There is a cost to building AI systems that do what we want. Not a theoretical cost, not a future cost — a present, measurable, compounding cost that shows up in every benchmark comparison between safety-focused labs and their less scrupulous competitors. We call it the alignment tax. And we have been pretending, for years, that it is smaller than it is.
The alignment tax is the performance penalty you pay when you constrain a model to behave in ways that are safe, honest, and beneficial. Constitutional AI, RLHF, red-teaming, interpretability requirements — each of these interventions makes a model more trustworthy and less capable. Not always. Not on every task. But on the tasks that matter most, on the benchmarks that predict real-world performance, the gap is real and it is growing.
The Numbers Don't Lie
I have spent the past six months reviewing internal evaluation data from three frontier labs — data that will never be published, shared under conditions of strict anonymity. The pattern is consistent across all three. Systems trained with aggressive safety constraints score, on average, 8 to 14 percent lower on the benchmarks that researchers use to predict real-world capability. That gap has been stable for two years. It is not closing.
Eight to fourteen percent sounds modest. It is not. In a competitive landscape where the difference between the leading model and the second-best model is often measured in single-digit percentage points, an 8 percent handicap is the difference between winning and losing. It is the difference between your model being deployed in critical infrastructure and your competitor's model being deployed instead.
An 8 percent capability handicap is not a rounding error. In a winner-take-most market, it is an existential threat to every lab that takes safety seriously.
The Incentive Structure Is the Problem
The labs that are most committed to safety are not naive. They understand the alignment tax. They pay it anyway, because they believe — correctly, I think — that the alternative is worse. But belief is not a competitive strategy. And the market does not reward virtue.
What the market rewards is capability. Capability attracts customers, customers generate revenue, revenue funds compute, compute produces better models. The feedback loop is tight and fast. A lab that pays the alignment tax is running the same race with a heavier pack. It can win. But it has to be significantly better at everything else to compensate.
The deeper problem is that the alignment tax is not static. As models become more capable, the interventions required to keep them safe become more expensive. The tax rate rises with capability. This means that the labs most committed to safety face an increasingly steep climb precisely as the stakes get highest.
What We Should Do About It
The honest answer is that I do not know. The policy interventions that would help — mandatory safety standards, liability frameworks that penalize unsafe deployment, international coordination on capability thresholds — are all either politically impossible, technically unverifiable, or both.
What I do know is that pretending the alignment tax does not exist, or that it will be solved by better research, is not a strategy. It is a wish. The researchers working on alignment are brilliant and dedicated. They are also, by their own admission, years behind the capability researchers. The gap is not closing fast enough.
We are paying the alignment tax right now, in the form of a competitive landscape that systematically disadvantages the labs most committed to building AI that is safe. The bill will come due eventually. The question is whether we will have the institutional capacity to pay it when it does.