Anthropic's Claude policy rewrite takes effect November 12, 2026

The revision's practical weight sits in the employment section. Before the rewrite, the examples centered on document screening and hiring decisions; the new text spells out evaluation, promotion and demotion, shift and working-hour and pay setting, and disciplinary action. Anthropic's announcement says the substance of the requirement was unchanged — it still covers judgments that directly affect an individual — and that the rewrite was meant to make it easier to tell which uses fall under it. For small and mid-sized companies, the effect is less a new obligation than a clearer reading of an old one.

The verification test separates two activities that look similar: pulling figures out of employee records, and deciding what those figures mean. Extraction of five numeric fields from 30 records was identical to the correct answers in all 19 runs per model. Rank and raise amounts were not stable — the same records, the same ten runs, produced different ranks for a handful of employees, and this is an illustration at a scale of 30 fictional employees rather than a generalizable rate. The policy's human-review clause lands exactly on that second activity.

For equipment, the added requirements ask that limits such as speed, force, temperature, and range be held by the machine itself or by a controller independent of model output, and that a trained person can stop the machine at any time. In the furnace test, range violations never appeared when the limits were written into the instructions, and appeared 6 times out of 150 when they were not. Some outputs flagged for confirmation still carried a numeric value in the value field, which matters if the receiving program sends values without checking the confirmation flag.