Three findings you should not have to discover alone.
Exact numbers, sources visible. Open each one. Every claim on this page is verified against the cited study.
The models did this from dialect alone. No race labels appeared anywhere in the input. The prejudice is covert: it operates on how someone writes, not on any stated identity.
Hofmann, Kalluri, Jurafsky, & King (2024), Nature, 633, 147-154.
89 of 91 essays were flagged by at least one detector. Native-speaker essays passed at nearly 100 percent. The tools built to catch AI writing misfire on real human writing, and they misfire along language lines.
Liang, Yuksekgonul, Mao, Wu, & Zou (2023), Patterns, 4(7).
The stereotyping was specific: critiques for Latino students, for example, were framed around family and culture. The argument itself went unengaged. This is the study behind the two feedbacks on the opening page.
Sparks (2026, June), Education Week.
What AI literacy actually is
Three modes of engagement: understand, evaluate, and use. Grounded in two core values: human judgment and centering justice (Mills et al., 2024, Digital Promise).
Our program permits these tools everywhere and requires this literacy nowhere. So every one of us is self-taught, and self-teaching is not evenly distributed.