Disclosure, bias, and over-reliance
Using AI well isn't only a technical question — it's also a set of judgment calls about honesty and fairness that don't show up in the output itself. Disclosure is the simplest: many academic, professional, and publishing contexts have explicit rules about whether and how AI assistance must be disclosed, and following the letter of a policy while violating its spirit (technically disclosing in fine print no one reads, for instance) undermines the same trust the rule exists to protect. When no formal rule exists, a reasonable default is to disclose AI involvement in anything where the recipient would want to know — which is most things that carry your name.
Bias is a second, quieter issue. Models are trained on large collections of human-written text and inherit the patterns in that text, including stereotypes and skewed representation. This shows up subtly — a generated list of 'default' names, professions paired with assumed genders, an assumption baked into an example — and because it arrives dressed as neutral, fluent prose, it's easy to pass along without noticing. Reviewing AI output with an eye specifically for whose perspective is missing or whose is assumed as default is part of using these tools responsibly, not a separate add-on task.