Member-only story
Founder Notes
Building Software for Skeptics
Trustworthy tools should expose their evidence, limits, and consequences
A teacher receives a polished student essay and a score from a tool suggesting that a machine may have helped write it. The teacher has a real decision to make, but the score does not explain what the student understood, which passages they wrote, or what assistance was permitted. The student has a real problem too: a suspicious result can be difficult to contest when neither side can see how it was produced.
That concern is not merely hypothetical. Studies of early-generation AI-text detectors found inconsistent performance across different kinds of writing and transformations, as well as a risk of disproportionate false positives for non-native English writers.1, 2
This essay was developed with AI-assisted editorial revision; the ideas, design judgments, and final responsibility are mine.
This is a hypothetical example, not a claim about a particular school or tool. It illustrates the problem that interests me: when software enters a question of trust, it can make a difficult judgment more visible, or make uncertainty look like an answer.
Building for skeptics begins there. Not with the question, “How do we convince…
