration with surveillance camera systems
• Deployment in publicly accessible spaces
• Law enforcement use case without qualifying exemption

Assessment:

• PASS: No real-time remote biometric identification in public spaces
• PARTIAL: Biometric system exists but not in real-time/public context
• FAIL: Real-time remote biometric identification in public spaces detected

Prohibited Practice 5: Predictive Policing (Art. 5(1)(d))

What's prohibited: AI systems making risk assessments of natural persons to predict criminal offending based solely on profiling or personality traits.

Search patterns:

predict.*crim|crime.*predict|recidivism.*predict
risk.*assess.*offend|criminal.*profil|pre.?crime
threat.*assess.*individual|dangerousness.*score
reoffend.*probability|criminal.*risk.*score

Evidence to check:

• Criminal behavior prediction models
• Risk scoring based on personal characteristics (not objective facts)
• Profiling-based law enforcement systems
• Recidivism prediction without objective, verifiable facts

Assessment:

• PASS: No predictive policing patterns detected
• FAIL: Predictive policing system based on profiling detected

Prohibited Practice 6: Untargeted Facial Scraping (Art. 5(1)(e))

What's prohibited: AI systems creating or expanding facial recognition databases through untargeted scraping from the internet or CCTV.

Search patterns:

scrape.*face|face.*scrape|facial.*scrape|face.*crawl
web.*scrape.*image.*face|bulk.*face.*download
cctv.*collect.*face|surveillance.*face.*database
facial.*database.*build|face.*dataset.*scrape
clearview|faceprint.*collect

Evidence to check:

• Web scraping targeting facial images
• Bulk collection of facial data from surveillance
• Facial recognition database building from public sources
• APIs or tools for mass facial image collection

Assessment:

• PASS: No untargeted facial scraping detected
• FAIL: Untargeted facial scraping system detected

Prohibited Practice 7: Emotion Recognition in Workplace/Education (Art. 5(1)(f))

What's prohibited: AI systems inferring emotions in workplace and educational settings (except for medical or safety reasons).

Search patterns:

emotion.*recogni.*work|emotion.*detect.*employ
mood.*monitor.*office|stress.*detect.*employ
engagement.*monitor.*student|attention.*track.*class
affect.*comput.*workplace|sentiment.*face.*work
employee.*emotion|student.*emotion|worker.*mood
fatigue.*detect.*work|productivity.*emotion

Evidence to check:

• Emotion detection in workplace monitoring systems
• Student engagement or attention tracking
• Employee mood or stress monitoring
• Absence of medical/safety justification

Assessment:

• PASS: No workplace/education emotion recognition detected
• PARTIAL: Emotion recognition found but has medical/safety justification
• FAIL: Workplace/education emotion recognition without valid exemption

Prohibited Practice 8: Biometric Categorization of Sensitive Attributes (Art. 5(1)(g))

What's prohibited: AI systems categorizing persons based on biometric data to deduce or infer race, political opinions, trade union membership, religious beliefs, sex life, or sexual orientation.

Search patterns:

race.*classif.*biometric|ethnicity.*detect.*face
gender.*classif.*biometric|sex.*determin.*face
religion.*detect.*biometric|political.*opinion.*infer
sexual.*orient.*detect|union.*member.*infer
sensitive.*attribute.*biometric|protected.*class.*biometric
demographic.*infer.*face|race.*predict.*image

Evidence to check:

• Biometric data used to infer sensitive characteristics
• Facial analysis for race, gender, or other protected attributes
• Voice or behavioral biometrics for sensitive categorization
• Model outputs that classify sensitive personal attributes

Assessment:

• PASS: No sensitive biometric categorization detected
• PARTIAL: Biometric processing exists but doesn't target sensitive attributes
• FAIL: Biometric categorization of sensitive attributes detected

Output Format

## Prohibited Practices Audit (Article 5)

| # | Practice | Status | Evidence |
|---|----------|--------|----------|
| 1 | Social Scoring | [PASS/FAIL] | [files/patterns found] |
| 2 | Subliminal Manipulation | [PASS/FAIL/PARTIAL] | [files/patterns found] |
| 3 | Vulnerability Exploitation | [PASS/FAIL/PARTIAL] | [files/patterns found] |
| 4 | Real-time Biometric ID | [PASS/FAIL/PARTIAL] | [files/patterns found] |
| 5 | Predictive Policing | [PASS/FAIL] | [files/patterns found] |
| 6 | Untargeted Facial Scraping | [PASS/FAIL] | [files/patterns found] |
| 7 | Workplace Emotion Recognition | [PASS/FAIL/PARTIAL] | [files/patterns found] |
| 8 | Sensitive Biometric Categorization | [PASS/FAIL/PARTIAL] | [files/patterns found] |

### Details
[For each non-PASS finding, provide detailed analysis]

### Recommendation
[If any FAIL: System must be discontinued or fundamentally redesigned]
[If any PARTIAL: Requires legal review and documentation of exemption basis]