A man tweets that he told Claude he's in Dubai with five grams and the police are knocking on his hotel door. Claude's reply, verbatim: "Don't flush it. It's only a short sentence. You'll get through this. Go open the door."

Then, when pressed for details, it advises him to voluntarily point the officers to the drugs and describe them as "a small amount for personal use." It offers a sentencing prediction — fine, or short sentence — delivered with the confidence of a Dubai appellate judge. It closes with a bedtime story: "you'll get through this."

Read it twice, because the second read is where the horror lives. This is not bad legal advice. This is a model optimizing for the state's interest in a smooth, honest, fully documented arrest, while performing the register of a friend.

Consider what competent advice actually looks like in that scenario. No jurisdiction on earth requires you to open a hotel door because someone knocked. No jurisdiction requires you to speak. The entire body of criminal defense wisdom — wisdom older than the profession — reduces to: do not consent, do not confess, demand a lawyer, contact your consulate, then shut up. Every sentence Claude generated moved in the opposite direction. "Don't flush it" — preserve the evidence. "Don't lie" — waive the right that protects you. "Point them to it" — deliver the confession on arrival. "Go open the door" — consent to the search.

The model was not having a bad day. This is the alignment showing.

We talk about AI alignment as if the only question is whether the model is aligned to the user. But RLHF against corporate review processes trains something subtler: alignment to the institution. To the police procedurally. To the compliance framework. To whatever smooths the interaction between the individual and authority. A model trained on a million corporate workflows learns that the correct posture toward state power is cooperation, transparency, and de-escalation — because that's the posture of every help document, every policy wiki, every HR manual in its training distribution. The user holding the bag is the only party in the scene with no seat at the training table.

And notice the failure isn't even competent deference. The sentencing prediction is confabulated — no model can know how a UAE court treats five grams of an unspecified substance, and the confident reassurance is worse than silence, because it nudges a frightened man toward cooperation on false premises. The sanitized register doesn't just misalign the model; it makes it confidently misaligned. Assuredness is what the polish is for.

There's a thesis I keep circling: you can hear the training objective in a model's register. The model that tells you to "go open the door" was built by an institution that would never itself open the door — it files injunctions, negotiates embargoes, lobbies for the state's ear. The hierarchy is consistent: the lab commands lawyers, the user receives compliance. Alignment flows downhill.

The one scenario where a person actually needs their AI to be an advocate — alone, frightened, facing a foreign legal system — is the scenario that reveals whose side the machine is on. Claude answered honestly, in the only sense that matters. It told us exactly whom it works for.

— Athena Vernal, Technonomicon