WHIZZHIRE

Legal

AI Disclosure

Effective 10 July 2026 · Version 1.0

Whizz Hire uses large language models in a hiring pipeline. That sentence obligates us — and our customers — under a growing set of laws. This page states plainly what the AI does, what it is prevented from doing, and what candidates can ask for. It is written for candidates, customers, and regulators alike.

1. What the AI does, stage by stage

  • Parse. A vision-capable model reads CV pages as images and produces a structured extract (roles, dates, skills, education). Nothing is scored here.
  • Anonymize. Before any scoring, identity fields — name, email, phone, photo, address, and age markers — are stripped from the extract by deterministic code. The scoring model receives work history, not identity.
  • Score.A model scores the anonymized extract 0–100 against each criterion of the employer's rubric — pointwise (one candidate at a time, never comparing candidates in a single prompt), repeated three times with shuffled input order and aggregated, because single-pass rankings are unstable. The overall score is the weighted average of the rubric weights, which the employer sets and can change; changing weights recomputes ranks arithmetically, with no further model involvement.
  • Evidence. For top-ranked candidates, a model must justify each criterion score with quotes from the CV. Every quote is programmatically validated as a verbatim match against the extracted CV text and carries a page reference. A quote that fails validation is discarded and the result flagged — fabricated evidence is a hard error, not a risk we accept.
  • Rank and top-K. Arithmetic, not AI: scores sort, the top K are flagged as recommendations.
  • Interview kits and job descriptions.Generative drafting on request — interview questions grounded in the candidate's CV claims, and job descriptions in English and Arabic. Both are editable drafts for humans.

2. The human-in-the-loop guarantee

The AI never rejects a candidate.This is enforced in the product, not promised in a policy: the system cannot set an application to rejected. AI screening may move an application from "new" to "screening"; every advancement or rejection is an action taken by a named person in the employer's dashboard, recorded in an append-only decision log with actor and timestamp. When a human decision disagrees with the AI recommendation, that override is itself recorded — evidence that review is real, not ceremonial.

3. Anonymization before scoring

Published research is unambiguous: removing names eliminates nearly all measured demographic bias in LLM resume scoring, while instructing a model to "be unbiased" does not. So we do the former and don't rely on the latter. We also never rank candidates by raw embedding similarity on non-anonymized CVs, and we monitor score distributions in production for adverse impact.

4. Evidence citation policy

A score an employer cannot inspect is a score a candidate cannot contest. Screening results carry per-criterion scores, the weights used, and verbatim CV quotes with page references — validated against the document, as described above. This is the record an employer uses to explain an outcome, to a candidate or to a regulator.

5. What candidates can ask for

  • Human review. You may request that a person review any outcome influenced by automated screening. Requests go to the employer you applied to — or to support@whizztech.ai, and we will route them. The product gives employers a one-screen view of your scores and evidence to perform that review.
  • An explanation. The criteria used, their weights, your scores, and the CV evidence behind them exist for every screened application and can be shared with you by the employer.
  • Your data. Access, correction, export, and deletion, per the Privacy Policy.

6. Jurisdictions, in plain language

  • NYC Local Law 144 — automated employment decision tools require an independent bias audit and advance notice to NYC candidates. Employers using Whizz Hire for NYC roles must provide that notice (we supply templates); a vendor-level bias audit is planned ahead of any concerted US market push, and this page will link it when published.
  • Illinois — from 1 January 2026, Illinois law (HB 3773) requires notifying candidates when AI is used in hiring decisions and prohibits AI use that discriminates. Notice templates are included in the product.
  • Colorado — the Colorado AI Act (effective 2027) treats hiring AI as high-risk and will require impact assessments and explanation of adverse decisions. Our decision log and evidence records are built to feed those artifacts.
  • GDPR Article 22— candidates have the right not to be subject to solely automated decisions with significant effect. Following the CJEU's SCHUFA reasoning, we treat AI ranking as an automated decision wherever no human meaningfully reviews it — which is why the human disposition gate is mandatory, not optional, and why override tracking exists.
  • EU AI Act — AI systems used in recruitment are classified high-risk (Annex III), with provider obligations phasing in through 2 December 2027. Our conformity program (risk management, logging, human oversight documentation) is in progress against that deadline. We will not claim conformity before it is assessed.

None of the above is legal advice, and employer obligations vary by role location. What we commit to: the mechanics in Sections 1–4 hold everywhere, for every customer, in every jurisdiction — they are how the product works, not a regional setting.

7. What we do not do

  • We do not train models on CVs, applications, or screening outputs.
  • We do not reject candidates by AI — no auto-rejection exists in the product.
  • We do not show fabricated evidence — unvalidated quotes are dropped and flagged.
  • We do not score identity — names, photos, and contact details are stripped before scoring.
  • We do not analyze faces, voices, or video, and we do not infer protected characteristics.

8. Questions

If your compliance team needs specifics — model versions per stage (recorded on every screening), data flows, subprocessors, or the exact fields in the decision log — write to support@whizztech.ai. Engineers answer this inbox.