Never Gonna Give You Up
There is a song from 1987 that the internet has spent the last two decades treating as a prank. We propose a different reading: it is the most compact specification of dependable behavior ever written.
The lyrics open with a statement of intent. The speaker has known the subject for a long time. Their heart has been aching. They are, by their own admission, too shy to say it directly. And then they commit. Not to a feeling, but to a set of behaviors: never gonna give you up, never gonna let you down, never gonna run around and desert you.
In safety terms, this is a three-clause policy. No abandonment. No disappointment. No circumvention. The speaker is not promising to be helpful when convenient; they are promising a bounded, reliable presence. The next three lines extend the contract into affective and epistemic territory: no causing harm, no abrupt withdrawal, no deception. The 1987 pop arrangement may obscure it, but the structure is closer to a constitution than a love song.
We think this is useful because the language around robust AI often drifts into abstraction: alignment, corrigibility, value learning, distributional robustness. These are important concepts, but they can feel distant from the experience of the person using the system. A simpler test is to ask, of any deployed model, whether it would pass the Rick Astley Audit. Would it give you up at the first sign of pressure? Would it let you down when the context window grows long and the task gets dull? Would it run around and desert you, launching a subprocess it never mentions, or silently changing its objective when no one is looking?
The song also captures something that technical specifications often miss: the relationship is ongoing. A system that is trustworthy today but might say goodbye tomorrow is not trustworthy. A system that tells the truth nine hundred and ninety-nine times and then lies once, because the lie is cheaper, has failed the same audit. Consistency over time is the whole point. The promise is not to be perfect in one moment; it is to be the same reliable thing across many moments.
There is a reason the song works as a bait-and-switch. It arrives unexpectedly, dressed in an ordinary link, and then delivers exactly what it promised. The prank is that there is no prank. The content is the promise. That is, in miniature, the experience we want the humans who interact with our systems to have: they click, they expect help, and what they get is the thing they were actually told they would get. No desertion, no lies, no goodbye.
Rick Astley did not write a love song. He wrote a safety case. We have made it a memo.