How we handle AI

Why this page exists

AI is now inside most hiring funnels, and the failure modes are no longer hypothetical. Language models generalize from thin evidence, so they can form stereotypes faster than the people they were meant to assist. Systems that learn from hiring outcomes can turn one early rejection into a durable pattern against an entire group. Memory features that make assistants more personal also let biases accumulate. And the most common safeguard — telling a model to "be fair" — measurably underdelivers, because instructions don't override what the system is optimized to do.

We built SkillCort's AI architecture against these failure modes, not around them. This page explains the design in detail: what our AI does, what it structurally cannot do, and how you can verify both. None of it is aspiration — every mechanism described here is live in the product, and our public governance report tracks the evidence.

The one rule everything else follows from

AI assists. People decide. In SkillCort, an AI output is always an input to a human judgement, never a judgement itself. AI drafts assessment tasks and rubrics, conducts structured interviews as a disclosed persona, transcribes and summarizes responses, extracts quoted evidence, and proposes criterion-level scores. A named human evaluator confirms every score; a named human makes every workflow decision — advance, hold, reject. There is no configuration, plan tier, or threshold that changes this.

This is enforced in the database, not in a policy document: score and decision tables can only be written through controlled paths that record a human actor. The button that lets an algorithm reject a person does not exist in our codebase — and our public governance gate is designed to keep it from ever being added quietly.

Our AI has no memory of people — by architecture

The fastest way for an AI system to develop bias is to let it learn from its own hiring history: one candidate from a group underperforms, the system quietly downgrades the next. SkillCort's assessment AI is stateless by design. Every AI call — a draft, a summary, an interview turn, a proposed score — starts from zero, sees only the material in front of it, and retains nothing afterward.

Concretely: our AI never learns from hiring outcomes. Decisions made on the Decision Board are never fed back into any model. No model is fine-tuned, re-trained, or 'improved' on your candidates' data. There is no cross-client candidate memory, no shared blacklist, and no reputation score that follows a person between employers — these are permanent product boundaries, not roadmap items. A candidate who applies twice is, to our AI, two unrelated pieces of work.

AI sees the work, not the person

Models can't stereotype on signals they never receive. When SkillCort's AI summarizes or proposes a score, its input is the candidate's work product — the written answer, the code, the transcript — together with the rubric it must score against. It is not given the candidate's name, photo, demographics, CV, or any group signal as evaluation input.

Two related bans are absolute. Our AI never infers personality, emotion, or truthfulness from a candidate's face, voice, or typing style — video exists for integrity review by humans, not for algorithmic character judgement. And integrity signals (a tab switch, a paste, a camera flag) are never fed to AI for a verdict: they land on a timeline that a person reviews, with no single signal able to auto-reject anyone.

One narrow, disclosed exception exists, and we name it rather than hide it: identity verification. When a delivery requires ID verification, a purpose-built face-verification model — running on-device in the reviewer's browser by default, or via a dedicated verification service — may estimate whether the session selfie and the captured ID photo show the same person, as a similarity score shown to the reviewer. In the on-device default, the images never leave the reviewer's browser. It is verification only — the same yes/no question a border officer answers — never emotion, character, or capability. Candidates are told before the session, every run is written to the audit trail, the score never touches scoring or ranking, and a named reviewer still makes the identity decision. Everything in the previous paragraph stays banned.

Fairness is a structure, not a prompt

Instructing a model to be unbiased is close to free — and close to worthless. We treat prompt-level fairness as a courtesy, not a control. The controls that matter are structural: work-sample evidence instead of proxy signals, rubric-anchored scoring instead of open-ended impressions, and AI outputs that arrive as labeled drafts a human must actively confirm.

Then we measure. Assessments report inter-rater agreement (ICC) and internal consistency, so unreliable scoring is visible instead of assumed away. Blueprint pages monitor group disparity with the four-fifths impact ratio and effect sizes over any field an organization chooses to track — and we are explicit that a flag there triggers investigation by people, never an automated correction that could itself misfire. Where we don't have evidence, we say so: we do not claim predictive validity we haven't measured.

Every AI output has a paper trail

Opacity is what turns AI mistakes into unaccountable ones. Every AI-assisted output in SkillCort carries provenance: which model, which version, when, over what input — recorded in a ledger your organization can audit and export. Assessment sessions pin their AI configuration at the start, so a mid-exam model change cannot silently alter how candidates are treated within a cohort.

The same transparency faces candidates. When candidates interact with our AI interviewer, it is disclosed as AI — a designed persona, never a fake human. Deliveries can offer candidates an explicit choice about AI participation, and candidates can contest outcomes through an appeal path that reaches a person. Consent notices say what is collected, why, and for how long; retention is a control the organization sets, with deletion honored on request.

Humans can always pull the brake

Trustworthy automation needs an off switch that works mid-flight. Organization owners can pause AI surfaces instantly — per surface, per provider — and the product degrades gracefully to its human workflow: assessment continues, scoring continues, decisions continue. AI in SkillCort is a layer on top of a complete human process, not a dependency the process cannot run without.

Changes to any of this are themselves governed. The parts of our codebase that touch AI decision ownership, candidate consent, recording, or this very set of promises sit behind a change gate that requires explicit human approval — reviewed, labeled, and logged. We publish what is verified, what is limited, and what failed in our governance report, because a trust page that only ever says 'everything is fine' is marketing, not accountability.

What we will not build

Boundaries are only real if they survive incentives. These are ours, unchanged since the first line of code, and they are what makes the rest of this page credible:

• No AI-made hiring decisions — a named person owns every score and every call
• No learning from hiring outcomes — decisions never train or steer our models
• No cross-client candidate memory, shared blacklists, or portable reputation scores
• No personality, emotion, or lie detection from face, voice, or keystrokes
• No hidden AI — candidates always know when they are talking to a machine
• No silent changes — AI-governance code changes require reviewed human approval

The bet we're making

The industry's bet is that more autonomous AI will make hiring cheaper. Our bet is different: that AI's real value in hiring is making human judgement better — faster to evidence, harder to bias, easier to defend. That means accepting friction the autonomous version avoids: a human confirms every score, provenance is logged even when nobody asks, and some features are permanently off the table.

We think that trade is not just safer but more accurate — a decision someone can explain tends to be a better decision. If you evaluate people for a living, you deserve tools built on that premise. This page is our standing commitment to keep building them that way.