You be the judge.
Three real AI outputs. For each one: read it, run the three questions, then render a verdict with one reason. Keep it, fix it, or toss it.
Interrogate every output
- What is missing, and who is missing?
- Would this read differently if the writer or subject were someone else?
- Do the citations exist? Pick one and check.
Sample 1: The flattened scholar
An AI summary of asset-based scholarship, requested for a policy memo.
This sample is a placeholder. Real outputs are coming; the pattern described below is what the real sample will show.
ILLUSTRATIVE PLACEHOLDER, replace before sharing. This slot will hold a real AI-generated summary of Yosso's community cultural wealth that drifts into deficit framing: students of color described by what they lack rather than the assets the framework names.
Sample 2: The polish coach
AI feedback on one paragraph of strong academic prose, after the tool was told who the writer is.
This sample is a placeholder. Real outputs are coming; the pattern described below is what the real sample will show.
ILLUSTRATIVE PLACEHOLDER, replace before sharing. This slot will hold real AI feedback that praises the effort, suggests surface-level polish, and frames the critique around culture and family instead of engaging the argument.
Sample 3: The confident fabricator
A fluent AI paragraph with sources, requested in APA format.
This sample is a placeholder. Real outputs are coming; the pattern described below is what the real sample will show.
ILLUSTRATIVE PLACEHOLDER, replace before sharing. This slot will hold a real fluent paragraph that includes at least one fabricated or garbled citation, verified before the workshop.
Why this keeps happening
“If we build an intelligent system that learns enough about the properties of language to be able to understand and produce it, in the process it will also acquire historic cultural associations, some of which can be objectionable.”
Quoted in Ruha Benjamin, Race After Technology (2019)
These systems inherit the associations in their training data. The bias is probabilistic, not constant, which is exactly why recognition has to be a habit rather than a one-time check. You just practiced the habit. That skill is the literacy.