AI News Digest, April 7: Why Every AI Hiring Bias Audit Just Got More Urgent

Three stories collided this week and they all point at the same problem: AI is moving into the workforce faster than the guardrails around it. A new look at the University of Washington recruiter study makes the case for an AI hiring bias audit on every screening tool you touch. Brussels is quietly buying itself an extra year on the AI Act. And in Bengaluru, a $1.55 billion sovereign-AI bet is changing what “frontier model” even means. Here is the briefing for HR leaders and founders who do not have time to read ten newsletters.

Top Story: The Case for an AI Hiring Bias Audit Just Got Stronger

HR Brew this week revisited a University of Washington study that should be on every talent leader’s desk. Researchers had 528 people screen resumes alongside large language models with varying degrees of racial bias baked in. Without AI suggestions, recruiters showed almost no bias in their picks. Once a biased model started recommending candidates, recruiters mirrored its choices in roughly 90% of severe-bias cases, even when they said afterward they had spotted the problem. The finding was reported by HR Dive and first published by the University of Washington in November 2025.

If you are running talent acquisition at a 50- to 500-person company, this is the part that should worry you. The “human in the loop” you are relying on as your fairness check is not the safety net you think. Your recruiters are not calmly overruling a biased model. They are nodding along, then writing notes that justify the AI’s pick. The legal exposure you face under the EU AI Act, NYC Local Law 144 and the new wave of state bills does not care whether the bias came from the algorithm or the human reading the algorithm’s output. You own both.

The practical move is an annual AI hiring bias audit on every screening tool in your stack. That means measuring four-fifths-rule disparity by race and gender at each stage, not just at offer. It means logging which candidates the model recommends versus which ones recruiters actually advance, so you can see the mirroring effect in your own data. And it means demanding the model card, training-data sources and bias testing methodology from your AI applicant tracking system vendor, in writing, before the next renewal. If your vendor cannot produce a recent third-party audit, that is your answer.

Sarvam AI Raises $350M for India’s Sovereign AI Stack

India is no longer just an AI buyer

Bengaluru-based Sarvam AI closed a round of $300-350 million led by Bessemer Venture Partners, with Nvidia, Amazon and Prosperity7 joining at a valuation of roughly $1.55 billion, according to Outlook Business. Sarvam is building voice-first agentic models across 22 Indian languages and was selected last year to anchor the IndiaAI Mission’s sovereign foundation model effort.

So what? If you are hiring or running payroll across India, the vendor map is about to look very different. Expect a wave of HR-tech startups to ship Sarvam-powered chatbots, voice screening and policy assistants in regional languages over the next twelve months. For founders building globally, this is a useful counter-data-point to the “everything routes through OpenAI” assumption: language coverage, data residency and procurement leverage all change when there is a credible local frontier lab. This also raises the bar for any AI recruitment tool that wants to win deals in India in 2026.

EU Quietly Pushes High-Risk AI Rules to 2027

The European Commission’s Digital Omnibus proposal would slide the application date for high-risk AI Act obligations from August 2, 2026 out to December 2, 2027 for stand-alone systems and August 2, 2028 for high-risk AI built into products, per OneTrust’s read of the file and the European Parliament’s legislative train. Brussels is framing the slip as alignment with standards bodies that have not yet shipped the harmonised standards companies need to comply.

So what? The relief is real but narrow. The general-purpose AI rules and prohibited practices already in force do not move. If you sell AI hiring software into the EU, you still have to meet transparency, bias-testing and human oversight expectations, you just get more runway to gather evidence and build your compliance package. Do not pause your audit program. Use the extra months to actually finish the work, and watch the Council vote, because member states are split on whether the delay is responsible recalibration or regulatory capture.

Payroll Becomes the Quiet Front Line of an AI Hiring Bias Audit

ADP’s Potential of Payroll 2026 research, summarised by PYMNTS and reflected in ADP’s own November briefing, finds that agentic AI is reshaping payroll into a real-time discipline. ADP-built agents now flag payroll anomalies as they happen, draft localised policies in response to changing state laws and surface turnover-risk scores by department. ADP’s APAC SVP Jessica Zhang put it bluntly: payroll is no longer defined by accuracy and timeliness alone, it is becoming the workforce-intelligence layer the rest of HR runs on.

So what? If you have been treating payroll as a compliance utility, the same fairness questions you ask of your AI payroll automation system are about to land on your desk. An agent that “drafts localised policies” is making interpretive calls about labour law. An agent that scores turnover risk by department is producing data your managers will act on. Founders shopping for a payroll stack should put AI explainability and audit-log access at the top of the RFP, not the bottom. The hiring bias problem and the payroll automation problem are turning into the same problem.

Quick Hits

• OpenAI has crossed $25 billion in annualised revenue and is reportedly preparing for a late-2026 public listing. Watch the eventual S-1 for headcount and HR-cost disclosures.
• OpenAI, Anthropic and Google are coordinating through the Frontier Model Forum to slow Chinese model distillation, an unusual rivals-cooperate moment.
• Anthropic announced an all-stock acquisition of biotech startup Coefficient Bio worth roughly $4 billion, its largest deal to date. Lab-to-vertical M&A is no longer hypothetical.

What This Means for Your Roadmap

If you take one thing from this week, make it the AI hiring bias audit. Put it on the Q2 plan, name an owner, and start with the screening tool that touches the most candidates. The cost of running an audit you do not need is small. The cost of skipping one is a class-action complaint and a regulator who has read the same study you should be reading right now. If you are still mapping which AI tools your HR team actually uses, our guide to top AI tools for HR is a useful starting inventory, and if you are evaluating payroll vendors, our breakdown of payroll software in India covers what to ask about agentic features.

FAQ: AI Hiring Bias Audit

What is an AI hiring bias audit?

An AI hiring bias audit is a structured review of an AI screening or assessment tool that measures whether outcomes differ across legally protected groups, typically using the four-fifths rule. It looks at the model, the training data, and the way recruiters interact with the model’s recommendations.

Is an AI hiring bias audit legally required?

It depends on jurisdiction. New York City Local Law 144 requires an annual independent bias audit for automated employment decision tools. The EU AI Act will require risk management and bias testing for high-risk AI systems used in employment, with the application date now likely shifting to December 2027 under the proposed Digital Omnibus.

How often should we run an AI hiring bias audit?

At least annually for any tool that screens candidates, and again whenever the model is retrained, the vendor pushes a major update, or the candidate population changes meaningfully. Many compliance teams now run lightweight monthly checks on disparity metrics and reserve the full third-party audit for the annual cycle.

Not to be considered as tax, legal, financial or HR advice. Regulations change over time so please consult a lawyer, accountant  or Labour Law  expert for specific guidance.