AI use and bias

Draft: not legal advice, pending review by counsel. This page is a working draft, not a published AI use statement. Placeholders in square brackets must be replaced and the whole document reviewed by qualified counsel before launch.

1. What AI does in boson

boson uses language models to help grade some screening modules. Grading is decision support. It is not an employment decision, and it does not replace the recruiter or hiring team.

A grade is intended to be explainable and evidence-based: the scorecard shows the dimensions, the reasoning, the evidence it relied on and a confidence level. Each grading record stores the model and prompt version used, so a result can be traced back to the configuration that produced it.

2. No automatic rejection

No candidate is automatically rejected by the grading system. The product has no automatic "rejected" state. Recruiters review results and make the hiring decision. Where a result is low-confidence or the evidence is weak, the product is designed to flag it for human review.

Human overrides require a reason and are recorded in the audit trail. A recruiter can disagree with a grade and record why.

3. How we try to reduce bias

• Grading prompts are designed around job-related evidence rather than impressions.
• Some prompts remove names, contact details and protected attributes from the context sent to the model.
• Scores are broken into dimensions so a reviewer can see which part of the submission drove the result.
• Every score is linked to evidence, so a reviewer can check the source rather than trusting a number.
• Low-confidence results are surfaced for human review instead of being treated as final.

These are risk-reduction measures. They do not make a model unbiased, and they do not remove the need for human judgement.

4. What we do not claim

We do not publish accuracy, adverse-impact or fairness statistics on this page because we have not validated them for publication. We do not claim that a score predicts job performance. We do not claim that the system is free of bias.

[Add any validated evaluation results here only when the methodology and numbers can be supported. Do not add estimates.]

5. Limitations and human oversight

Language models can be wrong, can miss context and can be influenced by the wording of a job description or rubric. Recruiters remain responsible for their hiring decisions and for complying with the employment, discrimination and AI laws that apply to them.

Customers should:

• review the job description and rubric for job relevance and unlawful criteria;
• inspect the evidence behind a score before relying on it;
• keep a human in the decision for any adverse outcome;
• follow their own accommodation and appeals processes.

[Confirm the customer obligations that belong in the terms and any required human-oversight language.]

6. Monitoring and improvement

We review grading behaviour when prompts or models change and keep versioned records so results can be compared. [Confirm the review cadence, who signs off a model or prompt change, and how incidents are handled.]

7. Contact

Questions about AI use or a specific grade: [contact email].

This draft is not legal advice.