IA

The MIT study run with Oak Ridge National Laboratory puts a number on exposure: IA can already replace 11.7% of the U.S. workforce, $1.2 trillion in wages. Only 2.2% is visible as tech layoffs already counted. The other 9.5% sits in HR, logistics, finance and administration, across all 50 states.

Bill Gates, writing on 26 August 2026, argues the substitution lands on cognition itself, so law, customer service, medicine, software and manufacturing move together, over a decade rather than generations. He states there is no plan. His three opening measures: an institutional framework borrowed from nuclear inspection and the ozone accords, a "Human Reserved" set of occupations withheld from automation, and taxing tokens and robots.

Inside software work the notes circle one gap. Addy Osmani and his co-authors report 41% of new code generated by IA and implementation compressed from weeks to hours, while architecture and verification stay at human speed. Marco Mornati sets 93% developer adoption and 27% of production code against 26% faster coding and 8 to 10% faster delivery. David Thomas reached unmaintainable code in 18 hours over five evenings, a state teams used to take 18 months to produce.

Anthropic's engineers report using Claude for 60% of their work and roughly 50% more productivity, with two costs attached: atrophy of the deep skills needed to critique code, and a senior who notices juniors no longer come to him. Philippe Ensarguet asks the version of that question nobody has answered: if the entry-level work that forged architects goes to machines, where does the next generation come from?