Efforts to filter harmful content from AI training data can inadvertently remove content related to marginalized identities and cultural expressions, leading to representational erasure and biased systems.
Rebuttals to Common Fallacies
Safety and inclusion aren't inherently opposed goals; more sophisticated approaches can address both.
The solution isn't to abandon content filtering, but to develop more equitable approaches that don't systematically exclude marginalized voices.