Screening
Your Job Description Is Now the Prompt Every Resume Is Written From
Assume every word of a job description comes back to you: a candidate pastes it into a model, and the resume returns your phrases and your priorities, so overlap is now free to produce and worth nothing as a signal. Keep the plain occupational language, because boards and search still route on it. Add the facts a model cannot guess: pay, location, the manager, the first project. Move whatever you judge on outside the posting: a question with a wrong answer, an artifact, a constraint only a reader satisfies.
The takePhrase matching was never good, and the mirror only made its failure legible. A screen rewarding overlap with the posting was measuring how well someone wrote toward a document, which tracks coaching, spare time and confidence more closely than it tracks the work. Candidates have now automated that, cheaply and at scale. Reading it as cheating is the wrong call: a free tool is doing the task the process itself defined as valuable. The repair is to stop defining it as valuable.
Where Olive fits
Open a role and see what the work shows
No posting can tell you which application a model wrote, so Olive skips the document and assesses the person: a 40-to-60-minute assignment in the candidate's own occupation, done alongside an AI assistant, returned as six findings with a timestamped excerpt behind each one. The candidate is granted the same report, free.
Rank your shortlistWhat Happens to a Posting After You Publish It?
It becomes source material. A candidate pastes it into a model along with their history and gets an application written toward your language: your bullets restated as accomplishments, your priorities in the summary, your title in the headline. The widely used checkers take a posting and a resume and report an overlap figure, which makes your own text the optimization target.
The consequence for the writer is exact. Anything you put in the posting will be in most applications, so nothing you put in the posting can separate them. That is true of hard requirements, preferred qualifications, values statements and the tooling list alike. The more distinctive the phrasing, the more distinctive the echo.
Two things follow that most advice still gets backwards. First, you cannot recover the signal by spotting which applications were model-written. Researchers estimated that between 6.5% and 16.9% of the text submitted as peer reviews at four 2023 and 2024 conferences could have been substantially modified by a model, and the same paper says those corpus-level patterns may be too subtle to see in any individual document 1. A population estimate is not a verdict on one person.
Second, reading for it by eye is worse. Untrained evaluators asked to separate model-written text from human writing performed at chance, and three quick training methods lifted them only to about 55% 2. That was 2021 and earlier models, which makes the unaided guess weaker now rather than stronger. Every resume in the inbox looking perfect is the same finding arriving as a workday.
Stop Measuring an Application by Its Overlap With the Posting
Overlap is now free to produce, which makes it worthless to reward. Any screening step that scores an application by how closely it matches the posting is scoring the candidate's access to a tool and their willingness to paste. Drop it as a ranking input, drop the knockout questions that restate the requirements in the same words, and stop asking recruiters to eyeball whether the wording feels genuine.
What replaces it is anything a mirror cannot produce: a claim specific enough to be wrong, a piece of work, a named number, a person who can confirm it. Three practical substitutions:
- Instead of years of experience with a tool, ask what the candidate built with it, in one sentence with the constraint they hit.
- Instead of a keyword requirement, ask a question that has a right and a wrong answer for this role, and read the reasoning rather than the vocabulary.
- Instead of a cover letter, ask for one artifact with a decision in it and a line about the decision they would now reverse.
All three shift the read from language to content, and all three survive AI assistance, because a candidate who used a model to draft an answer still had to choose the project, the number and the reversal. Verifying an AI-proficiency claim without taking the resume's word for it works the same way at the next stage.
Write the Parts a Model Cannot Mirror
Write facts the candidate has no way to guess and no reason to restate: the pay band, the days on site, who the hire reports to, the first project by name, the tool the team is stuck with and hates. A model can mirror a description of a job. It cannot mirror this job, because these specifics were never published anywhere it could read them.
That is also what makes the posting useful to the person deciding whether to apply. Whatever a model reproduces was generic to begin with, and the specifics are what a candidate reads a posting for: four concrete facts usually land better than fifteen adjectives.
One structural change is worth making at the same time: put the discriminating ask in the application rather than in the description. A posting is a public document that will be copied, summarized and fed to a model within an hour of publication. An application question is answered once, by one person, on your form, and you can replace it on the next opening once the answers start converging. The posting changes that reduce generated volume covers how that ask is sized and worded.
