Interview AI + transcript scoring

What this is

Interview AI takes recorded video, audio, or transcripts of candidate interviews + scores them against a competency rubric. Typical features: text-transcript competency extraction (drafter / classifier), structured-question-response scoring (closed-enum rubric), candidate-comparison ranking against a job-opening profile.

Less common but legally riskier: facial-expression analysis, voice-stress / emotion classification, behaviour-style prediction. These sit close to AI Act Art. 5 prohibitions (emotion recognition in workplace + education is largely prohibited from August 2026) — most reputable HR-tech SMBs avoid them or use them in carefully-scoped narrow contexts only.

For the legitimate transcript / structured-scoring use case: the substrate records each interview-scoring decision as an L12 entry. The rubric (spec) is hash-committed to every entry — any retroactive rubric tweak is detectable. The chain export underpins disparate-impact analyses by candidate cohort, and supports both NYC Local Law 144 audits + GDPR Art. 22 challenges.

How Promethean addresses this

• Rubric stability (specHash binding)

  The interview rubric is encoded as the spec's outputSchema. Each L12 entry carries specHash — a rubric change produces a new hash, visible in the chain. Retroactive rubric changes to make outcomes look better aren't possible without breaking the chain.

• Behaviour-analysis claims — what the substrate does + doesn't do

  Promethean records the decisions; it does NOT make behaviour-analysis a legal AI use. If your interview AI does facial-expression / emotion inference in EU workplace contexts, AI Act Art. 5(1)(f) prohibits it (effective 2 February 2025 with the rest of Chapter II) — and no evidence layer makes a prohibited practice lawful. For transcript-based or competency-rubric-based scoring (which is lawful), Promethean provides the audit trail.

• Candidate notification + data-deletion (IL AI Video Interview Act)

  Notification is operator-side workflow (form text, email cadence). For data deletion: when the candidate requests deletion, your platform handles the source-data deletion; Promethean's chain retains only SHA-256 hashes (no PII), so chain retention doesn't itself trigger the deletion right — the operator-side data store is the right venue.

• Reviewer-gate on hiring-influencing scores (Art. 22)

  Reviewer-gate primitive: always-human on any decline; on-low-confidence on advance decisions. The L12 entry records reviewer identifier + verdict. Override rate over time is the empirical evidence of meaningful intervention.

• Multi-employer-customer audit isolation

  R4 tenantId per L12 entry. Per-employer chain export filterable by tenant; bias-audit firms see only their employer's data via verify.mjs.

What this looks like in practice

Hypothetical: HR-tech SMB during AEDT bias-audit + Aug-2026 AI Act §5 audit

An HR-tech SMB ships transcript-based interview-scoring AI to 80 employer customers, including several NYC-headquartered ones (NYC AEDT applies) + several EU operations (AI Act + GDPR applies). In Q1 2027 the SMB faces: a NYC bias audit for its NYC-employer customers + a GDPR Art. 22 complaint from an EU candidate + an AI Act Art. 5 audit checking the product doesn't do prohibited emotion recognition. Without Promethean: three concurrent log-archeology exercises, partial answers to each regulator. With Promethean: chain-export filtered by tenantId for the NYC bias audit; chain-export for the affected candidate for the Art. 22 inquiry; the spec definitions (in the chain) prove the AI does transcript-based competency scoring, not emotion recognition. All three regulators verify with verify.mjs. The SMB exits with corrective recommendations rather than enforcement.

Frequently asked