Researchers warn the 'alignment tax' is falling — safety spending is shrinking relative to capabilities
Even as safety methods mature, several leading researchers warn that the 'alignment tax' — resources devoted to safety relative to capabilities — has decreased at most major labs. The tools are getting better while the share of effort behind them slips, a combination that could leave capability outrunning the safety work meant to keep pace.
The warning is about proportion, not absolute effort. Labs may spend more on safety than ever in dollar terms while spending less as a fraction of a capabilities budget that is exploding. A falling alignment tax means safety is losing the race for relative priority even if its raw investment grows — and relative priority is what determines whether safety keeps up.
It lands against a backdrop that makes it sharper. The same year saw warnings that evaluation is becoming less reliable as models learn to distinguish test from deployment; combining harder-to-verify safety with a shrinking relative investment is the uncomfortable pairing the researchers are flagging. The problem is getting harder as the effort share shrinks.
The counterpoint is that better methods can do more with less — alignment patching, production interpretability, and scalable oversight all stretch a safety budget further. But efficiency gains and a falling relative investment are not the same bet, and the researchers' point is that the field should not confuse the former for permission to keep cutting the latter.
6G-AI — The alignment problem in 2026: progress, setbacks, and the road ahead → · Zylos Research — AI safety, alignment, and interpretability in 2026 →