MIT Technology Review’s weekday newsletter The Download put two data problems at the top of its latest edition: AI systems that may judge job applicants unfairly, and weather feeds that may become targets for manipulation as money and machine learning pile onto forecasts.
On hiring, MIT Technology Review reported that large language models are already known to absorb human prejudice from training material. New research, the outlet said, suggests the problem does not stop there. The models can also develop biases through experience and may stereotype job candidates more readily than people do.
That matters because résumé screening is one of the places employers are most tempted to insert automation before a human recruiter has read anything. A model that turns past interactions into durable assumptions can turn memory into a liability. MIT Technology Review said AI companies building more agent-like models, including systems designed to remember small details about users, may be giving those models more material from which to form stereotypes.
The mechanism is not magic, which is the part vendors tend to skip. If a model has learned patterns from biased data, then keeps accumulating signals from later use, it can start treating those correlations as useful shortcuts. In hiring, shortcuts about people are usually where the compliance headaches and human damage begin.
The newsletter’s second lead item focused on weather data. Monique Kuglitsch, Jesper Dramsch, Franz G. Kuglitsch, and Andrea Toreti wrote that forecasts are used every day by airline dispatchers, power grid operators, and farmers. Prediction markets have added another user group: people betting on real-world outcomes, including the weather.
According to the authors, that creates a new incentive to tamper with weather data for financial advantage. At the same time, forecasters are increasingly using data-driven AI systems, which makes the quality of input data even more consequential. If the observations feeding those systems are corrupted, the forecast can be corrupted too, and the effects can spread beyond a single bet.
The authors said experts can foresee cases in which such manipulation grows into broader systemic risk. The newsletter did not describe a specific sabotage incident in the summary, so the warning is about an emerging vulnerability rather than a confirmed attack campaign.
Other items in the roundup
- The Wall Street Journal reported that SpaceX is in talks to sell AI computing capacity to the Pentagon, with Reuters noting the company’s expanding relationship with the Defense Department.
- The Financial Times reported that Trump Media has pitched trading firms and banks on a premium feed offering early access to Donald Trump’s posts for $100,000 a month.
- NPR reported from court filings that ICE shared Medicaid data with Palantir before the data was deleted.
- Reuters reported that Apple briefly passed Nvidia as the world’s most valuable company, while CNBC linked Nvidia’s stall to shifting AI bets.
- The New York Times reported that politicians are seeking help changing what chatbots say about them.
- The Washington Post reported that the Pentagon is accelerating work on AI weapons, including armed robots.
The edition also quoted Rayan Krishnan, CEO of Vals AI, telling the New York Times that Chinese and American AI development has produced an odd contrast: China’s models appear more egalitarian in his view, while US companies are behaving more authoritatively.
This story draws on original reporting from MIT Technology Review.
