Responsible AI

We are not going to tell you our AI is unbiased. Nobody can honestly claim that about a large language model, and a vendor who does is either not measuring or not telling you. What we can do is be specific about the design choices that reduce the obvious failure modes, and equally specific about the ones we have not addressed.

• Drafts assessments. From a recruiter's description of a role, it derives the competencies the role needs and writes questions against them. The recruiter reviews and edits before anything is sent.
• Scores answers. Multiple-choice questions are scored deterministically against a key. Written answers and interview responses are scored by a model against a rubric, which also produces the strengths, gaps, and written rationale shown to the recruiter.
• Conducts video interviews. For interview-type assessments, a model conducts a spoken conversation in real time against a prepared brief.

It does not screen résumés, infer anything about a candidate from their profile, or rank candidates against each other. Each submission is scored against the rubric on its own.

These are architectural, so they hold whether or not anyone remembers this page exists.

• The scorer is not told who you are. A scoring request contains the question, the rubric, the answer, and the role context — and nothing else. No name, no email address, no photograph, no country, no education history, no employment history. A model cannot act on a signal it was never given, and this is the single most effective control available.
• Interviews are scored from the transcript as text. The scorer receives the words that were said, never the audio and never the video. Appearance, accent, voice, speech rate, and fluency of delivery are not inputs to the score, because the component that produces the score never sees or hears them.
• The identity photo never reaches a model. It is stored, shown to the recruiter, and deleted on schedule. No model, ours or a provider's, is ever given it, and no facial recognition, template, or matching of any kind is performed on it.
• Assessments are derived from the role, not from a candidate profile. Questions come from the competencies a job requires. The system has no mechanism for scoring someone on a proxy for background — pedigree, school, employer, or years of experience — because it never receives them.
• Scores are sampled, not asserted once. Each written and interview score is generated several times and the median is taken; how much the samples agreed becomes a confidence signal shown to the recruiter. A low-confidence score is visibly a weak signal rather than a number that looks as authoritative as any other.
• Every score comes with its reasoning and its evidence, including verbatim quotes from the answer it was based on. A rationale a recruiter can read is a rationale a recruiter can disagree with, and it is the difference between a defensible decision and a number.
• Empty answers fail closed. A missing answer or an interview with no candidate speech is recorded as unscored rather than scored from nothing. Models will confidently invent a rationale for silence, and a fabricated score is worse than an absent one.

A NorthAssay score is a model's judgement of one piece of work against a rubric a model drafted. It is not a measurement, it has no validated relationship to job performance, and we have not conducted a validation study. Two runs on the same answer can differ, which is why we sample and publish a confidence signal instead of hiding the variance.

It is most useful as a reason to read someone's answer, and least useful as a threshold. A recruiter who sorts by score and contacts the top three has not used the product as designed and has taken on the legal risk that comes with that.

This section is the same length as the previous one on purpose.

• We have not conducted a bias or adverse-impact audit. We have not tested our scoring for disparate impact across race, sex, age, disability, or any other protected characteristic, and we have no results to publish. If you hire in New York City, an independent bias audit is a legal requirement under Local Law 144 and it is one you would have to commission yourself.
• We have not validated scores against job performance. There is no study showing a NorthAssay score predicts anything.
• We have no demographic data, which is what makes the controls in §03 effective and also what makes a fairness audit impossible for us to run without collecting data we have chosen not to collect. This is a genuine trade-off and we have taken the side of collecting less.
• We have not measured how our models treat non-native speakers, dialects, or non-standard phrasing in written answers and interview transcripts. Language models are known to score such text lower on fluency-adjacent criteria. The transcript-only design removes accent from the input; it does not remove phrasing.
• We have not audited the questions the model writes for bias at scale. The recruiter's review is the control, and a recruiter who does not read them has no control.
• We provide no built-in accommodation workflow — no extra-time setting, no alternative format, no screen-reader-verified assessment experience. A candidate who needs an accommodation depends on their recruiter arranging one outside the product.
• We have no formal AI governance: no model risk documentation, no evaluation suite for fairness, no red-team programme, and no third-party review of any of the above.

• You are entitled to know an AI was involved, which is why this page is public rather than in a contract.
• You can decline the identity photo without being blocked from the assessment.
• You can ask for the reasoning behind your score. It exists in written form. Ask the recruiter who invited you; if they do not respond, write to hello@northassay.com and we will tell you what we hold about you.
• You can ask for a human review or an alternative process. Depending on where you live you may have a right to one, and in several jurisdictions a right not to be subject to a decision based solely on automated processing. The recruiter decides your application; we will pass the request on and tell you that we did.
• You can ask us to delete your data, on the terms in the privacy policy.

The legal duties that come with assessing candidates are yours, and they are set out in §05 of our terms. The short version: read the questions before you send them, keep a human in every decision, offer an alternative to anyone who needs one, and find out whether your jurisdiction requires a bias audit or advance notice — because several do, and this page is not one.

If you believe a NorthAssay assessment or score treated someone unfairly, we want to hear about it specifically enough to investigate: email hello@northassay.com with the assessment and what happened. We would rather learn about a failure mode from you than not learn about it. As this page becomes less honest — as the gap list in §05 shrinks — we will update it, and the date at the top will move with it.