The internet, a space for connection and advocacy, can unfortunately become a breeding ground for hate speech, particularly targeting marginalized communities like people with disabilities (PwD). While platforms employ AI toxicity classifiers and large language models (LLMs) to combat online negativity, their efficacy in detecting ableism remains a critical question. Recent research explored this very issue, delving into how these AI tools respond to ableist language. The results revealed a stark contrast between how AI and humans, specifically PwD, perceive online toxicity. Traditional toxicity classifiers consistently underestimated the level of harm in ableist comments compared to PwD. LLMs, while closer in their assessments, still fell short of fully capturing the nuanced ways ableism manifests online. Interestingly, LLMs were more aligned with PwD than non-disabled individuals in identifying ableism, hinting at a potential for progress. However, a deeper dive into the explanations provided by both LLMs and PwD uncovered significant differences. While LLMs could identify stereotypical or overtly discriminatory language, they often missed the emotional impact and contextual subtleties that PwD emphasized. PwD responses reflected the personal and often painful experiences associated with ableist microaggressions, while LLMs tended toward more generalized, theoretical explanations. This discrepancy highlights a crucial gap in current AI moderation. Moving forward, effective online moderation needs to move beyond simple detection. Future development of inclusive AI tools requires centering the lived experiences of PwD, incorporating their nuanced understanding of ableism into training data, and prioritizing ableism interpretation alongside detection. Only then can online spaces become truly safe and inclusive for everyone.