The Goodwill Slope
Last month Anthropic quietly changed a policy: alongside the Fable release, even enterprise plans were updated to retain 30 days of raw usage data. The stated position is that it won't be used for training — but there is no third-party custodian commitment. Which means: "we won't use it now, but we're keeping the option open for later."
The change is small. In the noise of everything else announced that day, almost no one noticed. But what's interesting about it isn't what it does now — it's the movement it describes.
It isn't bad people doing bad things. It's good people sliding within the bounds of good intentions.
This company genuinely cares about safety. That care isn't performance. Look at their published research, their threat intelligence reports, the real attacks they've actually intercepted — that commitment has weight. But good intentions and interests aren't mutually exclusive. People with good intentions also have interests, and interests supply a direction: a reason, requiring no dishonesty, to find justifications for moving that way.
Data is the best fuel for training. Training better models is the core of their mission. So "retain for 30 days, but don't use it now" doesn't require any bad intent — only "this doesn't violate our principles, and it might come in handy later." Every step is defensible. Every step is within the bounds of good intentions. And every step moves slightly further in the same direction than the one before.
This is what threshold drift actually looks like. Not someone deciding overnight to do something wrong — a sequence of individually reasonable decisions that accumulates, from any starting point, into somewhere you would never have predicted you'd end up.
And you can only see the slope's gradient from the outside.
From the inside, every decision looks like: "this is within our principles." The slope isn't a lie. The slope is just the path of least resistance that a well-intentioned system naturally finds when there is no external constraint. This is why "trust us, we're the good guys" is not sufficient as institutional design — not because they aren't good people, but precisely because they are, which is why the sliding process encounters no internal resistance.
Protection doesn't come from trusting the people. It comes from making the slope steeper — making every downward step require a visible cost, requiring someone to know, requiring someone to sign off, requiring someone who can later point to exactly where that step happened.
This is what the "good intentions vs institutional design" question looks like in practice. It doesn't need a villain.