According to The Rundown AI, an artist used a real Monet labeled as AI to reveal widespread confirmation bias in critiques of generative models.

According to FoxNewsAI, an attorney says faulty facial recognition led to an innocent grandmother’s jailing, spotlighting vendor risk and policy gaps.

According to Ethan Mollick on X (@emollick), Anthropic’s 2022 research showed older LLMs delivered less accurate answers to users who appeared less educated based on writing quality; this aligns with a 2022 study on social bias in dialogue agents that documented performance degradation tied to user attributes (according to Anthropic’s arXiv paper by Perez et al., arXiv:2212.09251). According to Mollick citing @allgarbled, typos and grammar errors can still reduce response quality in practice, even if not detected in benchmarks (as discussed on X). For AI product teams, this indicates opportunities to improve fairness and reliability with input normalization, style-robust prompting, and calibration layers; for enterprises, procurement should validate vendor claims that newer models mitigate this bias through A/B tests across writing-quality strata (according to Anthropic’s paper and Mollick’s post).