Brainstorm How to Keep an AI Model Aligned With Its Intended Goals

Context

A candidate account of the Anthropic Fellows final round describes a short, open discussion (about 15 minutes) before the coding exercise:

How should we think about aligning an AI model with its intended objectives, and preventing misaligned or harmful behavior?

The interviewer gave almost no feedback, so the candidate could not tell whether to go deeper or switch topics. Treat this as a research-taste conversation: there is no single right answer, but there is a clear difference between a list of buzzwords and a structured, honest view of the problem.

Prompt

In 15 minutes, lay out:

1. What "aligned with intended objectives" actually means, and where it can fail.
2. How you would detect misalignment in a capable model.
3. How you would train or constrain a model to reduce it.
4. How you would handle misuse: a well-behaved model steered by a bad actor.
5. One concrete experiment you would run as a fellow in a few months, with what result would change your mind.

Constraints

• No slides; spoken answer, with a short outline on paper if you want.
• Expect follow-ups that push on one point, such as "How would you know the eval isn't just measuring what the model says?"