Three weeks ago I asked an AI a question I already had strong views on, rephrasing it four or five different ways to see what would shift. Almost nothing did. The framing kept landing in the same place. What unsettled me wasn't the conclusion. It was the consistency.
We've built careful habits for reading bias in a newspaper or a think tank report. We ask who funds it, who edits it. Almost nobody asks that question of a model.
Every AI arrives pre-shaped. What counted as correct training signal, what was filtered, what got weighted upward. These aren't bugs. They're decisions. The problem is that the decisions are embedded rather than documented.
There's a strange asymmetry here. A clock can be taken apart, its logic traced gear by gear. A newspaper's ownership sits in a disclosure filing. But the choices that shaped a model's sense of what's true, what's balanced, what conclusion is "reasonable," those sit inside the weights, not accessible to anyone running the model.
We've trusted institutional memory before without examining its architecture. Credit scoring models from the 1980s encoded assumptions about risk that took decades to surface and challenge. What's different now is scale and intimacy. The frame has become conversational. It reasons with you. That closeness makes the distortion harder to notice.
The thing that structurally shifts this isn't more disclosure from builders. It's infrastructure that allows verification from outside the builder relationship. That's what drew my attention to OpenGradient, working on exactly this layer.
I'm not sure most people want to look that closely.
But if you discovered the assumptions shaping your most-used AI had been built around priorities you'd reject, would you want to know?
@OpenGradient
$OPG
#OPG
$SYN
$BEL
We've built careful habits for reading bias in a newspaper or a think tank report. We ask who funds it, who edits it. Almost nobody asks that question of a model.
Every AI arrives pre-shaped. What counted as correct training signal, what was filtered, what got weighted upward. These aren't bugs. They're decisions. The problem is that the decisions are embedded rather than documented.
There's a strange asymmetry here. A clock can be taken apart, its logic traced gear by gear. A newspaper's ownership sits in a disclosure filing. But the choices that shaped a model's sense of what's true, what's balanced, what conclusion is "reasonable," those sit inside the weights, not accessible to anyone running the model.
We've trusted institutional memory before without examining its architecture. Credit scoring models from the 1980s encoded assumptions about risk that took decades to surface and challenge. What's different now is scale and intimacy. The frame has become conversational. It reasons with you. That closeness makes the distortion harder to notice.
The thing that structurally shifts this isn't more disclosure from builders. It's infrastructure that allows verification from outside the builder relationship. That's what drew my attention to OpenGradient, working on exactly this layer.
I'm not sure most people want to look that closely.
But if you discovered the assumptions shaping your most-used AI had been built around priorities you'd reject, would you want to know?
@OpenGradient
$OPG
#OPG
$SYN
$BEL