PAP - Public Administration Profile | Democratic Accountability for Government AI

Why PAP is Necessary

Current challenges in public AI systems and how PAP addresses them

Lack of Accountability

AI declares "ineligible" without explaining why. Citizens cannot even begin to appeal without understanding the basis for the decision.

Algorithmic Bias Entrenchment

AI trained on historically discriminatory data perpetuates and reproduces unfair decisions systematically.

Loss of Democratic Control

Unelected algorithms make decisions that affect citizens' fundamental rights, without democratic oversight or accountability.

Illustrative Problem Scenarios (Generalized)

Case 1: Welfare Benefits AI Black Box

Benefits eligibility determination

[Applicant] → [AI Decision: "Ineligible"] → [Appeal] → [Request for Decision Rationale]

↓

"The AI's reasoning is proprietary information."

"Details of the scoring model cannot be disclosed."

↓

The citizen loses all means to contest the decision.

Case 2: Recidivism Risk AI Racial Bias

Criminal justice risk assessment

[Defendant] → [AI Risk Score: "High Risk"] → [Bail Denied / Harsher Sentence]

↓

Post-hoc analysis revealed:

"ZIP code" and "arrest history" were primary factors

→ Effectively using proxies for race and socioeconomic status

↓

No mechanism existed to verify this at the time of the decision.

Case 3: Immigration AI Arbitrary Decisions

Visa application processing

[Visa Applicant] → [AI Review: "Denied"] → [Reason: "Comprehensive Assessment"]

↓

"Which factors were problematic?" Unknown

"What should be improved?" Not indicated

"Will reapplication yield the same result?" Unpredictable

PAP Application Scope

Target systems and recorded events

Target Systems

💳

Credit Scoring

Loan, Credit Card Review

Required

🏛️

Welfare Benefits

Social Security, Unemployment, Disability

Required

🛂

Immigration Control

Visa, Refugee Status, Border Control

Required

⚖️

Judicial AI

Recidivism Risk, Sentencing, Bail

Required

👔

Employment & HR

Hiring AI, Promotion, Termination Risk

Required

🏫

Education & Admissions

Admission Review, Scholarships

Recommended

🏥

Medical Resource Allocation

Organ Transplant Priority, Triage

Recommended

Recorded Event Chain (Causal Chain)

Application Receipt

Application ID
Timestamp
Applicant Attributes

Data Collection

Input Data
Data Sources
Missing Value Handling

AI Evaluation

Scoring
Contributing Factors
Model Version

Decision Generation

Decision Result
Reason Code
Threshold Applied

Notification

Notification Content
Disclosure Scope
Receipt Confirmation

Appeal

Review Result
Re-decision Basis
Change History

1. Application Receipt

2. Data Collection

3. AI Evaluation

4. Decision Generation

5. Notification

6. Appeal

PAP Architecture

Data flow and system structure

┌─────────────────────────────────────────────────────────────┐
│                  Public Administration System                │
│  ┌──────────┐   ┌──────────┐   ┌──────────┐                │
│  │ Citizen  │──▶│ AI/ML    │──▶│ Decision │                │
│  │ Data     │   │ Scoring  │   │ Output   │                │
│  └──────────┘   └────┬─────┘   └──────────┘                │
│                      │                                       │
│              ┌───────▼───────┐                              │
│              │  PAP Logger   │ ← Records all decision events│
│              │  (Sidecar)    │                              │
│              └───────┬───────┘                              │
│                      │                                       │
└──────────────────────┼───────────────────────────────────────┘
                       │
               ┌───────▼───────┐
               │  Hash Chain   │ ← Cryptographically chained
               │  + Signature  │ ← Digitally signed
               │  + Factor Log │ ← Complete decision factor record
               └───────┬───────┘
                       │
               ┌───────▼───────┐
               │ Audit Archive │ ← Long-term storage
               │ (GDPR-aware)    │ ← Access control
               └───────┬───────┘
                       │
               ┌───────▼───────┐
               │Citizen Portal │ ← Self-information disclosure
               │(Accountability)│ ← Appeal support
               └───────────────┘

Illustrative Event (concept for Track A, which is not drafted)

{
  "event_id": "019234ab-7c8d-7def-8123-456789abcdef",
  "timestamp_ns": 1734567890123456789,
  "event_type": "ELIGIBILITY_DECISION",
  "agency_id": "AGENCY_XXXXX",
  "provenance": {
    "actor": {
      "type": "AI_MODEL",
      "identifier": "benefit_eligibility_v2.3.1",
      "model_hash": "sha256:abc123...",
      "training_date": "2024-03-15",
      "bias_audit_date": "2024-06-01",
      "bias_audit_result": "PASSED"
    },
    "input": {
      "applicant_id_hash": "sha256:applicant_pseudonymized...",
      "data_sources": [
        {"source": "TAX_AUTHORITY", "freshness_days": 30},
        {"source": "EMPLOYMENT_REGISTRY", "freshness_days": 7},
        {"source": "RESIDENCE_REGISTRY", "freshness_days": 1}
      ],
      "input_hash": "sha256:def456...",
      "missing_fields": ["PRIOR_BENEFIT_HISTORY"],
      "imputation_method": "CONSERVATIVE_DEFAULT"
    },
    "context": {
      "applicable_law": "SOCIAL_WELFARE_ACT_2023_SEC_12",
      "policy_version": "2024-Q2",
      "budget_constraint_active": false,
      "human_override_available": true
    },
    "action": {
      "decision": "ELIGIBLE",
      "confidence": 0.89,
      "eligibility_score": 72.5,
      "threshold_applied": 65.0,
      "explainability": {
        "method": "SHAP",
        "top_contributing_factors": [
          {"factor": "income_below_threshold", "contribution": 0.35, "direction": "positive"},
          {"factor": "dependent_children", "contribution": 0.25, "direction": "positive"},
          {"factor": "employment_status", "contribution": 0.20, "direction": "positive"},
          {"factor": "asset_level", "contribution": -0.10, "direction": "negative"}
        ],
        "counterfactual": {
          "if_income_increased_by": 15000,
          "then_decision_would_be": "INELIGIBLE"
        }
      },
      "appeal_rights": {
        "deadline_days": 30,
        "appeal_authority": "ADMINISTRATIVE_TRIBUNAL",
        "required_documentation": ["INCOME_PROOF", "EXPENSE_RECORDS"]
      }
    }
  },
  "prev_hash": "sha256:789xyz...",
  "signature": "ed25519:...",
  "citizen_disclosure": {
    "simplified_explanation": "Your application has been approved. Main reason: You meet income criteria.",
    "full_explanation_available": true,
    "explanation_request_url": "/citizen/explanation/019234ab..."
  }
}

Regulatory Mapping

PAP design targets against selected regulations (draft)

Regulation	Jurisdiction	Requirement	PAP Support
EU AI Act Article 6(2)	EU	AI systems in the Annex III use cases are high-risk (subject to the Art. 6(3) exceptions)	Design target (draft)
EU AI Act Article 13	EU	Transparency and provision of information to deployers (the right to an explanation of individual decisions is Art. 86)	Design target (draft)
EU AI Act Article 14	EU	Human oversight obligation	Design target (draft)
GDPR Article 22	EU	Right to contest automated decision-making	Design target (draft)
GDPR Article 15	EU	Right of access (self-information disclosure)	Design target (draft)
NYC Local Law 144	NYC	Audit obligation for hiring AI	Design target (draft)
Illinois BIPA	Illinois	Biometric information use restrictions	Design target (draft)
Administrative Procedure Act	Japan	Obligation to provide reasons for administrative decisions	Planned