menu_book Part 3 · Honest AI Briefing
What the evidence supports.
Every prompt pack you've seen promises AI will get you hired. Here is what the strongest study says, what it does not say, and the two failure modes nobody selling prompts mentions.
What the evidence supports
The strongest study on this question is a field experiment with 480,948 job-seekers, published in Management Science in 2025. Candidates given algorithmic writing assistance on their resumes were hired 8% more often, at 10% higher wages. Employers were no less satisfied with the people they hired.1
That's a real result from a large sample, and it's the reason this kit exists.
What it does not support
The same study tested grammar and clarity correction — not a language model drafting your bullets from scratch. As of now, no published experiment measures callback or hire rates for generative-AI-written resumes against human-written ones. Not one.
Clearer, error-free, well-structured writing gets hired more often, and AI is one way to get there. The claim "ChatGPT will improve your callback rate" has never been tested. Anyone stating it as fact is guessing.
This distinction matters because it tells you what to use these prompts for. Use them to sharpen writing you could have written yourself with more time and a better editor. Don't use them to manufacture a professional identity you can't defend in a room.
Risk one · the self-preferencing trap
A 2025 study found that when language models are used to evaluate resumes, they favour resumes generated by the same model — self-preferencing in 67% to 82% of comparisons. Candidates whose resume came from the same model doing the screening were 23% to 60% more likely to be shortlisted, with content held equivalent.2
This is a preprint and hasn't completed peer review, so hold it loosely. But if it holds, the implication runs against everything prompt packs tell you. Optimising heavily for one model's style is a bet on which model your next employer happens to screen with. You can't know that. Write for a human reader, keep your own voice in the output, and you're not exposed to the bet either way.
Risk two · detectors, and who they punish
AI-text detectors are unreliable in a way that isn't evenly distributed. A peer-reviewed analysis found accuracy ranging from 55% to 97% depending on text type, length, and language — with the errors falling disproportionately on non-native English speakers.3
Read that again if English isn't your first language. Writing that is careful, correct, and slightly formal — exactly the register a second-language professional writes in — is the register these tools most often misclassify as machine-generated.
There is no clean defence, and pretending otherwise would be dishonest. What reduces exposure: keep specific details, real names, real numbers, and the odd irregular sentence that no model would produce. Generic polish is what gets flagged. Specificity is both better writing and better cover.
What AI cannot do for you
- Know what you did. It has no access to your work. Every number comes from you.
- Judge what's defensible. It'll happily produce a metric you can't survive being questioned on.
- Understand your market. It doesn't know what your industry pays or values this year.
- Make the ask. No model sends the referral message. That's still you.
Five rules
- Never ship a number you didn't supply. If it appears in output and not in your input, it's invented. Delete it.
- Read every line aloud before it goes live. If it doesn't sound like you speaking, it won't sound like you in the interview.
- Keep the irregularities. The slightly odd phrasing that's genuinely yours is worth more than smooth text.
- Use AI on the draft, not the decision. What to claim, where to apply, who to ask — yours.
- Assume a human reads it last. Because one does, and that's the only reader who makes an offer.
Bookmark this section. Come back to it every time you're tempted to paste raw output into your profile.