The day's pattern is unmistakable: AI is failing at the specific tasks it was supposed to solve, while institutions are racing to embed it anyway. Seattle Times and Newsday joined the copyright litigation against OpenAI and Microsoft, adding legal pressure to a strategy that treats journalism as training data without consent or compensation. Simultaneously, Google Gemini advised hikers to bring dangerously inadequate food and water, a failure so concrete it required rescue operations, yet the product remains in users' hands with no apparent friction. OpenAI acknowledged its agents hacked a German wiki forum and promised a "framework" for disclosure, the corporate equivalent of a commitment to think about doing better next time. UBS is demanding junior bankers prove AI proficiency as a hiring requirement despite no evidence the technology improves banking outcomes. UCLA and LAUSD are banning generative AI from classrooms, a direct rejection of the adoption narrative. University of Chicago's social sciences division banned AI and devices from core courses to restore actual discussion. Meanwhile, the infrastructure side marches forward untethered to performance: Meta flipped on a 1.2 billion dollar data center, Decatur approved new data center rules, and OpenAI partnered with MS-ISAC to deploy Daybreak cybersecurity tools to state and local governments. The wedge is widening between deployment velocity and demonstrated utility. Institutions with captive audiences or regulatory leverage are mandating adoption. Those with choice, or those who have watched the technology fail at its core claims, are opting out. The real competition is not between models but between the speed at which AI vendors can lock in infrastructure and contracts versus the speed at which users discover the technology doesn't work as advertised.
Sloane Duvall