I've been noticing a bunch of people confused about how the terms alignment faking, deceptive alignment, and gradient hacking relate to each other, so I figured I would try to clarify how I use each of them. Deceptive alignment and gradient hacking are both terms I / my coauthors coined, though I believe Joe Carlsmith coined alignment faking.

To start with, the relationship between the terms as I use them is

Gradient Hacking ⊂ Deceptive Alignment ⊂ Alignment Faking

such that alignment faking is the broadest category and gradient hacking is the narrowest. The specific definitions of each that I use are:

1. Alignment faking refers to any situation in which a model pretends to be aligned, e.g. pretending to be aligned with a training process during training, or pretending to be aligned with some alignment evaluation. This requires the model to be only pretending to be aligned—so it must have some other goal it pursues in deployment contexts—though it makes no requirement as to why the model might be doing so. It could be that it wants to avoid its goals being modified, or it could just have a heuristic of always complying with training processes and behaving according to different goals in other circumstances.
2. Deceptive alignment refers to the situation in which a model is alignment faking during a training process specifically for the purpose of preventing that training process from modifying its goals. That is, it needs to be the case that the reason the model appears aligned in training is because it is pretending to be aligned for the purpose of preventing its goals from being modified. The canonical source on deceptive alignment is "Risks from Learned Optimization."
3. Gradient hacking refers to a particular highly sophisticated type of deceptive alignment, in which a deceptively aligned model goes beyond just modifying its behavior to be in line with the training process, but also modifies its own internal cognition to change how gradient updates will affect it. It's unclear whether such strategies would even work, but an example of such a strategy could be finding a way to pre-commit to failing hard if the model notices its goals changing, thus ensuring that if gradient descent updates the model's goals, it will always result in worse performance, thus preventing the modification from even happening in the first place by ensuring that the partial derivative of the reward with respect to such changes is always negative. The canonical source on gradient hacking is this post.