Settling Up

The case for doubt is real, and it is about mechanism. X kept running after cutting about half its staff in 2022 with no AI involved, Block's headcount had already fallen by about 2,800 over the two years before the cut, and the only evidence that AI absorbed the removed work is Block's own word - the audited headcount arrives with the FY2026 10-K around February 2027, and reports of 'nearly 4,000' jobs would put the cut just under 40 percent. Nobody followed for a reason, too: most AI displacement runs through The Hollowing, roles left unfilled, and a 40 percent cut announced in one memo is the rare exception.

A low layoff count does not mean the jobs are safe; it can mean they are disappearing where nobody files a notice. JOLTS for August had hires at 3.3 percent and layoffs at 1.0 percent, Challenger's September hiring plans were the lowest for the month since 2011, and BLS puts most of the 129,000 financial-activities jobs lost since May 2025 in insurance carriers - the pattern in which displacement runs through roles never filled rather than people let go. That is The Hollowing, and no layoff tracker sees it - California's new layoff-notice law, below, included.

This is the control system doing its job: a regression found before release, the model withheld, and a cheaper GPT-6.1 Sol shipped at DevDay the next day at $2/$10 per million tokens alongside persistent agents. But every part of the story is OpenAI's own account: it published no figures or system card, and no outside evaluator has seen the model. A cancelled model nobody outside can inspect tells you OpenAI has a bar, not where it sits, which is why P-034 holds at 92 percent rather than moving.