Calibrate AI trust by subject, not by tool

Students tend to hold one global trust setting for AI: it's magic, or it's garbage. Neither survives contact with reality, because reliability is domain-shaped: strong on summarising a well-documented process, shakier on arithmetic done in prose, prone to invention on niche sources and local detail. The OECD and European Commission's AI literacy framework treats this kind of applied, subject-level judgement as a core competence rather than a generic disposition (OECD & European Commission, 2026), and Manning's disciplinary-ai-literacy skills take the same line: AI literacy is taught inside a subject, not beside it.

In practice

• Build a trust map for your subject with the class: three columns, "usually solid", "check it", "don't trust it", filled with your subject's actual task types.
• Populate it with evidence, not opinion: each claim about a column gets tested live (see "Teach them to catch the model being wrong").
• Revisit each term. Models change, and watching a "don't trust" row migrate to "check it" is itself a lesson in how this technology moves.