AI System Transparency Report

1. AI System Identification

2. Intended Purpose

HumanBaseIQ assists employers in the recruitment process by providing AI-powered analysis of candidate applications. The system provides decision-support information only. No hiring decision is made automatically by the AI system. A qualified human reviewer must always make the final decision.

AI capabilities:

• CV parsing and structured data extraction
• Skills-based candidate-to-role matching
• Text-based AI interview (20 role-specific questions)
• Voice analysis (limited-risk): audio recording → transcript only; no paralinguistic/emotional-tone inference
• Video analysis (limited-risk): video recording → transcript only; no facial-expression, body-language, or emotion inference
• Bias detection: automated flagging of potentially discriminatory wording in transcripts

3. Data Processing Methods

3.1 Text Analysis (CV + Interview Answers)

Processed by Claude AI (Anthropic, US-based). Text is sent via API, processed, and response returned. No text is retained by the AI provider beyond the API call.

3.2 Voice Analysis (Opt-in)

Processed by AssemblyAI (US-based, GDPR DPA in place). Audio is uploaded for transcription and sentiment analysis. PII is automatically redacted (names, emails, phone numbers, dates of birth). Transcripts and audio are deleted from AssemblyAI servers immediately after analysis.

3.3 Video Analysis (Opt-in) — Limited-Risk (Article 50)

Video recording with explicit candidate consent. Audio track is transcribed (AssemblyAI, deleted after transcription). The video stream is retained encrypted (AES-256) for the data-retention period configured by the employer, and is available only for human reviewer playback alongside the transcript. No AI analysis is performed on the video stream itself.

4. Human Oversight (Article 14)

• All AI scores are capped at 85/100 — no candidate can be fully scored without human review
• Bias flags trigger mandatory human review before any decision
• The employer's HR team makes all final decisions; the AI provides supporting data only
• Candidates who refuse biometric analysis are not penalised — text-only interview is always available
• Employers can override or disregard any AI recommendation

5. Bias Mitigation & Fairness

• Automated detection of sensitive topics: age, family status, religion, nationality, disability, marital status, children
• Any mention of protected characteristics triggers mandatory human review
• Scoring algorithm does not use protected characteristics as inputs
• Quarterly bias audit available (NYC Local Law 144 compliant)
• UK Equality Act 2010 Schedule 19 compliance embedded
• Disparate impact analysis available for employer review

6. Candidate Rights

7. Data Retention

8. Legal Compliance by Jurisdiction

9. Technical Architecture

Video Upload (encrypted TLS 1.3)
  │
  ├─→ Face Analysis (on-server, self-hosted)
  │     • face-api.js + TensorFlow.js
  │     • 7 emotion categories
  │     • 68 facial landmarks → eye contact
  │     • Frames: RAM only, never persisted
  │
  ├─→ Voice Analysis (AssemblyAI, DPA)
  │     • PII auto-redacted before analysis
  │     • Transcript + audio deleted post-analysis
  │     • Sentiment, pace, pause patterns
  │
  ├─→ Transcript Analysis (Anthropic Claude, via AssemblyAI LeMUR)
  │     • Normal path for every interview
  │
  ├─→ [FALLBACK ONLY] OpenRouter → Google Gemini 2.5 Flash
  │     • Fires only if the analysis above is unavailable
  │     • First 6,000 characters of transcript sent
  │     • Degraded output: no per-segment sentiment or pace
  │     • Routed with data_collection: deny — providers that
  │       retain or train on prompts are excluded
  │
  └─→ Scoring Engine (on-server)
        • Max score: 85/100 (human review enforced)
        • Bias flags → mandatory review
        • No protected characteristics used

10. Contact & Complaints