The Resume Has a Second Audience

A resume used to be written for one reader: a person. A piece published yesterday by HR Executive describes what that document looks like when the first reader is a model. Candidates increasingly write with AI screeners in mind, and some go further. In a Duke University study of roughly 200,000 resumes, about 1% contained prompt injections: hidden white-on-white text instructing the AI screener to rank the candidate favorably, regardless of what the resume actually says.

You can see the logic. Employers adopted AI screening because a single posting can draw more applications than any recruiter can meaningfully review. Once the machine reads first, writing for the machine is rational. Hidden instructions are just the aggressive version of the same move.

The temptation to answer with detection is where the piece gets uncomfortable. It cites Stanford research finding that leading AI detectors misclassified more than 60% of essays written by non-native English speakers as AI-generated. The tool built to catch manipulation flags honest writers instead.

My read: the weak point sits in the screening layer, not in the resume. Two qualifications stay honest. The 1% comes from a single dataset, so the true rate across hiring is unknown. And the evidence does not show which screening tools are actually vulnerable to injected text.

The useful move depends on your seat. If you screen with AI, the piece recommends stripping formatting before any document reaches the model, and I would treat detector scores as noise at that false-positive rate. If you apply, assume anything hidden gets stripped and make sure the resume still works as plain text.

Plain text is the one version both sides can trust. Whatever signal survives the strip is real; the rest was formatting.