In the world of electronic components, which is where you came into my orbit at a conference, transistors are biased with either a positive or negative electrical charge to establish a stable operating point. I read this piece with the understanding that bias is both positive and negative, but with a strong bias to believe that those whom stand to benefit the most from AI are biased to build AI tools biased to maintain the status quo. A source I am biased to trust on this topic is social scientist Christian Ortiz, Forbes article here: https://www.forbes.com/sites/janicegassam/2025/09/26/big-tech-ignored-bias-in-ai---justice-ai-gpt-says-it-solved-it/
I’m not so sure model manufacturers are uniquely misaligned to our collective interests, and I actually disagree that they’ve ignored their product’s biases; BUT I do think that explainability standards (enforced by governments and industry watchdogs) are critical.
In the world of electronic components, which is where you came into my orbit at a conference, transistors are biased with either a positive or negative electrical charge to establish a stable operating point. I read this piece with the understanding that bias is both positive and negative, but with a strong bias to believe that those whom stand to benefit the most from AI are biased to build AI tools biased to maintain the status quo. A source I am biased to trust on this topic is social scientist Christian Ortiz, Forbes article here: https://www.forbes.com/sites/janicegassam/2025/09/26/big-tech-ignored-bias-in-ai---justice-ai-gpt-says-it-solved-it/
I’m not so sure model manufacturers are uniquely misaligned to our collective interests, and I actually disagree that they’ve ignored their product’s biases; BUT I do think that explainability standards (enforced by governments and industry watchdogs) are critical.
Zack, this connects directly to our earlier conversation on the rules of AI.
You nailed the first layer: make bias visible, and it becomes manageable.
I've been building the second layer: semantic guardrails that prevent reinterpretation once visibility happens.
It's called ClarityOS. And your explainability thesis is why it works.
This essay should be required reading for everyone building it.
Would be worth a deeper conversation, curious if you see the same dynamics in alignment work.