A Resume Stopped Being a Writing Sample

What Is a Resume Still Evidence Of?

Claims and priorities. A claim is any line with an issuer standing behind it, so somebody other than the candidate can settle it and there is a cost if it turns out false. A priority is the arrangement: what sits at the top, what got three lines, what is missing. Neither one is prose, which is why neither one moves when the prose is regenerated.

Three things it stopped being evidence of, and they are the three most screening advice still rests on:

• Writing ability. The document is now a collaboration by default, and there is no way to separate the halves.
• Care and attention to detail. A tidy layout and consistent tenses cost one prompt. An untidy one might mean a phone, a screen reader, or a plain-text export.
• Effort or interest. A tailored resume takes seconds, so tailoring no longer signals that anyone wanted the job in particular.

What replaces them is duller and more useful. A checkable claim gives the later stages something to act on. An ordering choice cannot be settled by anybody, and it still tells you what this person thinks their strongest work is, which is exactly what a first conversation should open on.

Why Grading the Prose Grades the Tool

Because writing is the task that got cheapest first. In a pre-registered experiment, 444 college-educated professionals completed occupation-specific writing tasks, and those given ChatGPT finished 37% faster with grades 0.45 standard deviations higher, with the largest gains going to the weakest writers ¹. A document that compresses the distance between a strong writer and a weak one cannot be used to tell them apart.

Those were paid online tasks with no colleagues, no revision cycle and no consequences, and the published version of the same study reports the effects differently, so the decimals are not the part to carry. The direction is, and the direction is enough. Prose quality on a resume now measures access and habit.

The fallback, reading the document and forming a private view, measures less than it feels like it does. Asked to tell GPT-3 text from human writing across stories, news and recipes, untrained evaluators performed at chance, and three quick training methods lifted them only to about 55% ². That was an older model generation, and the evaluators were crowdworkers rather than hiring managers reading in their own field. It sets a floor under the problem and settles nothing about any particular reader.

The cost of acting on the feeling anyway is not evenly distributed. Running detectors over 72 journal abstracts, one tool that was highly accurate on the clean human-versus-machine split over-flagged AI-assisted text from non-native authors at 25% against 11% for native authors ³. Small sample, academic abstracts, and detector versions from late 2024. The asymmetry is the part that transfers, and the mixed case, a person's own work polished by a model, is the case that actually arrives in a hiring inbox. When every document in the pile looks equally good, what to screen on once the resumes all look perfect is the practical version of this problem.

Read It as a Claims Index and a Priority Statement

Read it twice, for different things. First pass: mark every claim with an issuer behind it, so employers, dates, titles, degrees, licences, named products shipped. Second pass: read only the order and the space. What sits at the top, what got three lines, what is missing between two dates. The first pass produces a verification list and the second produces the questions worth asking.

The order is more informative than it looks, with one correction. A resume tailored to the posting will echo the posting's own priorities, so read the ordering against the job description rather than in isolation. Where the two match exactly, that is the posting talking. Where they differ, where someone led with the smaller project or gave the prestigious employer one line, a person made a choice and can explain it.

The omissions are the other half of the same reading. A gap, a role held for five months, a degree with no dates: none of those is a finding, and each is a good first question. The claims themselves go into a different pipeline entirely, because only some of what an application asserts can ever be checked, and knowing which is which stops a screen from spending its attention on the parts that will never resolve.

What this changes on Monday is small and specific. Stop scoring the summary paragraph. Stop counting quantified bullets as evidence of rigour, since a number is the easiest thing in the document to generate. Start writing one line per candidate naming the claim you would check and the question you would ask.

Which Questions Go to the Conversation?

The ones the document cannot settle, which is anything about judgment, ownership and choice. Send the checkable claims to verification, and carry everything else into a structured conversation where every candidate gets the same questions rated the same way. That format is worth roughly double what an unstructured chat about the same resume is worth, and the gap between two interview formats is larger than the gap between two methods ⁴.

Three question shapes do most of the work:

• Ownership. Which part of this did you decide, and what did you hand off. Works on any bullet, and it is a normal question rather than an accusation.
• Consequence. What broke, what did it cost, what changed afterwards. A resume rarely carries any of that, drafted by hand or not, so the answer has to come from the person.
• Rejection. Name something you produced and threw away. The answer is unavailable to anyone who was not there.

None of that requires knowing who typed the resume, which is the point. The three questions work identically on a hand-written document and a generated one, so the authorship question stops mattering the moment the conversation starts.

One thing worth checking separately is whether the screen itself is still earning its keep. If a large share of applicants is being cut on a document that now says less than it used to, finding out whether the resume screen is throwing away the wrong people matters more than tuning how it is read. The intake side has the same problem, and it is fixable in an afternoon: which application fields still produce signal determines what you are reading before any of this begins.

See a sample report >