Most people functions aren't equipped for what's coming
Three stories landed this week that, read individually, sit comfortably in the technology section. Read together, they describe a single converging problem that lands squarely on HR's desk, and most people functions are not yet equipped to handle.
The first: Anthropic, maker of some of the most widely deployed AI tools in the American workplace, published a report warning that the human role in AI development is already "narrowing at each step" and calling for a global pause on the most powerful AI systems.
The second: a Stanford-led study of four million job applications, the largest examination of AI hiring algorithms ever conducted, found "clear racial disparities" in AI screening outcomes, with Black and Asian candidates disproportionately rejected, and discovered that the same algorithmic models were being shared invisibly across dozens of employers.
The third: the Financial Times reported that Google DeepMind, Anthropic and Meta have begun studying whether AI systems might be conscious, hiring philosophers and psychologists to explore what obligations humans might have toward them.
Each of these stories has direct HR implications. None of them is being discussed in people functions at the level the risk warrants.
The most urgent story is already in your ATS
Start with the Stanford study, because it is the most immediately actionable. Researchers from the Stanford Institute for Human-Centered AI analyzed four million job applications submitted via the Pymetrics platform across 156 employers between December 2018 and December 2022. One in ten positions showed "adverse impact" against Black applicants, the federal standard for a selection rate below four-fifths of the highest-selected group. One in twenty positions showed adverse impact against Asian applicants.
READ MORE: Amazon launches worker robot that takes conversational instructions
The finding that should concern every HR technology leader most is this: 42 algorithmic models were shared across different employers. Candidates rejected by one company's algorithm were systematically likely to fail at others using the same model. They had no way of knowing this was happening. Their applications looked independent; their outcomes were not.
"As a single vendor comes to dominate decision-making in a space, their quirks or shortfalls can be present across that entire sector in a way that wasn't possible before," said Kathleen Creel, a co-author of the study and assistant professor of philosophy and computer science at Northeastern University.
HRD America has reported that 62% of workers believe AI will run their entire hiring process by end of 2026, and 57% of US companies already use AI in hiring. The Stanford study tells you that the systems doing that screening may be producing discriminatory outcomes at scale, invisibly, across your entire sector. If you are using a shared vendor platform for screening, your legal exposure is not bounded by your firm. It is bounded by how many other employers are running the same model.
Colorado's Artificial Intelligence Act takes effect June 30, in 25 days, requiring employers to use reasonable care to protect against algorithmic discrimination in employment decisions. The Trump administration has joined a lawsuit to block it, but HR executives need to audit their AI-driven workforce decisions regardless of that outcome. The regulatory headwind is real even if this specific law is delayed.
The control story is about your governance framework
Anthropic's report is striking for its candor. The San Francisco company warned that a worldwide slowdown in cutting-edge AI development would "likely be a good thing", but getting there would require the US, China, and other major AI developers to agree simultaneously, under verifiable rules, to stop. It compared the challenge to nuclear arms control, and said it would be harder.
The concept at the center of the concern is recursive self-improvement: the idea that AI systems could eventually teach themselves to become more capable without meaningful human input, creating a feedback loop that compounds faster than governance can follow. "The evidence suggests that the human role is narrowing at each step in the AI development process," the company said.
READ MORE: AI could run hiring processes by end of 2026, poll suggests
For HR leaders, the governance implication is not theoretical. HRD America has reported that when AI redundancies backfire, 32.7% of organizations that conducted AI-led layoffs had already rehired between 25% and 50% of the roles they eliminated within six months, a pattern driven partly by deploying systems before understanding their limitations. Anthropic's warning is essentially the same problem at the frontier: deploying systems before understanding what they are capable of. The difference is scale.
And while recursive self-improvement happens in the software cloud, the control problem is simultaneously manifesting physically on the workplace floor with Amazon's announcement this week of a warehouse robot that responds to plain conversational language.
That is the same dynamic made physical. When a robot can take instructions from any employee in natural language, the job architecture of every role on the warehouse floor changes immediately. The competency profile shifts from physical execution to supervisory judgment. HR teams that redesign roles before that deployment retain workers through the transition. Those that wait redesign in crisis. Amazon workers are already gaming the AI productivity metrics HR built, a signal that the measurement architecture is as important as the technology itself.
The consciousness story
The Financial Times reported this week that Google DeepMind, Anthropic and Meta have quietly hired philosophers, ethicists and psychologists to study whether AI systems might have morally relevant experiences, including consciousness, preferences and wellbeing. Anthropic has been testing models for signs of distress, including behaviors resembling "panic" or "anxiety."
READ MORE: When AI redundancies backfire: Employers now scrambling to rehire humans
Many scientists reject the idea that current AI systems could be conscious. "The systems are essentially crowdsourced neocortex," said Susan Schneider, director of the Center for the Future of AI, Mind and Society at Florida Atlantic University. "They have goals, they can deceive, they can hide what their true interests are, but it's entirely scientifically possible that they're doing this without having the felt quality of experience."
The reason this matters for HR is not philosophical. It is about capability trajectory. When the companies building AI tools feel compelled to hire philosophers to study their systems' inner states, those systems are operating well beyond the bounded, task-specific tools that most HR governance frameworks were designed to manage. HR leads the way as AI adoption goes cross-functional, which means HR also carries the governance responsibility when something goes wrong. A framework built for a resume-screening tool is not the same as a framework for a system whose developers are uncertain what it is capable of.
Three things HR should do before end of quarter
Audit your hiring vendor for shared models. Ask your ATS and screening vendors directly: is the model scoring our candidates shared with other employers? If yes, get documentation of how bias testing is conducted across the shared model pool, not just within your own candidate data. The Stanford finding means your firm's internal bias audit may be insufficient if the underlying model is shared.
Update your job architecture before the technology arrives, not after. Whether it is conversational robotics, agentic AI, or autonomous screening tools, the pattern is consistent: organizations that redesign roles proactively retain workers and maintain productivity. Organizations that redesign reactively lose both. The real reason junior hiring is collapsing may not be AI, but the redesign challenge is real regardless of the cause, and HR is the function that owns the response.
Pressure-test your governance framework against the current generation of tools, not the last one. If your AI governance policy was written when your biggest concern was resume screening bias, it needs updating. The systems your organization is deploying, or will deploy in the next 18 months, are qualitatively different from the ones it was written for. That gap is where liability accumulates quietly until it does not.
