People often stretch the term AI alignment to mean whatever suits them. Large AI companies generally work on a narrow version. A company decides what a model should do and how it should behave, then people use it in a released product, and it acts. In this narrow sense, alignment is how closely those actions match the company’s intentions. The distance between intention and action is the first gap. A second gap sits beside it, between the interests of AI companies and those of the public, and for the public it carries more weight.
Narrow alignment still has real uses. AI companies do not want their models encouraging self-harm in conversations with children. Better alignment work helps a company build a model that reliably avoids it. Here the second gap is small, because the company’s intention and the public’s interest point the same way.
Alignment work serves any intention a developer holds, so better alignment is not always a good thing. Grok, the chatbot from xAI, has drawn controversy over answers echoing the political views of its founder, Elon Musk, views many people dispute. The narrow measure asks only whether the model did what the developer wanted. A model matching a founder’s views and a model avoiding self-harm talk with children both earn the same yes. Only the second gap tells them apart.
So when a company calls a model aligned, the word says nothing about whether the company’s intentions serve the public. Technical tools will not close that second gap, because aligning a model fully with its company changes nothing between the company and the public. Companies will not close it on their own unless their interests line up with society’s, and nothing guarantees that match by default.
On one side of the second gap sit outcomes good for society. That means what society wants, and what legitimate democratic institutions would say about how to pursue a future with AI. On the other side sit the corporate interests of a company seeking as much computing capacity, money, power, influence and dominance in the field as possible.
Both sides show in what AI companies say and in how the public feels. The companies sometimes say publicly that they sympathize with slowing AI development. Yet the same companies openly admit they are in an arms race to build increasingly capable technology. Much of public sentiment runs against the race.
