AI Ethics Guide — Using AI Responsibly in Hiring
What "responsible AI" means concretely in a hiring context, not as an abstract principle.
"Use AI responsibly" is easy to say and hard to operationalise. This page states, concretely, what that means in the specific context of AI-assisted hiring — grounded in the architecture choices this platform actually makes, not general principles.
On this page
- Principle 1 — never let AI produce the final decision-bearing number
- Principle 2 — strip identity before the model sees it
- Principle 3 — publish the methodology
- Principle 4 — keep a human in the loop, structurally not just nominally
- The regulatory landscape, briefly and honestly
- A practical checklist for evaluating any AI hiring tool
- Frequently asked questions
01Principle 1 — never let AI produce the final decision-bearing number
An LLM asked to output a hiring score directly is a well-documented failure mode: the same inputs can yield different outputs on different runs, because generative models predict tokens, not compute deterministic functions. Responsible use means constraining AI to extraction/reading, with fixed code computing anything decision-bearing.
02Principle 2 — strip identity before the model sees it
Personal identifiers should be removed from a CV before any AI reads it for evaluation purposes — reducing the surface area for the model to (even unintentionally) pattern-match on demographic proxies rather than qualifications.
03Principle 3 — publish the methodology
A scoring system's inner workings should be inspectable by the people affected by it, not a trade secret — Expertini publishes the CMS formula as a citable academic paper specifically so this principle isn't just asserted.
04Principle 4 — keep a human in the loop, structurally not just nominally
"Human in the loop" should mean a human can meaningfully review and override, not a rubber-stamp step — every score comes with dimension-level rationale a recruiter can actually evaluate, not just a bare number to accept or reject blindly.
05The regulatory landscape, briefly and honestly
Automated hiring tools are moving from unregulated to regulated faster than most HR teams track. NYC Local Law 144 requires bias audits and candidate notice for automated employment decision tools used on NYC candidates. The EU AI Act classifies AI systems used in recruitment as high-risk, bringing documentation, transparency, and human-oversight obligations. Illinois and Maryland regulate AI analysis of video interviews; GDPR Article 22 already restricts purely automated decisions with significant effects. The common thread across all of them: you must be able to explain what the tool does, show a human meaningfully decides, and produce documentation when asked. A tool whose scores can't be reproduced makes every one of those obligations harder. This isn't legal advice — but if your counsel asks "can we explain this tool's decisions," the architecture on this platform is built so the answer is yes.
06A practical checklist for evaluating any AI hiring tool
Six questions worth putting to every vendor, including us. One: does the model produce the decision-bearing number, or does deterministic code? Two: is the methodology published, or "proprietary"? Three: what identity information does the model see when it evaluates? Four: can a past score be reproduced and explained months later, to an auditor or a candidate? Five: does anything get auto-rejected without human action? Six: what happens to candidate data, and when is it deleted? Expertini's answers are documented on the pages linked throughout this guide — scoring architecture, privacy practice, ranking behaviour. Any vendor who answers all six crisply is taking the problem seriously; vague answers to questions one and four are the reddest flags, because those are the properties you can't retrofit.
Frequently asked questions
Does using Expertini ATS make us compliant with LL144 / the EU AI Act automatically?⌄
Is it ethical to use AI in hiring at all?⌄
At a glance
- Never let AI output the final decision-bearing number directly
- Strip identity before AI reads a document for evaluation
- Publish the methodology — inspectable, not a trade secret
- Human review must be meaningful, with real rationale to evaluate
- Recruitment AI is high-risk under the EU AI Act — plan for documentation
- Six vendor questions that separate substance from marketing
See ai ethics guide — using ai responsibly in hiring on your own hiring.
Bring a real job description to a 30-minute demo — free trial included.
Book a demo