Why black-box hiring AI is already a liability.
Five jurisdictions have made AI-driven hiring decisions legally risky in the last three years. A pipeline that can't be explained fails every one of these regimes — not in theory, in active enforcement. This page is a short primer on what you owe, and how Axon is structured so you satisfy each of them by construction.
AEDT bias audits for NYC hires.
Any automated employment decision tool (AEDT) used to substantially assist a hiring decision for a role located in or managed from New York City must:
- Undergo an annual bias audit by an independent auditor, published publicly.
- Disclose the use of an AEDT to candidates at least 10 business days before use.
- Allow candidates to request an alternative non-AEDT selection process.
Penalties start at $500 per violation and escalate to $1,500 per subsequent violation per day. "We used ChatGPT to screen résumés" is a per-candidate violation.
Every Axon decision carries a pipeline version hash. Auditors re-run your stored candidate-input on the stored pipeline version and verify the scoring is deterministic. Bias testing happens on the same interface, not as an afterthought PDF.
Hiring AI is high-risk. Treat it that way.
The EU AI Act classifies AI used in employment, worker management, and access to self-employment as high-risk (Annex III, point 4). High-risk systems must have:
- Risk management over the full lifecycle (Art. 9)
- Record-keeping of each operation with tamper-resistant logs (Art. 12)
- Transparency and information to deployers and users (Art. 13)
- Human oversight built into the design (Art. 14)
Maximum administrative fines for non-compliance reach €35M or 7% of global turnover. "The vendor's model did it" doesn't shift liability to the vendor.
The audit log is append-only and hash-chained — tamper-resistant by design. Every decision records the reviewer (the human-oversight surface) next to the pipeline recommendation. The criteria UI makes it impossible to ship a rubric that hasn't been reviewed by a person.
AI in video interviews requires consent + disclosure.
The Artificial Intelligence Video Interview Act requires employers using AI to analyze interview videos to: notify applicants, explain how it works, obtain consent, limit sharing, and destroy copies on request.
Axon doesn't touch video in the pilot, but the disclosure-and-destruction scaffolding applies equally to résumé analysis and is exposed via per-candidate data retention controls.
Facial recognition in hiring requires consent.
Maryland prohibits using facial recognition services during pre-employment interviews without written candidate consent.
Axon's node palette does not include a facial-recognition primitive. The audit log would surface such a node immediately if introduced.
Colorado AI Act: duty of care + impact assessments.
Colorado will impose a duty of reasonable care on deployers of high-risk AI to prevent algorithmic discrimination, including mandatory impact assessments and documentation obligations very similar to the EU AI Act's.
Axon produces impact-assessment-ready artifacts by default: pipeline version graphs, exclusion lists per criterion, distribution of scores per run, reviewer-override rates.
This is an alpha. Not legal advice.
Axon is a portfolio pilot. It is structured to satisfy the above regimes, but any production deployment requires counsel review, real audit partners, and per-jurisdiction rollout plans. Talk to your lawyers.