How to make AI hiring decisions auditable
An AI hiring decision is auditable when any candidate’s score can be reproduced months later and traced back to the logic that actually produced it, not to an explanation written afterwards. Dr. Fredrik Törn, Chief Product Officer at Hubert, sets out the four tests recruitment AI must pass.

An AI hiring decision is auditable when you can take any candidate, months later, and reconstruct exactly why they scored as they did, and when that reason is the real mechanism behind the score rather than an explanation written after the fact. Most tools sold as explainable AI recruitment fail that test. The EU AI Act treats recruitment AI as high-risk and, under Article 13, requires these systems to be transparent enough for a person to interpret their output. That is the floor. Real explainable and auditable AI hiring clears it on four counts. Scores you can reproduce, scores that measure something real, explanations that match the actual logic, and a record of every decision and override.
What compliance theatre looks like
A lot of algorithmic hiring transparency is surface deep. A vendor shows you a confidence score and a short written rationale, then calls the process defensible. The problem is what sits underneath. If the model is a general-purpose LLM, the rationale you see is generated separately from the score. The model assigns a number, then writes a reasonable-sounding account of why. This is post-hoc plausibility. It reads well but is produced after the ranking, with little connection to the maths. A rejected candidate, or a regulator who asks why, gets narrative dressed as evidence.
Reproducibility is the first test
Before an explanation means anything, the score behind it has to be repeatable. Feed the same answers in twice and you should get the same result both times. LLM-based scoring often does not clear this bar. These models are probabilistic, and peer-reviewed research in 2025 found that minor, job-irrelevant changes to a candidate's input can produce considerable shifts in how an LLM ranks them. Hubert scores with deterministic models, so the same input always returns the same output, and the models are versioned and locked for the length of a hiring cycle. A candidate's result does not depend on which server processed their application.
The score has to measure something real
A consistent score is worthless if it measures the wrong thing. Auditability means you can show a score reflects competence, not noise. Decades of selection research rank structured, competency-based interviews among the strongest predictors of job performance, and that is the method Hubert automates. Its assessment models are trained against more than 100,000 evaluations from expert human raters, so the benchmark for a good answer is set by experienced professionals rather than by another language model marking its own work, an approach research shows diverges substantially from human expert judgement. Predictive validity is tracked against real hiring outcomes and refined over time, not assumed once and filed away. When you audit a score, you can point to the method behind it and the evidence that it works.
Faithful explanations, not plausible ones
Auditable recruitment decisions need the reason shown to a recruiter to be the reason the score was produced. Hubert assesses competency responses against behavioural anchors in the STAR framework, with explicit weightings, so the action a candidate describes carries about half the score and the situation, task, and result account for the rest. Hubert applies the weights in the same way for every candidate. What a recruiter sees is that calculation, not a story laid over the top, and you can trace a score back to the words that earned it.
The audit trail that actually matters
AI hiring compliance lives in the record. An auditable process logs the scoring logic, every recruiter action, every override, and the decision history behind each hire. It treats fairness in hiring algorithms as an ongoing duty, with bias checks during model development and monitoring for drift after deployment, not a single certificate from launch day. Candidate data stays within the EU and is never used to train third-party models. The final hire decision stays with a person, because a model cannot stand in front of a tribunal and account for its reasoning. A human can.
A fair process is usually a better one. Human screening carries a well-documented bias. A meta-analysis of hiring field experiments found White applicants receive 36% more callbacks than equally qualified Black applicants, with no improvement across 25 years. An algorithm that discriminates ranks on something other than competence and hands you a weaker shortlist, and auditable AI is what lets you prove you are not repeating that pattern.
So ask any vendor to reproduce the same score twice and to show you the log behind every decision. A real system does both. The ones selling compliance theatre cannot, and auditability is not something you can bolt on once a decision is challenged.
Sources
- EU AI Act, Article 13, Transparency and provision of information to deployers. European Commission AI Act Service Desk. https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-13
- Seshadri, P., Chen, H., Singh, S. and Goldfarb-Tarrant, S. (2025), "Small Changes, Large Consequences", IJCNLP-AACL 2025. https://arxiv.org/abs/2501.04316
- Varshney, A. and Ganuthula, V.R.R. (2025), "Signal or Noise? Evaluating Large Language Models in Resume Screening", arXiv. https://arxiv.org/abs/2507.08019
- Sackett, P.R., Zhang, C., Berry, C.M. and Lievens, F. (2022), "Revisiting Meta-Analytic Estimates of Validity in Personnel Selection", Journal of Applied Psychology, 107, 2040 to 2068. https://doi.org/10.1037/apl0000994
- Quillian, L., Pager, D., Hexel, O. and Midtbøen, A.H. (2017), "Meta-analysis of field experiments shows no change in racial discrimination in hiring over time", PNAS, 114. https://doi.org/10.1073/pnas.1706255114




