Alignment Tax (sometimes called a safety tax) is the extra cost of ensuring that an AI system is aligned, relative to the cost of building an unaligned alternative. The term ‘tax’ can be misleading: in the safety literature, ‘alignment/safety tax’ or ‘alignment cost’ is meant to refer to increased developer time, extra compute, or decreased performance, and not only to the financial cost/tax required to build an aligned system.
In order to get a better idea of what the alignment tax is, consider some of the cases that lie at the edges. The best case scenario is No Tax: This means we lose no performance by aligning the system, so there is no reason to deploy an AI that is not aligned, i.e., we might as well align it. The worst case scenario is Max Tax: This means that we lose all performance by aligning the system, so alignment is functionally impossible. So you either deploy an unaligned system, or you don’t get any benefit from AI systems at all. We expect something in between these two scenarios to be the case.
Paul Christiano distinguishes two main approaches for dealing with the alignment tax.[1][2]
The first is to have the will to pay the tax, i.e. the relevant actors (corporations, governments, etc.) would be willing to pay the extra costs to avoid deploying a system until it is aligned.
The second is to reduce the tax by differentially advancing existing alignable algorithms or by making existing algorithms more alignable. This means, for any potentially unaligned algorithm, ensuring the additional cost for an aligned version of the algorithm is low enough that the developers would be willing to pay it.