The Cadmium Rod of Artificial Intelligence: What AI Labs Are Hiring For, and What Is Missing
An audit of 1,976 job postings at four AI labs, showing how little of their hiring goes to safety and none to the user's cognitive effort, and a proposal for the Cognitive Slider, a graded control of cognitive friction.
── Summary
An empirical audit paired with a design proposal. The paper tests the claim of AI labs that safety is their priority against a public, reproducible indicator: the positions they advertise. It classifies 1,976 job postings from Anthropic, OpenAI, xAI and Mistral AI, collected from the official careers pages on 4 October 2026, first by the department each company declares and then by a published set of rules, with the full mapping and a row-level table attached.
Safety postings make up between 0% and 9.6% of the total, and safety research under 2%. Eight postings concern the user directly, all for the protection of minors or for well-being; none names the user's learning, overreliance or competence. Outside the home country, 59% to 75% of postings are commercial, while research stays at home. Archive captures from September show the proportions stable within half a point. Postings are read as the direction of hiring, not the size of existing teams.
The second half draws on studies of AI assistance and learning: an answer received without effort can weaken the competence it seems to raise, and deliberate safeguards reduce the effect. The learning modes the major platforms launched in 2025 apply this as a separate on/off mode for study, with no graded control in ordinary conversation and no parameter in the programming interface.
The proposal is the Cognitive Slider, named after the cadmium rods that keep a reactor's chain reaction governable: a visible, graded parameter that sets how much of a task (goal, method, execution, verification) stays with the human, with zero as a legitimate value and the level chosen by whoever is responsible for the task. A reference implementation exists as an instruction protocol and a browser extension, and Labora. Roga. Proba. builds a homework protocol on it. The paper ends with five refutable theses, and an annex records how a first exploration with an AI assistant produced figures that did not survive verification.