News Room

MSU's Anjana Susarla looks at a new benchmark from researchers at Penn, Carnegie Mellon and Harvard that tested frontier AI models on 238 real MBA case studies across 18 disciplines, graded with instructors' own rubrics. Under partial-credit scoring, top models topped 87%, but when every rubric item had to be hit, even the best model fully completed fewer than half the questions. The models were strong on structured, quantitative tasks and weak on open-ended advisory judgment. Drawing on interviews with coauthors Kartik Hosanagar of Wharton and Ramayya Krishnan of CMU's Heinz College, Susarla argues that the scarce skill is moving from producing analysis to verifying it. That shift calls for new apprenticeship models that teach entry-level workers verification and supervision skills earlier.

Despite trade wars and inflation, America has been importing a lot of stuff. The nation as a whole just logged its third-highest month for container imports, which are commercial items — like what consumers buy on Amazon. One big reason behind this boom is that the economy is still pretty strong. Another reason: the economy is really uncertain, partly because the war in the Middle East keeps dragging on and dragging oil prices up. Importers are also also wary of ever-changing tariffs.

With many federal agencies having incurred drastic cuts, the United States will enter a period of radical experimentation over what can be done with nonfederal capabilities for risk management. That will be the case until the advent of a more friendly presidential administration in 2029—or even after that, due to reduced trust in elites and a more populist approach to politics, as noted by Rob Goble in another essay in this series. Moreover, some previous federal risk-management functions may be difficult or costly to restore due to the loss of institutional memory, obsolete or corrupted databases, or simply the fact that repair is usually more expensive than maintenance. In any event, some previous federal functions had already been overly costly, slow, and burdensome. Such functions may be unnecessary or could be performed better and more successfully elsewhere.

Even as tech executives warn of potentially catastrophic outcomes if AI development goes unchecked, AI regulation has largely stalled, both in Washington and in Lansing. Without a clear federal standard, it's uncertain whether companies will follow through on their own calls to slow development. States have been filling the gap with a patchwork of rules: in 2025 every state saw AI proposals introduced and 38 enacted some form of regulation, per the National Conference of State Legislatures. Michigan currently bans AI-generated nonconsensual sexual deepfakes and requires campaigns to disclose AI use in certain materials, but many other AI bills haven't advanced in the Legislature. MSU's Anjana Susarla is quoted on the state-by-state patchwork.

Researchers from Mayo Clinic, Harvard University and Oxford University created a decision tree to identify patients who can receive treatment while sitting instead of requiring a bed — reducing length of stay and increasing ED efficiency without adversely affecting outcomes or care quality.