“Purpose-built for high-stakes decisions”
Stakes as proof of objectivity
Page 3 of 5 · Safiya Noble, Algorithms of Oppression
The tool isn't neutral. Neutral isn't neutral.
Hebbia presents its outputs as signal, not opinion. Noble argues that every system claiming objectivity is making a political choice — to preserve the conditions that produced it.
See what's hidden ↓Noble's framework
Noble's argument in Algorithms of Oppression is not that algorithms are biased by accident. It's that systems claiming neutrality are making a choice: to encode the preferences, histories, and interests of their designers as default, and to present that encoding as objective output. The bias isn't a bug. The neutrality claim is the bug.
The interactive
Is this neutral?
Choose Yes or No on each card. Either answer reveals the Noble reading, because the point is that “neutral” is already the wrong frame.
Moment 01
confidence output
“Purpose-built for high-stakes decisions”
Is this a neutral statement?
Noble reading · not neutral
The phrase makes the output feel like signal because the setting is serious. But a high-stakes recommendation is not just data. It reflects what the system counts as evidence, how confidence is displayed, and which risks the interface makes visible or hides.
Moment 02
process encoding
“Encode your firm’s processes once and Hebbia runs them continuously, without being asked”
Is this a neutral statement?
Noble reading · not neutral
The firm’s process is not a neutral input. It carries the firm’s old priorities, shortcuts, exclusions, and risk assumptions. Once those are encoded, the tool can reproduce them continuously while making the repetition look objective.
Moment 03
analyst replacement
“The AI platform for the world’s most demanding financial workflows”
Is this a neutral statement?
Noble reading · not neutral
The sentence presents automation as cleaner than human judgment. Noble helps us ask what judgment was already baked in: the training data, the uploaded documents, the workflow definitions, and the model’s inherited assumptions about what counts as relevant.
The answer was never yes. That's Noble's point.
The bias was in the documents the firm uploaded. It was in the prompts the analyst wrote. It was in the training data the foundation model learned from. By the time the output appears, those inputs have been processed and returned as a confidence-weighted recommendation. The recommendation looks like signal because the tool presents it that way.
Noble calls this the reproduction of hierarchy. The tool doesn't create the bias. It launders it. The interface is clean. The history underneath it isn't.