Pipeline
Put the AI Rule in the Job Post and Watch Who Still Applies
Say in the job posting whether candidates can use AI. Put the sentence in the process section, with the stages, not in the requirements list: the assignment is AI-open, and the following call spends twenty minutes on the choices you made in it. That names which stages are AI-open and what happens after them. It costs a line. It makes the rule enforceable, because it was published before anyone applied. And it changes who applies, which no requirements bullet manages.
The takeFairness is the smallest reason to publish it. The obvious case, that candidates should not have to guess your rules, is true and undersells the sentence badly: it is the only filter in the posting that costs nothing, needs no new stage, and works before a resume has been read. Saying the assignment is AI-open and will be discussed live moves the people who cannot discuss it out of the pipeline on their own initiative, at the cheapest possible moment.
Where Olive fits
Open a role and see what the work shows
A published rule sets what a candidate may use, and it says nothing about what they do with it. Olive puts that part in front of them as work: an occupational assignment with an AI assistant that will do the whole thing if nobody stops it, returned as six findings a human reviewer wrote with the timestamp behind each one.
Rank your shortlistWhere does the sentence go, and what does it say?
In the process section, one or two sentences, sitting with the stages. It names a stage and a consequence: which parts of the loop are AI-open, which are not, and what the following conversation will ask about. Keep it out of the requirements list, where a sentence about AI stops describing your process and starts asserting a skill bar nobody has defined.
Three postures cover almost every role:
- Open, discussed: use whatever tools you use. The next call spends twenty minutes on the decisions in your submission.
- Open, declared: use what you use, and add two lines saying which tools did what.
- Closed for one stage: the live exercise runs without an assistant, and the take-home does not.
Each posture is a sentence, not a paragraph. The consequence clause is the part that does the work, because it converts a rule into something the candidate can prepare for and something an interviewer can act on without accusing anybody of anything.
This is a different sentence from an AI requirement in the job description, and mixing them is the common error. A requirement asserts a bar the candidate must clear before applying and has to be defensible as job related. A process rule describes what happens inside your loop. The first one needs a definition you can defend; the second one needs only to be true.
Why an unpublished rule cannot be enforced
Because enforcing it after the fact means proving something about a document, and that does not work. Untrained readers asked to tell GPT-3 text from human writing performed at random chance, and three quick training methods lifted them only to about 55%, inconsistently across the text types tested 1. That study is from 2021 and the models have improved since, which does not help the reader.
Tools do not close the gap. A 2023 study testing 12 publicly available detection tools plus Turnitin and PlagiarismCheck, in an academic-integrity setting, concluded the available tools are neither accurate nor reliable, and that they lean toward calling text human-written 2. On code the picture is worse: an ICSE 2025 evaluation of five detectors plus one built specifically for code found accuracy mostly below 0.6 across three benchmarks, which the authors call ineffective 3. Those were benchmark functions in Java, C++ and Python, a cleaner target than any take-home submission, so the result understates the problem.
So the enforceable version is contractual, not forensic. A rule published before anyone applied is a term of the process every applicant accepted, and you enforce it by talking to the candidate about the work. Nobody has to be accused of anything: the candidate who used four tools well can say so, and the one who cannot describe the choices in their own submission has told you what you needed to know.
It is also why the trick going around, hiding an instruction in the posting to trap AI-written applications, buys nothing a published rule does not buy honestly.
Write the rule stage by stage
Set a posture per stage rather than one rule for the whole loop, because the stages ask different things. The application and the assignment can be open. A live exercise may be closed for a defined window and open after it. A reference conversation is neither. Two sentences cover a four-stage loop, as long as the postures were chosen before the wording was.
A worked example for an analyst role:
- Application: use any tools you like. Nothing here is checked for authorship.
- Assignment, 90 minutes: AI-open. Bring the working as well as the answer.
- Panel: no assistant during the exercise, then fifteen minutes discussing how you would have used one.
- Offer stage: no AI rule applies.
Two details decide whether the wording survives. Ask what the tools did, and skip the prompt logs, because a log is a compliance artifact nobody reads and the answer you want is about judgment. And write the closed stage as a limit on the exercise: this one runs without an assistant. That is a statement about the hour, and it accuses nobody. Deciding all four postures at once is easier than it looks, and setting the rule stage by stage is the version most teams end up at.
A blanket ban is also a posture, and it should be published in the same place if that is the decision. The difference is that it inherits the enforcement problem from the section above, in full.
What changes in the pipeline once it is published
Composition changes before volume does. The people who withdraw wanted the assignment to be a finished artifact they could hand over, and they withdraw before a reviewer spends an hour on them. Expect the raw application count to hold roughly steady and the first-screen pass rate to move, which is the number actually worth watching.
Three measurements make the change readable, and all three are already in your ATS: applications per requisition, assignment completion rate, and the share of assignment submissions whose author can talk through their own decisions. The third one is a judgment call by the interviewer, so put it on the scorecard as a yes or no, where a note would get skimmed.
Instrument it rather than trusting the argument. No published study measures what one sentence in a posting does to an applicant pool, and anybody claiming a number for it is guessing. Run the sentence on half your open requisitions for a quarter and compare, which is cheap because the change is one line of template text.
The other half of the trade is who you stop attracting, and that is worth deciding on purpose before the posting goes up. A posting saying the loop is AI-open reads as an invitation to people who work that way and as a warning to people who do not, which is the same mechanism at work when a posting decides whether to require AI experience at all.
Common questions
Does the rule belong in the posting or in the assignment email?
Both, with the posting first. The assignment email is where the rule gets operational detail, such as the time budget and what to submit. The posting is where it becomes a term everyone accepted by applying. Repeating it costs a sentence, and a candidate who read it twice has no basis for surprise at the follow-up conversation.
What should the rule say about a live interview?
Write it as a fact about the exercise rather than a prohibition on the person: this exercise runs without an assistant. Then say what follows, usually a short discussion of how the candidate would have used one. That wording avoids the monitoring question entirely, which is the right outcome, because nothing you can do in a remote interview settles it anyway.
Can I ask candidates to disclose which tools they used?
Yes, and keep the ask small. Two lines naming what the tools did and what the candidate changed is useful input for the conversation. A prompt log is not: it is long, nobody reads it, and it converts a discussion about judgment into a documentation exercise. Ask in the assignment brief, not in the application form, where it reads as a screening question.
What if the decision is a blanket ban?
Publish it in the same place, in the same plain form, and go in knowing the enforcement problem comes with it. Detection tools are not accurate or reliable enough to carry a decision about a person, so a ban is held up by the candidate's agreement rather than by your ability to check. That is an argument for narrowing the ban to the one stage that needs it.
Does publishing the rule invite people to use AI who otherwise would not?
Probably a few, and that is not the risk worth managing. What publishing changes is that the use becomes discussable, which is the difference between an assignment you can interrogate and one you have to take on trust. The stated posture also gives the interviewer permission to ask directly, and that question is the one worth spending the round's time on.
References
- 1. All That's 'Human' Is Not Gold: Evaluating Human Evaluation of Generated Text aclanthology.org Supports the claim that unaided readers cannot tell AI text from human text, which is why an unpublished rule cannot be enforced by inspection.
- 2. Testing of Detection Tools for AI-Generated Text arxiv.org Supports the claim that available AI-text detection tools are neither accurate nor reliable enough to carry a decision about a candidate.
- 3. An Empirical Study on Automatically Detecting AI-Generated Source Code: How Far Are We? arxiv.org Supports the claim that a submitted code assignment cannot be checked for AI authorship, including by a detector built for code.
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.