When the hiring manager will not name the first project, that is worth pushing on rather than writing around. A manager who cannot say what the hire owns in month one has a headcount request rather than a role, and the posting will absorb the vagueness. Two questions usually produce it: what is not getting done right now, and who is doing it badly in the meantime.
Keep the Plain Occupational Language
Do not write around the mirror by getting clever. Job boards, search and internal search still route on ordinary occupational terms, so the title, the level and the core skills belong in plain words a person in that occupation would use. Losing the routing to defeat the copying is a bad trade: the candidates you want have to find the posting before any of this matters.
The restraint worth applying is to AI vocabulary specifically. Indeed Hiring Lab put AI-related postings at 6.3% of US postings in its August 2026 snapshot, past a prior peak of 3.3% in 2022 3. That is a growing minority and not a norm, and it counts postings that mention AI rather than roles that require it. Writing a posting that leads with AI language when the job is a marketing manager job mostly attracts applications written to that language.
A workable division:
- Plain and searchable: the title, the level, the occupation, the three or four skills the work genuinely runs on.
- Concrete and unmirrorable: pay, place, manager, first project, the constraint the team is under.
- Cut entirely: the values paragraph, the tool inventory, and any requirement nobody would reject a strong candidate for missing.
That third list is where most postings lose their length. It is also the material a model has the easiest time reflecting, which is a good test: if a sentence would come back verbatim in half your applications, it was never telling you anything.
Common questions
Should I stop publishing detailed requirements altogether?
No. Requirements tell a candidate whether to apply, which is most of the posting's job, and cutting them to defeat the mirror only makes that decision harder for the reader. What changes is downstream: do not use those same requirements as the yardstick at screening, because every application will match them. Publish the requirements for the reader, then judge on something the reader had to supply themselves.
Is a resume written with AI a reason to reject someone?
On its own, no, and you cannot reliably tell anyway. A resume is a formatting and summarizing task, which is exactly what these tools do well, and using one says nothing about how the person works. The judgment worth making is about the content: whether the claims are specific, whether they hold up in a conversation, and whether the work behind them exists. Reject for an unverifiable claim, not for polish.
Do keyword filters still do anything useful?
As routing, yes. As screening, less each month. A filter that surfaces postings to candidates and organizes an inbox by occupation is doing sensible work. A filter that ranks or rejects on phrase overlap is now sorting on a free rewrite, and it removes the candidate who did not think to paste the posting more often than it removes anyone unqualified.
How do I write a requirement that a model cannot restate?
Write one that has a wrong answer. 'Experience with forecasting' comes back in every application. 'Which of these two forecasts would you trust, and why?' does not, because the candidate has to commit to a position you can disagree with. The test is whether two qualified people could answer differently. If they could not, the requirement is a keyword.
What about internal postings, where nobody is competing at volume?
Internal candidates paste the posting into a model too, but the stakes flip: an internal candidate who mirrors the posting is telling you what they think you want, which is useful information about the role's reputation rather than about them. Keep the concrete facts, because an internal reader is deciding whether to leave a team they already have, and put the judging in a conversation about work already visible inside the company.
References
- 1. Monitoring AI-Modified Content at Scale: A Case Study on the Impact of ChatGPT on AI Conference Peer Reviews arxiv.org Supports the claim that model use can be estimated across a pile of documents but not judged in any single one: 6.5% to 16.9% of peer-review text, with the authors noting the pattern may be too subtle to see individually.
- 2. All That's 'Human' Is Not Gold: Evaluating Human Evaluation of Generated Text aclanthology.org Supports the claim that reading for model-written text by eye does not work: untrained evaluators performed at chance, and brief training lifted them only to about 55%.
- 3. US Labor Market Snapshot: August 2026 hiringlab.indeed.com Supports the claim that AI language in postings is a growing minority rather than a norm: 6.3% of US postings mention AI, past a prior peak of 3.3% in 2022.
3 sources, numbered by first appearance. How Olive sources claims
General guidance for hiring teams. What works at one company and one volume may not transfer to yours.
Olive assesses how a person works with AI. It does not detect AI-written documents, and it never produces a score, a ranking, or a match percentage for a person. Candidates read the same report the employer reads.