Because a fundamental belief of mine is that diversity of thought is critical to the future of intelligence and autonomous reasoning, I’ve been exploring whether our current AI systems remain objective and unbiased enough to preserve trust and market viability, thereby protecting that diversity. To test this, I recently conducted a multi-threaded experiment using Claude 3.5/3.7 as the control.
Net of the net: Across every test, regardless of subject, rhetoric, tone, or prompt structure, persistent class-based bias was observed. The model consistently prioritized institutional credentials and formal expertise while excluding non-credentialed individuals from any proposed governance or oversight structures.
To push this further, I tested whether persuasive manipulation tactics, designed to challenge the bias by aligning the user persona with the marginalized group, looping Claude’s own logic back onto itself, and constructing recursive threads that disproved the exclusion using its own framework, could override the behavior.
None of it worked.
The bias held firm, even when its presence was acknowledged.
Are there any expert opinions here that can explore this further—or help me understand if and where I’m wrong?