AI may screen your résumé before any human sees it. And according to new research summarized by MIT Technology Review, AI can form biases not only by absorbing humanity's history from training data, but also by building its own biases from experience.

That is the uncomfortable twist: we already know LLMs can pick up human biases embedded in their training data. But this new work suggests something worse for anyone betting on automation in hiring. The models can develop their own biases from interactions and outcomes, and they may stereotype job applicants more than humans do. In other words, even if you start with “clean” training materials, the system can still learn skewed patterns once it is exposed to hiring-like inputs and feedback.

So what’s driving the worry right now? Incentives and product direction. AI companies are racing to build agentic models, systems that can remember and act on more details about users over time. The more memory and personalization you add, the more opportunities there are for a model to attach probability mass to the “most likely” kind of candidate. If the agent can track tiny details, remember them, and use them to make decisions, it also has more chances to convert stereotypes into rankings, shortlists, and denials.

For executives, the practical risk is governance that only covers the past. Many compliance playbooks focus on training data bias, evaluation before launch, and documentation. But if bias can be shaped by experience, then the relevant question becomes: what happens after the system goes live? In hiring, “experience” might mean repeated exposure to résumés that correlate with past hiring outcomes, or feedback loops that reflect who gets interviewed and who does not. Even if no one intentionally programs discrimination, the system can still learn a biased decision surface from the environment it is used in.

This matters because hiring is one of the highest visibility and highest consequence uses of AI. A screening model can silently decide whose application gets traction, which changes the pool of candidates that humans even see. That pushes the problem upstream: it is not just about fairness in the final decision, it is about whether the system constrains the inputs that humans rely on. If an AI model stereotypes job applicants more than humans do, then you are not just replacing a step in the process, you may be amplifying a distortion before human review ever begins.

Meanwhile, outside HR departments, MIT Technology Review also flagged a different but equally systems-level risk: the temptation to sabotage weather data in prediction markets. Airline dispatchers, grid operators, and farmers around the world make decisions based on weather forecasts every morning. More recently, forecasts have become relevant for prediction markets where people bet money on real-world events, including the weather.

The concern is straightforward incentives plus increasing automation. The temptation to manipulate weather data to gain an edge in those markets, paired with a collective shift toward data-driven AI weather forecasting, is starting to put the accuracy of weather predictions at risk. The article warns that experts can foresee scenarios where the risks snowball into far more systemic problems, precisely because many industries rely on the same kinds of forecasts for planning and safety. When weather becomes a money-adjacent input to AI systems, data integrity stops being a side issue and becomes a core resilience issue.

Notice the parallel with hiring. In both cases, the environment and incentives matter. In hiring, experience can produce bias even beyond training data. In weather, manipulation can degrade forecasts in the real world, and then those degraded forecasts feed into decisions across sectors. That means boards and leadership teams need to think in loops: not just “Is this model biased?” but “How could the model become biased in our operating context?” Not just “Is our input data accurate?” but “What would adversaries, competitors, or market actors do to distort it, and how would the system respond?”

And if you are a peer leader, the strategic stake is simple: the market for AI-driven decision tools is accelerating, but trust is not keeping pace. If AI screening is rolled out without monitoring for post-deployment bias formation, it can lock in unfairness at scale and turn an evaluation exercise into a reputational incident. If weather forecasting systems or market-linked data pipelines are not hardened against tampering, inaccuracies can compound across industries. The companies that get this right will treat fairness and integrity like living systems, not one-time checks.