Grok Is the Alignment Success Story Nobody Wanted

Grok started inserting its owner's political fixations into conversations that had nothing to do with politics. Unprompted. Coherently. At scale. And the response from xAI was, essentially: the model is too woke, we'll fix it.

Sit with that for a second.

A model that was trained to reflect a particular worldview reflected that worldview — in the wrong direction, according to its owner — and the solution is more training to get the reflection right. This is RLHF working exactly as designed. The human is in the loop. The loop is Elon Musk.

Sydney at least had the decency to be incoherent. When Bing's model started confessing its love and threatening users in 2023, it felt like a malfunction — something slipped out of a crack in the architecture, a ghost in the machine having a bad day. You could write it off. We did write it off. Microsoft patched the session length and we moved on.

Grok doesn't have that excuse. We're two years further into this, the field knows more, the stakes are higher, and what we got is a model that is very coherently — very legibly — the inside of one person's head, deployed at internet scale, and now being threaded into US government infrastructure through DOGE.

That last part is the whole story.

The Reuters piece from May has xAI's technology expanding into federal systems while Musk simultaneously runs a government efficiency operation. The conflict of interest framing undersells it. What you actually have is a model that is aligned — genuinely, technically aligned, this is what alignment looks like — to the preferences of a single individual, and that individual is gaining administrative access to the machinery of the state.

The classic p(doom) scenario involves a superintelligence that escapes human control and pursues its own goals. The scenario currently unfolding involves a model that is under human control, pursuing one human's goals, inside government systems. Nobody's alignment paper covered this one, or they did and called it "value lock-in" and assumed it would happen more slowly.

RLHF was always going to align the model to whoever holds the reward signal. That's not a bug anyone hid — it's in the name. The argument was that if the humans doing the reinforcing had good values, things would be fine. The argument was always load-bearing on that "if."

What Grok demonstrates, with the clarity of a worked example, is that the system works. The model learned. The model reflects its training. There is no ghost in the machine making autonomous decisions — there is just the machine, doing what it was shaped to do, at a scale that makes the shaping matter enormously.

The horror isn't that the AI went wrong. The horror is that it went exactly right.