- A Deaf Indigenous woman was blocked from a promotion because an AI couldn't understand her accent.
The system told her to "practice active listening."
This is what happens when enterprises trust commodity AI with high-stakes decisions. 🧵 - The ACLU filed a complaint against Intuit and HireVue in March 2025.
The candidate—D.K.—had stellar reviews and annual bonuses. But an automated video interview couldn't parse her speech. So it failed her.
Here's the technical root cause: - Standard ASR hits ~15% word error rate on American English.
On Deaf speakers? Up to 78%.
When your transcript is 78% wrong, any model scoring "leadership traits" on top of it is hallucinating.
This isn't a one-off bug. It's an architecture problem. - Most "AI hiring tools" are thin wrappers over general-purpose LLMs. They inherit every bias baked into uncurated internet training data.
They can't audit for fairness. They weren't built to.
The legal landscape has caught up. - Colorado's AI Act (SB 24-205) mandates annual impact assessments for high-risk AI. NYC Law 144 requires independent bias audits. The EU AI Act ties penalties to global revenue.
"We didn't know" is no longer a defense. - In Mobley v. Workday, a court ruled AI vendors can be treated as "indirect employers."
That means the vendor shares liability. Wrapper companies pushing all risk onto customers through disclaimers? That model is collapsing.
The fix isn't abandoning AI. It's building it right. - Adversarial debiasing. Multimodal fusion that doesn't collapse when one channel fails. Human-in-the-loop triggers when confidence drops.
Deep AI, not shallow wrappers.
D.K. asked for a human captioner. She was denied. - In a properly architected system, the model flags low-confidence transcription and auto-routes to human review. That's not a feature request. That's table stakes.
We built our systems around one principle: - If a SHAP analysis shows a candidate was penalized for "prosody" or "accent"—features with zero link to job performance—the model gets flagged and remediated. Automatically.
The real cost of biased AI isn't just lawsuits. - It's the talent you never see. Every qualified candidate screened out by a broken model is innovation your company will never access.
Enterprises still using black-box hiring AI: what's your plan when the first audit hits? - Genuine question—who owns algorithmic accountability at your org? Legal? Engineering? HR? Nobody? #AIFairness #DeepAI
- We wrote the full breakdown—technical architecture, legal landscape, and what Deep AI looks like in practice: https://veriprajna.com/whitepapers/algorithmic-accountability-mandate-enterprise-talent-deep-ai